The Framing of time-dependent machine learning risk prediction models among young individuals with acute coronary syndromes

The Framing of time-dependent machine learning risk prediction models among young individuals with acute coronary syndromes | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article The Framing of time-dependent machine learning risk prediction models among young individuals with acute coronary syndromes Luiz Sergio Carvalho, Gustavo Alexim, Ana Claudia Nogueira, Marta Fernandez, and 5 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-1824283/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 19 Jan, 2023 Read the published version in Scientific Reports → Version 1 posted 10 You are reading this latest preprint version Abstract Acute coronary syndrome (ACS) is a common cause of death in individuals older than 55 years. Although younger individuals are less frequently seen with ACS, this clinical event has increasing incidence trends and triggers considerable economic burden. Young individuals with ACS (yACS) are usually underrepresented and show idiosyncratic epidemiologic features compared to older subjects. These differences may justify why available risk prediction models usually penalize yACS with higher false positive rates compared to older subjects. We hypothesized that exploring temporal framing structures such as prediction time, observation windows and window shifts, coupled with subgroup-specific prediction, could improve time-dependent prediction metrics. In a large cohort of ACS individuals (n global_cohort =6341 and n yACS =2242), the predictive accuracy for adverse clinical events in ACS was optimized by using specific rules for yACS and splitting short-term and long-term prediction windows, leading to the detection of 80% of events, compared to 68% by using a rule designed for the global cohort. Acute coronary syndromes Risk prediction Clinical intelligence Population health management Figures Figure 1 Figure 2 Figure 3 Introduction In the last four decades, a major concern has arisen from the progressive increase in the incidence rates of acute coronary syndrome (ACS) among young individuals (yACS, i.e. before 55 years of age) 1 , 2 and the high recurrence rate of these events 3 . In addition to amplifying ACS-related reduction in quality of life and life expectancy, yACS also carries a heavy economic burden by reducing work capacity in early adulthood 3 , 4 , 5 . Young individuals with ACS are more likely to be men, smokers, obese, sedentary or to present familial-combined hyperlipidemia, and they more frequently consume cocaine or androgenic anabolic steroids than older ACS patients 6 , 7 . Furthermore, compared with older individuals, those with yACS have a higher proportion of traditional cardiovascular risk factors out of control 7 . Despite this clear distinction in the recurrence of coronary events and the prevalence of risk factors, effective risk prediction tools specific to individuals with yACS remain an unmet need. In clinical research, risk prediction tools infrequently explore the role of the observation window, i.e., the moment when predictors are captured, and the forecast window, i.e., the period from which the event is surveyed or sampled. As reviewed recently, the match between the clinical problem and these temporal framing structures is essential for high-quality predictive models 8 . This indicates that some “acute phase” information captured during ACS hospitalization may be useful to predict short-term outcomes but may not be useful to predict long-term outcomes 9 , 10 . We therefore hypothesized that splitting predictive rules into two (short- and long-term) would allow more accurate risk prediction. While for short-term outcomes it is commonly accepted that binary classification rules are reasonable, for instance major adverse cardiovascular events [MACE] vs non-MACE 9 , long-term predictive models need to consider time-to-event with competing events to minimize censoring bias 11 , 12 . To develop and validate models, we used interpretable and state-of-the-art algorithms for tabular data to predict in-hospital outcomes 13 and used survival analysis with competing risks 11 to predict long-term clinical events in a large cohort of yACS individuals. Furthermore, we studied the differences in risk factors and optimal prediction rules for yACS compared to ACS in older subjects. Results The study population had a mean age of 48±6 years, and 66% were male. Table 1 depicts the characteristics of yACS subjects (n=2242) compared to older individuals with ACS (n=4099). A total of 170 deaths (11.4 per 1000 patients-years), 132 STEMI (8.8 per 1000 patients-years) and 421 NSTEMI (28.2 per 1000 patients-years) occurred after a median follow-up of 6.67 years (95% confidence interval [CI] of 5.59-7.24) among yACS individuals. As described in Supplemental Table S1, in-hospital MACEs occurred in 180 individuals, and postdischarge MACEs occurred in 454 subjects with yACS. Among subjects older than 55 years old, in-hospital MACEs occurred in 493 individuals, and postdischarge MACEs occurred in 881. Younger subjects were more frequently smokers and obese and had a more frequent family history of premature CAD and personal history of alcohol or cocaine use (Table 1). Although type 2 diabetes mellitus (T2DM) was less frequently found and the global mean for HbA1c was lower among yACS, among subjects with T2DM, those with yACS presented a higher HbA1c (9.08±0.92%) than their older counterparts (8.12±1.09%; p<0.0001). Younger subjects were more frequently admitted due to STEMI, but the severity was generally lower than that in older subjects with ACS, as cardiac arrest before admission and Killip scores III or IV were less frequent. As expected, the burden of coronary artery disease was also lower among yACS individuals (Table 1). As shown, the yACS subgroup shows a highly distinctive epidemiologic profile compared with older ACS subjects. Therefore, we explored which are the key risk factors for MACEs in the short- and long-term among the two subgroups and found different patterns. Short-term MACE STWm has shown that there are significant differences in key predictors and model accuracy both by using stepwise LR (sLR) and more complex predictive algorithms. While the sLR model within the global cohort (training/validation with n=4439) showed an accuracy in the yACS test set (n=673) of 0.82 (95% CI of 0.79-0.84) and a C-statistic of 0.79 (95% CI of 0.77-0.81), an sLR developed specifically within the yACS individuals showed a significantly higher C- statistic of 0.88 (95% CI of 0.85-0.90, p for C- statistic comparison <0.001) in the yACS test set (Table 2). Supplementary Tables S2 and S3 show that the most important predictor variables in sLR to explain at least 90% of model variance were different in yACS and the global cohort. In yACS subjects, the odds ratios for MACE compared to the global cohort were higher for blood glycemia, prior chronic kidney disease (CKD), Killip class and syncope at ACS onset and lower for myocardial blush grade (MBG) and presence of dyskinesia (any wall). Increasing duration of catheterization (cath) was highly associated with MACEs in yACS and linked to intraprocedural complications such as coronary artery dissections (3.02% of yACS) and coronary rupture (0.1% of yACS). Late catheterization (12 h after symptom onset for STEMI and 24 h after symptom onset for UA/NSTEMI) was also an independent risk factor for MACE only in yACS. The sLR model trained in the yACS cohort performed as well as the random forest and XGBoost algorithms ( p for C- statistic comparisons of 0.68 and 0.77, respectively), and sLR was superior to the GRACE score-based model ( p =0.031) (Table 2). However, with a C- statistic of 0.92 (95% CI 0.89-0.95), the TabNet algorithm trained in the yACS cohort was superior to sLR ( p for C- statistic comparisons <0.001) and superior to TabNet trained in the global cohort (C-statistic of 0.90 (95% CI 0.88-0.92), p for C- statistic comparisons 0.011). As shown in Figure 3, 28 variables are included in the TabNet algorithm for the global cohort, and 20 are responsible for 91% of the model variance. In the yACS cohort, 24 variables were recruited, and 20 were responsible for 93% of the model variance. Among the top predictor variables that explain at least 90% of the model variance, Figure 1 shows very different patterns for the TabNet algorithm trained in yACS subjects and TabNet trained in the global cohort. Risk models share three variables in common (blood glycemia, BMI and right ventricular akinesia), and the algorithm trained in yACS contains characteristics related to microvascular thrombosis and intraprocedural complications of catheterization. Long-term MACE with competing risks Here, the clinical question is whether LTW m would be better suited for global follow-up (observation window of the first 48 hours) or whether it would perform better for postdischarge (from index ACS) risk prediction (observation window including in-hospital stay). Postdischarge models had 47 noncardiovascular deaths and 454 MACEs, while the global follow-up models had 92 noncardiovascular deaths and 631 MACEs among individuals with yACS. In the postdischarge models (available in Table 3), CS-Cox and Fine-Gray yielded the lowest C td indexes in the test set, 0.602 (95% CI 0.556-0.649) and 0.612 (95% CI 0.564-0.663), while DMGP and DeepHit reached 0.685 (95% CI 0.639-0.725) and 0.722 (95% CI 0.678-0.760), respectively. Global follow-up models (Table 3) produced generally lower concordance indexes in the test set, 0.597 (95% CI 0.552-0.643), 0.601 (95% CI 0.559-0.660), 0.687 (95% CI 0.640-0.728) and 0.681 (95% CI 0.654-0.703) for CS-Cox , Fine-Gray , DMGP and DeepHit, respectively. DeepHit in the postdischarge horizon yielded the highest C td index and the lowest IBS (0.0579), suggesting the highest accuracy. The CIFs of 12 random yACS individuals are depicted in Figure 2. C td -indexes for the global cohort were similar to yACS both in postdischarge and global follow-up horizons. Among the algorithms, only CS-Cox is easily interpretable; therefore, it was used to acquire a glance at the risk components for long-term MACE. As seen in Supplementary Tables S3 and S4, most short- and long-term MACE predictors differ significantly. Only the Killip class and prior CKD stood as predictors in both long-term CS-Cox and short-term sLR . The CS-Cox model in yACS individuals showed that drugs prescribed at discharge from index ACS, such as anticoagulants, furosemide and ticagrelor/prasugrel, are independently associated with MACEs. The atherosclerotic burden (Synthax score) and low ejection fraction were also linked to MACEs, but in yACS individuals, STEMI in index ACS showed reduced long-term risk compared to NSTEMI, and CABG as a treatment of index ACS was also associated with lower risk compared to PCI. Finally, the occurrence of non-fatal MACE during index ACS hospitalization was associated with an increased risk of long-term MACE. We did not observe differences in risk components for the global cohort and yACS individuals in the CS-Cox model. Discussion In this study, we found that individuals with yACS present different demographic characteristics and susceptibility to risk factors for MACEs compared to older subjects. We also identified that risk prediction models are optimized by using a compound strategy: (i) specific risk prediction rules for yACS individuals rather than targeted to the overall population; (ii) short-term predictions are highly efficient; and (iii) long-term prediction models should incorporate competing events and should be optimized by including in-hospital clinical data in the observation window. Roughly, the best model using this compound strategy led to the detection of 80% of events, compared to 68% by using general rules. As mentioned, risk prediction rules are improved by the optimal selection of observation windows. This issue was recently reviewed by Lauritsen et al. 8 and suggested that temporal framing structures are critical for successful risk prediction. In models for predicting sepsis, the authors suggested not only implementing optimal selection of observation/prediction windows but also including a sequential evaluation by using predictions made until the current timestep 14 . Indeed, Wong et al. suggested that a hospitalization-level risk score for sepsis based on the entire trajectory of predictions may enable more realistic evaluations 15 . However, in clinical cardiology, risk scores are typically less dynamic and employ temporal framing suboptimally. By setting a short-term endpoint, we could identify important predictors of in-hospital MACE with a set of data gathered from the first two days of index ACS onset. In parallel, long-term risk prediction taking into account competing risks was optimized by including predischarge information, including prescription at discharge and in-hospital clinical events. Another argument in favor of splitting two prediction windows is that we showed large differences between key predictors of MACE in the short term and the long term. In short-term models, the most important predictors of MACEs are symptoms at ACS presentation, microvascular thrombosis and intraprocedural complications of catheterization. Instead, in long-term models, the top predictors of worse clinical outcomes are mostly related to in-hospital outcomes, discharge medications, past medical history and severity of coronary artery lesions. In addition, splitting two prediction windows permits a flexible and dynamic way of dealing with clinical problems 14 . It is important to mention that binary classification can provide predictions for a predetermined duration (e.g., in-hospital stay), useful for short-term outcomes where time to event is not an issue. If transported to long-term risk modeling, binary classification typically ignores time censoring and may increase the risk for false negatives. Hence, in clinical problems with a substantial amount of censoring, the use of survival models tends to be advantageous 16 . On the other hand, if censoring bias is not accounted for or the context can neutralize censoring, binary classification tends to maximize accuracy compared to survival models 8 . Therefore, the way to better fit a real-world scenario was to combine short-term classification with long-term survival. For in-hospital MACE prediction, TabNet yielded the best results. The algorithm has been recently described and couples a deep neural network architecture and gradient descent-based optimization designed specifically for tabular data 13 . Together with the great predictive capacity, it also enables interpretability. Although no causality can be attributed to top predictors, they are consistent with the most prevalent risk factors for MACEs among yACS 4 , 17 . As observed by others 17 , we observed that variables of interest for predicting MACEs in individuals with premature ACS differed from the top predictors among the global cohort and older subjects. Among the long-term models, DeepHit was the most accurate. DeepHit is a multitask network that makes no linear assumptions during the predictive process, allowing for the possibility that the relationship between covariates and risks changes over time 11 . Although such architecture improves predictive ability and flexibility to deal with competing risks compared to CS-Cox and Fine-Gray models, it is not possible to interpret which variables are recruited at each step. However, among the long-term predictors of MACEs using CS-Cox , we identified that yACS may be at higher risk when prescribed at discharge drugs such as ticagrelor or prasugrel than clopidogrel. These observations contradict the findings from major clinical trials such as PLATO 18 and TRITON-TIMI-38 19 but should be explored in other real-world scenarios with appropriate techniques for neutralizing any potential selection bias. There are limitations in this study that should be acknowledged. First, the observational and retrospective design of this study limits any potential causal conclusions. Second, the definition of yACS is not consensus; while some consider an age threshold of 55 years old, others consider 50 or 45 years old 4 , 6 , 20 . Third, our models were trained in a relatively small cohort. Although the B-CaRe:QCO yACS cohort is among the largest cohorts of yACS, some algorithms, such as DeepHit , DMGP , and TabNet , were originally developed in datasets of > 10,000 individuals 11 , 13 , 21 . Our results suggest that these algorithms also perform well in smaller datasets, and we did our best to maximize external validity by using cross-validation and resampling techniques. The main advantage of our cohort is that we systematically included all subjects admitted due to ACS in public hospitals from Brasília (Brazil) who underwent coronarography up to 48 hours after hospital admission between January 2011 and February 2020. In summary, we found that individuals with premature ACS share considerable morbidity and show unique epidemiologic features compared to those of older subjects. In this study, we also identified that risk prediction models are optimized by using specific risk prediction rules for yACS individuals in two windows: a short-term window and a long-term window that incorporate competing events and in-hospital clinical data within the observation window. It is critical to better understand risk factors within this subgroup to allow public health initiatives that mitigate the economic burden aroused by yACS 4 , 6 . Risk prediction-enhanced clinical care could turn into a framework for intensified clinical surveillance in individuals predicted to be high risk 5 . Methods Study design and participants The set of individuals was selected from the B-CaRe:QCO ( Brasilia Cardiovascular Registry for Quality of Care and Outcomes ), a retrospective registry of 6341 subjects with ACS (n = 2242 with yACS). The B-CaRe:QCO study included consecutive individuals admitted to public hospitals in Brasília (DF) with ACS who underwent coronarography up to 48 h after hospital admission from January 2011 to February 2020. At that time, all coronarographies were carried out in Hospital de Base (Brasília-DF, Brazil) and Instituto de Cardiologia (Brasília-DF, Brazil). We excluded 17 individuals who died within the first 48 hours. Enrolled subjects experienced therapies based on guidelines for the treatment of ACS 22 . Attending physicians made all therapeutic decisions and were blinded to the study evaluations. Most individuals admitted due to STEMI (n = 1659 with premature ST-elevation myocardial infarction [STEMI]) were treated by primary percutaneous coronary intervention (pPCI) or pharmacoinvasive strategy. The methods were performed in accordance with relevant guidelines and regulations, and approved by the Institutional Ethics Review Board from Instituto de Gestão Estratégica em Saúde do Distrito Federal (IGESDF) (study protocol approval number [CAAE] 28530919.0.1001.8153). For predicting in-hospital MACE (defined as cardiovascular deaths or recurrent ACS) occurring 48 h after hospital admission, the observation window comprised the first 48 h after hospital admission. The yACS dataset was divided into a training/validation set (70%, n = 1569) and a test set (30%, n = 673). Short-term models (STW m ) were trained and validated in a 5-fold cross-validation framework with upsampling to mitigate outcome imbalance. STW m was then evaluated in the test set. To predict long-term outcomes with competing risks (noncardiovascular deaths vs MACE), two contexts were evaluated: (i) postdischarge , where an observation window included the whole period of index hospitalization (mean of 5 ± 2 days) and the outcomes were observed from hospital discharge to the end of follow-up (median of 6.67 years); (ii) global follow-up , where the observation window included only the first 48 h and the outcomes observation period began at 48 h and extended to the end of follow-up. A training/validation set (n = 1513) and test set (n = 648) included individuals alive at discharge and were used to train and validate long-term window models (LTW m ). LTW m was repeated over five cross-validation folds and then assessed in the test set. To better understand model accuracy and differences in key predictors for short-term MACE between the yACS and older subjects, we also created models using the global cohort (n = 6341) by splitting a training/validation set (n = 4439) and a test set including only 673 individuals in the yACS test set (remaining 1229 individuals older than 55 years were not included in the test set to prevent sampling imbalance). Again, we used 5-fold cross-validation with upsampling for STW m and evaluated the model in the yACS test set (n = 673). Clinical definitions and outcome assessment Current smokers were defined as those who had smoked at least 100 cigarettes during their lifetime and were smoking at least one year before ACS onset, according to the National Health Interview Survey (NHIS) definition 23 . Ex-smoking status was defined as smoking cessation for at least the last 6 months. Diabetes was defined as the use of antidiabetic medications, prior diagnosis of diabetes, or glycosylated hemoglobin (HbA1c) ≥ 6.5% at hospital admission. Patients were considered hypertensive if they were taking any antihypertensive medication or presented systolic blood pressure (SBP) ≥ 140 mm Hg or diastolic blood pressure (DBP) ≥ 90 mmHg. The anthropometric measurements obtained were body weight (kg), height (m), and waist circumference (cm). The Killip class and GRACE scores for in-hospital MACEs were evaluated in all enrolled patients 24 . Clinical outcomes were assessed by checking electronic health records (EHRs). Information about the cause of death and clinical events was obtained from the death certificate or medical records. The following adverse cardiac events for both STW m and LTW m were considered: cardiovascular deaths and recurrent ACS (MACE). For STW m , those who had any event during follow-up were marked as 1, and those who did not were coded as 0. For LTW m, we considered a competing event approach in survival analyses, i.e., individuals were followed until their deaths, the occurrence of recurrent ischemic events or the end of follow-up (last visit to the outpatient clinic registered in EHRs). Reinfarction was defined as the occurrence of new ischemic symptoms during the first 28 days after index MI associated with a > 20% increase in cTn levels after a 3-to-6-hour interval from symptoms 25 . Models and variable selection A domain-knowledge-driven approach was first used to select variables. From 186 variables at baseline, we excluded variables with no potential causal link with the outcomes and included those proven as predictors in previous models, leaving the remaining 108 variables. Variables were included only if they were unambiguous in their interpretation and recorded in a structured (numeric/binary) format. After this, a data-driven approach took place and consisted of an automated process based on actual data and the relevance of each variable to a specific outcome 26 . For most of the STW m and LTW m , we used a fully automated process incorporated into the algorithms. When selection could not be performed automatically, we followed guidelines as proposed by Belsley et al 27 : in the case of high correlation between variables (partial R 2 ≥ 0.5 in univariate regression with MACE[= 1] as the dependent variable or variance inflation factor [VIF] > 10), we dropped the variables with lower R 2 . Information-gain ranking was used to evaluate the worth of each variable by measuring the entropy gain with respect to the outcome, followed by ranking the attributes by their individual evaluations. Considering the tradeoffs between the cost of information and information gain, only attributes resulting in information gain higher than 0.01 were subsequently used in STW m and LTW m . Variable selection was performed in the training/validation dataset. Missing values (MVs) were relatively rare (2.7% of B-CaRe:QCO data). We handled MVs with multiple imputations directly in the training/validation dataset by using boosted trees. Only a few variables showed MV frequencies ≥ 10% (plasma TSH, free T4 and urea). Imputation using boosted trees fills each column by treating it as a regression problem. We did not impute missing values for the outcomes. Predictive algorithms For predicting short-term outcomes, we used XGBoost 28 , random forests 29 , and TabNet 13 . Random forests , based on decision trees, rank variable importance on the selection frequency of the variable as a decision node and generally show good performance for classification problems in tabular data with a single outcome 5 . XGBoost is also based on decision trees and uses gradient descent-based optimization 28 . TabNet has an interpretable canonical deep tabular data learning architecture, merging both deep learning and gradient descent-based optimization. The observation window was considered the first 2 days upon hospital admission and encompassed past medical history, emergency room data and coronarography. We compared models with the benchmark GRACE score 30 , recalibrated using regression coefficients of risk factors derived from logistic regressions (LR) as described elsewhere 5 (details in below). For long-term outcomes, we used the following survival algorithms with competing risks: cause-specific Cox-proportional hazards model ( CS-Cox ) 31 , Fine-Gray proportional subdistribution hazards model ( Fine-Gray ) 32 , deep multitask Gaussian process ( DMGP ) 21 , and DeepHit 11 . CS-Cox and Fine-Gray assume linear proportional hazards, DMGP assumes the underlying stochastic process to follow the Gaussian process, and DeepHit employs a network architecture that makes no assumptions about the relationship between predictors and outcomes. Each model’s hyperparameters were determined using the grid search method 33 and 5-fold cross-validation for STW m and LTW m . STW m were generated with upsampling to mitigate outcome imbalance. Performance in the validation set is reported as the mean of 5-folds. A full description of variable selection, hyperparameters and model architectures can be found below. Model Development Process To develop the prognostic models, B-CaRe:QCO data were extracted into a labelled dataset containing the independent variables (using the patients’ clinical records at their baseline dates or during index hospitalization) and all dependent variables (occurrence of a composite endpoint of death due to cardiovascular causes and recurrent ACS following the baseline date). We implemented a grid search for the hyperparameter optimization using the method reported by Bergstra and Bengio 33 . This requires the operator to specify a range of values for each hyperparameter, and all possible combinations of the hyperparameters are investigated, with the combination corresponding to the highest cross-validation performance metric (in this case, maximization of the C-statistics being chosen for the final model). The justification for selecting the hyperparameters that maximise the C-statistics is that this is less affected when the labelled data are unbalanced compared to using accuracy as a metric. When the classes are unbalanced, it is also a common strategy to oversample the rare label data and undersample the common label data, as many machine learning models can be sensitive to unbalanced data. Below, we describe in further detail the algorithms used. Short-term predictive algorithms for classification Random forests . For the hyperparameter grid search, we investigated ntree = 50, 150, and 350; mtry from 5 up to the maximum number of variables in increments of 5; max depth = 2, 4, 6, 8, and 10; and row samples of 90%, 95% and 100%. The chosen (optimal) random forest model had the following hyperparameters: ntree = 350, mtry = 25, max depth = 5 (up to 5 variable interactions were used by the model) and row sample fraction of 0.95 (95% of the data points were used to train each tree). XGboost. The grid search for the hyperparameters investigated in our models were ntree = 25, 50, 75 and 100; max depth = 2, 3, 4, 6 and 8; and the minimum observations per node was 5, 10, 20, and 40. The gradient boosting machine model was chosen to have a Bernoulli distribution, and the chosen model had the following hyperparameters: ntree = 50, max depth = 3 (up to 3 variable interactions were used by the model), and the minimum number of observations per node was 10. XGBoost was implemented in Python. TabNet. We used a canonical deep neural network (DNN) architecture for tabular data described by Arik et al 13 . Briefly, TabNet is trained using gradient descent-based optimization and uses sequential attention to choose which features to reason from at each decision step, enabling (i) interpretability, (ii) more accurate and faster learning and (iii) flexible integration into end-to-end learning. Through sparse and instancewise selection ( sparsemax is used for normalization of the coefficients) of features with the highest impact on outcomes, the learning capacity of a decision step is not wasted on irrelevant ones, and thus the model becomes more parameter efficient. TabNet also constructs a sequential multistep architecture, where each step contributes to a portion of the decision based on the selected features, improves the learning capacity via nonlinear processing of the selected features, and mimics ensembling via higher dimensions. The TabNet encoder is composed of a feature transformer, an attentive transformer and feature masking. A split block divides the processed representation to be used by the attentive transformer of the subsequent step as well as for the overall output. For each step, the feature selection mask provides interpretable information about the model’s functionality, and the masks can be aggregated to obtain global feature important attributions. The TabNet decoder is composed of a feature transformer block at each step. Each feature transformer block is composed of a 4-layer network, where 2 are shared across all decision steps and 2 are decision step-dependent. Each layer is composed of a fully connected (FC) layer, ghost batch normalization (BN) and gated linear unit (GLU) nonlinearity. We used standard classification ( softmax cross entropy) loss functions, and we trained the model until convergence using unsupervised pretraining. The final TabNet model was implemented in a PyTorch environment and had the following configuration: Adam optimizer with a learning rate of 0.02 and a decay rate of 0.9 every 10 interactions, Glorot uniform initialization, batch size of 256, Max epoch 1000, workers at zero, momentum of 0.9, N steps =8, γ = 2.0, and weight at 1 (automated sampling). Logistic regression models . We built a series of stepwise logistic regression models to predict in-hospital MACEs. Long-term predictive models – survival with competing risks Cause-specific Cox-proportional hazards model (Cox) and Fine-Gray proportional subdistribution hazards model (Fine-Gray). The Cox model relates the covariates to the hazard function of the outcome of interest and not directly to the survival times themselves. The covariates have a relative effect on the hazard function because of the use of the logarithmic transformation, and the regression coefficients are interpreted as log-hazard ratios. The hazard ratio is equal to the exponential of the associated regression coefficient 31 . Competing risks imply that a subject can experience one of a set of different events or outcomes. In this case, two different types of hazard functions are of interest: the cause-specific hazard function and the subdistribution hazard function. The cause-specific hazard function indicates the instantaneous rate of occurrence of the k th event in subjects who are currently event free (i.e., in subjects who have not yet experienced any of the different types of events). Considering two types of events, death attributable to cardiovascular causes and death attributable to noncardiovascular causes, the cause-specific hazard of cardiovascular death denotes the instantaneous rate of cardiovascular death in subjects who are still alive. It denotes the instantaneous risk of failure from the k th event in subjects who have not yet experienced an event of type k . There is a distinct cause-specific hazard function for each of the distinct types of events and a distinct subdistribution hazard function for each of the distinct types of events. In settings in which competing risks are present, two different hazard regression models are available: modeling the cause-specific hazard and modeling the subdistribution hazard function. The second model has also been described as a cumulative incidence function (CIF) regression model, which means that the subdistribution hazard model allows one to estimate the effect of covariates on the cumulative incidence function for the event of interest. However, it is recommended to use the Fine-Gray (FG) subdistribution hazard model when the focus is on estimating incidence or predicting prognosis in the presence of competing risks, since this model generally shows better accuracy than the Cox model. The (cause-specific) cumulative incidence function (CIF) expresses the probability that a particular event k* occurs on or before time t∗ conditional on covariates x* . Since true CIF is not known, the model utilizes estimated CIF to compare the risk of events occurring and to assess how models discriminate across cause-specific risks among patients. Model performance was calculated by using the time-dependent concordance index C td 34 (C td -index), which measures the extent to which the ordering of actual survival times of pairs agrees with the ordering of their predicted risk. Cox and FG benchmarks were run using the R libraries survival and cmprsk . We estimated the time-dependent C td index for the survival analysis methods under consideration using the function cindex of the R package pec . A deep multitask Gaussian process (DMGP) 21 is a nonparametric Bayesian model for survival analysis that relies on a conception of the competing risks problem as a multitask learning problem; i.e., it models the cause-specific survival times as the outputs of a random vector-valued function, the inputs to which are the patients’ covariates. This allows the model to learn a “shared representation” of survival times with respect to multiple related comorbidities. Inference of patient-specific posterior survival distribution is conducted via a variational Bayes algorithm. By using inducing variables to derive a variational lower bound on the marginal likelihood of the observed time-to-event data, which is maximized using the adaptive moment estimation algorithm ( Adam ). Hyperparameters Θ Z and Θ T were tuned using the offline B-CaRe:QCO dataset, and for any out-of-sample patient with all covariates, DMGP evaluates posterior probability density by direct Monte Carlo sampling. Hyperparameters were calibrated by maximizing the marginal likelihood of posterior probability density. DMGP was implemented in Python. DeepHit trains a neural network to learn the estimated joint distribution of survival time and event while capturing the right-censored nature inherent in survival data 11 . The network is trained by using a loss function that exploits both survival times and relative risks. DeepHit makes no assumptions about the underlying stochastic process and allows for the possibility that the relationship between covariates and risks changes over time. DeepHit is a multitask network that consists of a shared subnetwork and K cause-specific subnetworks, differing from that of a conventional multitask network in two ways: (i) it utilizes a single softmax layer as the output layer of DeepHit to ensure that the network learns the joint distribution of K competing events, not the marginal distributions of each event; (ii) it keeps a residual connection from the input covariates into the input of each cause-specific subnetwork. To train DeepHit , a total loss function L Total is specifically designed to handle censored data. This loss function is the sum of two terms L Total = L 1 + L 2 ; L 1 is the log-likelihood of the joint distribution of the first hitting time and event; L 2 incorporates a combination of cause-specific ranking loss functions that adapts the idea of concordance. The hyperparameters for L Total were selected based on the discriminative performance on the validation set. Early stopping was performed based on the total loss. DeepHit is a 4-layer network consisting of 1 fully connected layer for the shared subnetwork and 2 fully connected layers for each cause-specific subnetwork and a softmax layer as the output layer. For hidden layers, the number of nodes was set as 3, 5, and 3 times the covariate dimension for layers 1, 2, and 3, respectively, with the ReLu activation function. The network was trained by backpropagation via the Adam optimizer with a batch size of 50 and a learning rate of 0.0001. A dropout probability of 0.6 and Xavier initialization were applied for all layers. DeepHit was implemented in a TensorFlow environment in Python. Statistical analysis STW m were compared using accuracy and C-statistics for their performance on the test and validation datasets. We calculated the median performance and 95% confidence intervals (CIs) for the C-statistics for each algorithm. We built models with the training/validation set and finally evaluated the model on the test set to estimate performance. STW m was compared to the C-statistics obtained by the recalibrated GRACE score 30 . LTW m evaluates each individual’s cumulative incidence function (CIF), also known as the subdistribution function . CIF is commonly used in settings with competing risks and refers to the probability of a particular event during follow-up. CIFs are used to evaluate the case-specific concordance, and this concept is used to derive a performance metric to compare LTW m , the time-dependent concordance index C td 34 . The C td -index measures the extent to which the ordering of actual survival times of pairs agrees with the ordering of their predicted risk (further information is available in Supplemental Methods). A confidence interval for the C td index is derived using the jackknife method on correlated one-sample U-statistics. The integrated Brier score (IBS) was also used as an LTW m evaluation measure. Normally distributed data are presented as the mean ± SD, and skewed data are presented as the median [interquartile range (IQR)]. Normality of distribution and variances were checked using histograms, Kolmogorov-Smirnoff test, normal probability plots and residual scatter plots. Chi-square or two-tailed t -tests were used for comparison of baseline data. P-values < 0.05 were considered significant. Analyses were carried out using R[v4.0.1] and Python[v3.8], and the packages used are described in the Supplemental Methods. Declarations Conflict of Interest Disclosures: There are no conflicts of interest. Funding/Support: This work was supported by grant 2019/09068-3 from São Paulo Research Foundation (FAPESP) and grants 437413/2018-7 and 310718/2021-0 from the Brazilian National Research Council (CNPq). Role of the Funder/Sponsor: The funder had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication. DATA AVAILABILITY: Codes are available at https://github.com/lsergiocarvalho/openwindowACS. All requests for raw and analyzed data and related materials, excluding programming codes, will be reviewed by the Clarity Healthcare Intelligence legal department to verify whether the request is subject to any intellectual property or confidentiality obligations. Requests for patient-related data can be considered upon request. Any data and materials that can be shared will be released via a Material Transfer Agreement. DECLARATIONS: The authors had full access to all of the data (including statistical reports and tables) in the study and can take responsibility for the integrity of the data and the accuracy of the data analysis. They consent to the submission of the manuscript as it is. There are no financial and nonfinancial competing interests for all authors. The authors declare that they do not have a conflict of interest regarding the study. Funding/Support: This work was supported by grants 310718/2021-0 from the Brazilian National Research Council (CNPq), 371/2021 from FAPDF and 2019/09068-3 from FAPESP. Role of the Funder/Sponsor: The funder had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication. IRB Approval and Patient Consent : The study proceedings are in accordance with the Helsinki Declaration and the study was approved by the Institutional Ethics Review Board (IRB) from Instituto de Gestão Estratégica do Distrito Federal (IGESDF) (study protocol approval number [CAAE] 28530919.0.1001.8153). Since this is a retrospective study, the IRB approved the waiver of participants informed consent as long as data is captured anonymously. AUTHOR CONTRIBUTIONS Concept and design : Sposito, Carvalho, Fernandez, Avila, Nogueira, Alexim, Rezende Acquisition of data : Carvalho, Alexim, Nogueira Analysis and interpretations of data : Carvalho, Sposito, Fernandez, Alexim, Rezende, Reis, Rezende Drafting of the manuscript : Carvalho Critical revision of the paper for important intellectual content : Sposito, Avila, Fernandez, Alexim, Rezende, Nogueira, Soares, Reis Statistical analysis : Carvalho, Reis Provision of study materials or patients : Carvalho, Alexim, Nogueira Obtaining funding : Carvalho, Sposito Administrative, technical, or logistic support : Carvalho, Sposito, Avila, Alexim, Nogueira Supervision : Sposito, Avila, Carvalho References Arora S, Stouffer GA, Kucharska-Newton AM, et al. Twenty Year Trends and Sex Differences in Young Adults Hospitalized With Acute Myocardial Infarction. Circulation. 2019;139(8):1047-1056. Gupta A, Wang Y, Spertus JA, et al. Trends in acute myocardial infarction in young patients and differences by sex and race, 2001 to 2010. J Am Coll Cardiol. 2014;64(4):337-345. Zeitouni M, Clare RM, Chiswell K, et al. Risk Factor Burden and Long-Term Prognosis of Patients With Premature Coronary Artery Disease. J Am Heart Assoc. 2020;9(24):e017712. Dreyer RP, Raparelli V, Tsang SW, et al. Development and Validation of a Risk Prediction Model for 1-Year Readmission Among Young Adults Hospitalized for Acute Myocardial Infarction. J Am Heart Assoc. 2021;10(18):e021047. de Carvalho LSF, Gioppato S, Fernandez MD, et al. Machine Learning Improves the Identification of Individuals With Higher Morbidity and Avoidable Health Costs After Acute Coronary Syndromes. Value Health. 2020;23(12):1570-1579. Sagris M, Antonopoulos AS, Theofilis P, et al. Risk factors profile of young and older patients with Myocardial Infarction. Cardiovasc Res. 2021. Yanase T, Sakakura K, Taniguchi Y, et al. Comparison of Clinical Characteristics of Acute Myocardial Infarction Between Young (< 55 Years) and Older (55 to < 70 Years) Patients. Int Heart J. 2021;62(1):33-41. Lauritsen SM, Thiesson B, Jorgensen MJ, et al. The Framing of machine learning risk prediction models illustrated by evaluation of sepsis in general wards. NPJ Digit Med. 2021;4(1):158. Kuo RN, Dong YH, Liu JP, Chang CH, Shau WY, Lai MS. Predicting healthcare utilization using a pharmacy-based metric with the WHO's Anatomic Therapeutic Chemical algorithm. Medical care. 2011;49(11):1031-1039. Lauffenburger JC, Mahesri M, Choudhry NK. Use of Data-Driven Methods to Predict Long-term Patterns of Health Care Spending for Medicare Patients. JAMA Netw Open. 2020;3(10):e2020291. Lee C, Zame WR, Yoon J, Schaar Mvd. DeepHit: A Deep Learning Approach to Survival Analysis With Competing Risks. Paper presented at: XXXII Association for the Advancement of Artificial Intelligence (AAAI) Conference2018. McCaw ZR, Claggett BL, Tian L, et al. Practical Recommendations on Quantifying and Interpreting Treatment Effects in the Presence of Terminal Competing Risks: A Review. JAMA Cardiol. 2021. Arik SO, Pfister T. TabNet: Attentive Interpretable Tabular Learning. Association for the Advancement of Artificial Intelligence; 2020. Lauritsen SM, Kalor ME, Kongsgaard EL, et al. Early detection of sepsis utilizing deep learning on electronic health record event sequences. Artif Intell Med. 2020;104:101820. Wong A, Otles E, Donnelly JP, et al. External Validation of a Widely Implemented Proprietary Sepsis Prediction Model in Hospitalized Patients. JAMA Intern Med. 2021;181(8):1065-1070. Kvamme H, Borgan Ø, Scheel I. Time-to-Event Prediction with Neural Networks and Cox Regression. Journal of Machine Learning Research 2019;20:1-30. Lei L, Bin Z. Risk Factor Differences in Acute Myocardial Infarction between Young and Older People: A Systematic Review and Meta-Analysis. Int J Cardiovasc Sci. 2019;32(2). Wallentin L, Becker RC, Budaj A, et al. Ticagrelor versus clopidogrel in patients with acute coronary syndromes. N Engl J Med. 2009;361(11):1045-1057. Wiviott SD, Braunwald E, McCabe CH, et al. Prasugrel versus clopidogrel in patients with acute coronary syndromes. N Engl J Med. 2007;357(20):2001-2015. Divakaran S, Singh A, Biery D, et al. Diabetes Is Associated With Worse Long-term Outcomes in Young Adults After Myocardial Infarction: The Partners YOUNG-MI Registry. Diabetes Care. 2020;43(8):1843-1850. Alaa AM, van der Schaar M. Deep multi-task gaussian processes for survival analysis with competing risks. 30th Conference on Neural Information Processing Systems; 2017. Ibanez B, James S, Agewall S, et al. 2017 ESC Guidelines for the management of acute myocardial infarction in patients presenting with ST-segment elevation: The Task Force for the management of acute myocardial infarction in patients presenting with ST-segment elevation of the European Society of Cardiology (ESC). Eur Heart J. 2018;39(2):119-177. Ryan H, Trosclair A, Gfroerer J. Adult current smoking: differences in definitions and prevalence estimates--NHIS and NSDUH, 2008. J Environ Public Health. 2012;2012:918368. Fox KA, Dabbous OH, Goldberg RJ, et al. Prediction of risk of death and myocardial infarction in the six months after presentation with acute coronary syndrome: prospective multinational observational study (GRACE). BMJ. 2006;333(7578):1091. Thygesen K, Alpert JS, Jaffe AS, et al. Fourth Universal Definition of Myocardial Infarction (2018). Circulation. 2018;138(20):e618-e651. Heinze G, Wallisch C, Dunkler D. Variable selection - A review and recommendations for the practicing statistician. Biom J. 2018;60(3):431-449. Belsley; DA, Kuh; E, Welsch RE. Regression Diagnostics: Identifying Influential Data and Sources of Collinearity. 2nd edition ed: Wiley Interscience; 2013. Chen T, Guestrin C. XGBoost: A Scalable Tree Boosting System. . Paper presented at: KDD '16: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining2016; San Francisco. Breiman L. Random forests. Machine Learning. 2001;45(1):5-32. Granger CB, Goldberg RJ, Dabbous O, et al. Predictors of hospital mortality in the global registry of acute coronary events. Arch Intern Med. 2003;163(19):2345-2353. Austin PC, Lee DS, Fine JP. Introduction to the Analysis of Survival Data in the Presence of Competing Risks. Circulation. 2016;133(6):601-609. Fine JP, Gray RJ. A Proportional Hazards Model for the Subdistribution of a Competing Risk. Journal of the American Statistical Association. 1999;94 (446):496–509. Bergstra J, Bengio Y. Random search for hyper-parameter optimization. Journal of Machine Learning Research. 2012;13:281–305. Antolini L, Boracchi P, Biganzoli E. A time-dependent discrimination index for survival data. . Statistics in Medicine 2005;24:3927–3944. Tables Table 1 Baseline characteristics and clinical outcomes of individuals with premature acute coronary syndrome (ACS, ≤55 years old) and older subjects with ACS (>55 years old) ACS >55 years-old ACS ≤ 55 years-old p n 4099 2242 Age (mean (SD)) 68.47 (8.11) 47.63 (5.84) <0.001 Male gender (%) 60.8 66.1 <0.001 Diagnoses Index diagnosis (%) <0.001 STEMI 60.9 74.0 NSTEMI 20.1 15.5 UA 19.0 10.5 T2DM (%) 34.0 27.5 <0.001 T2DM on insulin (%) 9.4 8.1 0.177 Smokers (%) 34.7 41.2 <0.001 Dyslipidemia (%) 20.2 19.3 0.550 Hypertension (%) 74.7 66.9 <0.001 Obesity (%) 5.2 7.9 0.001 Family history of premature CAD (%) 9.8 16.6 <0.001 Prior ethylic habit (%) 9.7 14.1 <0.001 Prior drug abuse (%) 1.3 6.9 <0.001 Prior AMI (%) 8.5 7.1 <0.001 Prior stroke (%) 4.1 2.2 0.003 Prior PAD (%) 4.3 2.9 0.025 Prior CKD (%) 8.5 2.5 <0.001 Prior PCI (%) 8.9 6.2 0.002 Prior CABG (%) 5.6 3.7 0.007 Prior cocaine abuse (%) 0.1 1.3 <0.001 Prior marijuana abuse (%) 0.1 0.8 0.027 Atrial fibrillation (%) 3.7 0.8 <0.001 Drugs prescribed at discharge Nitrate (%) 45.1 42.5 0.123 Statin (%) 83.0 81.9 0.385 Betablockers (%) 64.9 66.6 0.293 ARB or ACEi (%) 58.4 61.7 <0.001 CCB (%) 20.0 18.5 0.278 ASA (%) 91.0 89.4 0.153 Clopidogrel (%) 62.5 64.4 0.244 Prasugrel (%) 21.8 22.0 0.941 Ticagrelor (%) 3.5 4.0 0.422 Anticoagulant (%) 3.3 3.5 0.816 Spironolactone (%) 11.7 7.9 <0.001 Furosemide (%) 15.2 9.4 <0.001 Coronary artery stenoses (%) LCA % (mean (SD)) 5.50 (17.66) 1.97 (10.40) <0.001 Proximal LAD % (mean (SD)) 30.07 (39.15) 29.18 (45.19) 0.079 LAD % (mean (SD)) 36.77 (39.87) 27.93 (38.85) <0.001 D1 % (mean (SD)) 20.79 (40.38) 15.28 (30.34) <0.001 Intermedium % (mean (SD)) 4.56 (18.38) 4.42 (18.19) 0.077 Cx % (mean (SD)) 35.27 (39.66) 24.17 (36.35) <0.001 M1 % (mean (SD)) 18.08 (33.17) 11.12 (27.64) <0.001 RCA % (mean (SD)) 50.77 (40.71) 37.84 (41.10) <0.001 PDA % (mean (SD)) 6.94 (21.88) 5.19 (18.70) <0.001 RMA % (mean (SD)) 5.44 (19.88) 4.12 (17.51) <0.001 Severe coronary artery lesions 1-vessel with proximal LAD (%) 4.8 8.8 <0.001 1-vessel with LAD (%) 8.5 9.9 <0.001 1-vessel with RCA (%) 10.4 12.6 <0.001 2-vessels without LAD (%) 13.1 11.2 0.133 3-vessels with LAD (%) 13.3 8.9 <0.001 3-vessels without LAD (%) 16.1 10.6 <0.001 3-vessels with LCA and LAD (%) 1.8 0.4 < 0.001 3-vessels with LCA / without LAD (%) 1.2 0.3 0.007 PCI - index coronarography LAD (%) 13.5 18.5 <0.001 MINOCA (%) 2.4 1.7 0.223 Number of new stents (mean (SD)) 1.38 (0.72) 1.44 (0.70) <0.001 Echocardiography Apical dyskinesia 3.5 1.1 <0.001 Apical akinesia 29.6 23.7 45% 22.3 24.4 <45% 77.7 75.6 Admission metrics (STEMI individuals only; n=4042) Cardiac arrest before admission (%) 2.3 0.3 <0.001 Time pain-primary hospital, minutes (mean (SD)) 164.51 (142.39) 153.66 (131.44) 0.043 Time door-needle, minutes (median [IQR]) 70.00 [43.00, 120.00] 68.00 [43.00, 110.50] 0.269 Time pain-needle, minutes (median [IQR]) 225.00 [150.00, 335.00] 210.00 [140.00, 315.00] 0.005 Time tnk-coronarography, minutes (mean (SD)) 1195.82 (1269.01) 1270.69 (1137.35) 0.114 Coronarography duration, minutes (median [IQR]) 55.00 [40.00, 75.00] 50.00 [40.00, 65.00] <0.001 Pharmacoinvasive strategy (%) 88.9 90.2 0.698 Primary PCI (%) 11.1 9.8 0.452 SBP at admission, mmHg (mean (SD)) 121.86 (26.01) 143.43 (25.22) <0.001 DBP at admission, mmHg (mean (SD)) 75.46 (17.19) 88.18 (17.21) <0.001 HR at admission, beats/minute (mean (SD)) 81.00 (19.47) 77.40 (15.44) <0.001 Killip score (%) <0.001 I 43.5 84.0 II 25.9 14.2 III 16.2 0.9 IV 14.4 0.9 TIMI flow pre-PCI (mean (SD)) 2.05 (1.22) 2.20 (1.18) 0.003 TIMI flow post-PCI (mean (SD)) 2.58 (0.81) 2.69 (0.75) 0.002 MBG pre-PCI (mean (SD)) 1.61 (1.45) 1.84 (1.42) <0.001 MBG post-PCI (mean (SD)) 1.86 (1.35) 2.18 (1.22) <0.001 Clinical scores TIMI score (mean (SD)) 5.15 (2.35) 2.46 (1.56) <0.001 GRACE in-hospital death (mean (SD)) 150.80 (32.75) 89.47 (18.19) <0.001 GRACE score (6 months) (mean (SD)) 141.49 (24.28) 91.61 (15.96) <0.001 CRUSADE (mean (SD)) 35.75 (13.87) 19.44 (11.29) <0.001 Laboratory exams Troponin (peak) (mean (SD)) 9999 (10399) 6838 (6465) <0.001 Glycemia, mg/dL (mean (SD)) 159.51 (81.38) 137.86 (62.80) <0.001 HbA1c, % (mean (SD)) 6.87 (2.03) 6.63 (2.01) 0.014 Total cholesterol, mg/dL (mean (SD)) 194.74 (48.58) 204.71 (46.58) <0.001 HDL-cholesterol, mg/dL (mean (SD)) 42.01 (12.26) 41.02 (12.68) 0.069 LDL-cholesterol, mg/dL (mean (SD)) 122.82 (39.59) 132.19 (40.23) <0.001 AST, mg/dL (median [IQR]) 185.00 [80.00, 334.00] 148.00 [75.00, 272.75] <0.001 ALT, mg/dL (median [IQR]) 50.00 [29.00, 83.75] 44.00 [29.00, 69.00] <0.001 Triglycerides, mg/dL (mean (SD)) 152.54 (133.41) 166.06 (118.62) 0.009 Creatinine, mg/dL (median [IQR]) 0.98 [0.81, 1.23] 0.84 [0.71, 1.00] <0.001 Creatinine clearance, ml/min/1.73m2 (mean (SD)) 74.52 (28.94) 107.10 (33.94) <0.001 BMI, kg/m2 (mean (SD)) 26.65 (4.57) 27.15 (4.51) 0.005 Clinical outcomes (considering competing risks) In-hospital deaths (%) 5,7 3,9 <0.001 Post-discharge deaths (%) 9,6 4 <0.001 Global deaths (%) 15.3 7.6 <0.001 Global CV deaths (%) 5.2 3.5 0.018 Global non-CV deaths (%) 10.1 4.0 <0.001 MI during follow-up (%) <0.001 STEMI 12.8 5.9 NSTEMI 15.5 18.8 Follow-up time, days (median [IQR]) 2331 [2030, 2625] 2436 [2039, 2644] <0.001 CABG (in-hospital) (%) 4.7 2.8 0.017 CABG during long-term follow-up (%) 13.3 10.9 0.026 CABG: coronary artery bypass graft; CKD: chronic kidney disease; GFR: glomerular filtration rate (CKD-EPI); GPIIbIIIa: glycoprotein IIbIIIa; HbA1c: glycosylated hemoglobin; LDL-C: low-density lipoprotein cholesterol; LV: left ventricle; BMI: body mass index; NSTEMI: non-ST-elevation myocardial infarction; PCI: percutaneous coronary intervention; STEMI: ST-elevation myocardial infarction; SBP: systolic blood pressure; DBP: diastolic blood pressure; HR: heart rate; CV: cardiovascular. Table 2 Accuracy and C-statistics for short-term window models in predicting in-hospital death or recurrent ischemic events in 2,242 individuals with premature ACS (55 years old or younger), total number of events = 180 ( in-hospital CV deaths = 39, and MI = 141) Accuracy (95% confidence interval) C-statistics (95% confidence interval) Logistic regression with GRACE score risk factors Validation set (mean of 5-folds) 0.841 (0.819 - 0.866) 0.834 (0.819 - 0.866) Test set (n=673, 75 events) 0.820 (0.784 - 0.856) 0.819 (0.782 - 0.853) Logistic regression (top 20 predictors) Validation set (mean of 5-folds) 0.889 (0.867 - 0.904) 0.883 (0.867 - 0.904) Test set (n=673, 75 events) 0.880 (0.854 - 0.896) 0.872 (0.851 - 0.894) Random Forests (top 20 predictors) Validation set (mean of 5-folds) 0.901 (0.870 - 0.928) 0.908 (0.879 - 0.933) Test set (n=673, 75 events) 0.892 (0.852 - 0.930) 0.888 (0.846 - 0.927) XGBoost (top 20 predictors) Validation set (mean of 5-folds) 0.894 (0.870 - 0.929) 0.898 (0.872 - 0.931) Test set (n=673, 75 events) 0.876 (0.859 - 0.903) 0.861 (0.835 - 0.891) TabNet (top 20 predictors) Validation set (mean of 5-folds) 0.951 (0.924 - 0.979) 0.936 (0.904 - 0.955) Test set (n=673, 75 events) 0.946 (0.917 - 0.975) 0.921 (0.889 - 0.953) Table 3 Time-dependent C-statistics (C td -index) for predicting long-term noncardiovascular deaths or MACEs (cardiovascular deaths and recurrent ischemic events) with competing risks occurring (i) after discharge in 2,161 individuals with premature ACS (55 years old or younger), total number of postdischarge events: 47 noncardiovascular deaths and 454 MACEs; (ii) 48 h after index ACS hospital admission ( global follow-up models), total number of events: 92 noncardiovascular deaths and 631 MACEs. C td -index (95% confidence interval) Post-discharge horizon CS-Cox 0.602 (95% CI 0.556-0.649) Fine-Gray 0.612 (95% CI 0.564-0.663) DMGP 0.685 (95% CI 0.639-0.725) DeepHit 0.722 (95% CI 0.678-0.760) Global follow-up horizon CS-Cox 0.597 (95% CI 0.552-0.643) Fine-Gray 0.601 (95% CI 0.559-0.660) DMGP 0.687 (95% CI 0.640-0.728) DeepHit 0.681 (95% CI 0.654-0.703) Additional Declarations No competing interests reported. Supplementary Files SupplementFile124062022.docx Cite Share Download PDF Status: Published Journal Publication published 19 Jan, 2023 Read the published version in Scientific Reports → Version 1 posted Editorial decision: Major revision 25 Nov, 2022 Reviews received at journal 17 Nov, 2022 Reviewers agreed at journal 01 Nov, 2022 Reviews received at journal 08 Aug, 2022 Reviewers agreed at journal 18 Jul, 2022 Reviewers invited by journal 12 Jul, 2022 Editor assigned by journal 12 Jul, 2022 Editor invited by journal 11 Jul, 2022 Submission checks completed at journal 11 Jul, 2022 First submitted to journal 04 Jul, 2022 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-1824283","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":120275716,"identity":"5e3b74b5-d59d-4846-9cfd-43726b07b047","order_by":0,"name":"Luiz Sergio Carvalho","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA7UlEQVRIiWNgGAWjYBACPgbGBjCDjYHxAcMHEIOdgBY2hBZmA8YZIAYzQS1wwGzAzAOmCWmRSG7++KPinl0fezPjZ5tf2+T5mBkYP3zMwaclscGY50xxchvPYWbp3L7bhm3MDMySM7fh15LM2JaQzCaRf4w5t+c2I1ALGzMvAS0Hf/4DaUlmY7bsuW1PjJbGBt6GBDuwFoYftxMJa+F52MzMcywhgQ3oF8nehtvJbcyMzXj9ws+e/vjjj5oEe/n2ZsYPP/7ctp3f3nzww0c8WmAgsQFEMraByQbC6oHAHkL9IUrxKBgFo2AUjDAAAAXXR/Oyrq8vAAAAAElFTkSuQmCC","orcid":"","institution":"Clarity Healthcare Intelligence","correspondingAuthor":true,"prefix":"","firstName":"Luiz","middleName":"Sergio","lastName":"Carvalho","suffix":""},{"id":120275722,"identity":"adb60d5b-59a4-4604-9974-0bc6c7c3111d","order_by":1,"name":"Gustavo Alexim","email":"","orcid":"","institution":"IGESDF","correspondingAuthor":false,"prefix":"","firstName":"Gustavo","middleName":"","lastName":"Alexim","suffix":""},{"id":120275725,"identity":"c6c25080-cb97-4539-9f7c-74e060cac027","order_by":2,"name":"Ana Claudia Nogueira","email":"","orcid":"","institution":"Aramari Apo Institute for Education and Clinical Research","correspondingAuthor":false,"prefix":"","firstName":"Ana","middleName":"Claudia","lastName":"Nogueira","suffix":""},{"id":120275729,"identity":"7ba0d5ef-6569-4bdb-9055-2a5f2e5c3b45","order_by":3,"name":"Marta Fernandez","email":"","orcid":"","institution":"Clarity Healthcare Intelligence","correspondingAuthor":false,"prefix":"","firstName":"Marta","middleName":"","lastName":"Fernandez","suffix":""},{"id":120275731,"identity":"7b1f7345-d08d-4429-8163-de9c0859595e","order_by":4,"name":"Tito Rezende","email":"","orcid":"","institution":"Clarity Healthcare Intelligence","correspondingAuthor":false,"prefix":"","firstName":"Tito","middleName":"","lastName":"Rezende","suffix":""},{"id":120275733,"identity":"f8470cdf-27a7-411e-874a-6fdd7ef4fea8","order_by":5,"name":"Sandra Avila","email":"","orcid":"","institution":"UNICAMP","correspondingAuthor":false,"prefix":"","firstName":"Sandra","middleName":"","lastName":"Avila","suffix":""},{"id":120275735,"identity":"e985f876-3314-44ea-b661-e9da2b3fdd5c","order_by":6,"name":"Ricardo Reis","email":"","orcid":"","institution":"University of Brasília","correspondingAuthor":false,"prefix":"","firstName":"Ricardo","middleName":"","lastName":"Reis","suffix":""},{"id":120275737,"identity":"38129178-54ac-42b2-991a-9f7d4be6164b","order_by":7,"name":"Alexandre Anderson Soares","email":"","orcid":"","institution":"Aramari Apo Institute for Education and Clinical Research","correspondingAuthor":false,"prefix":"","firstName":"Alexandre","middleName":"Anderson","lastName":"Soares","suffix":""},{"id":120275739,"identity":"99c8af92-3eb3-4e57-ad27-dfc01429783e","order_by":8,"name":"Andrei Sposito","email":"","orcid":"","institution":"UNICAMP","correspondingAuthor":false,"prefix":"","firstName":"Andrei","middleName":"","lastName":"Sposito","suffix":""}],"badges":[],"createdAt":"2022-07-04 14:44:27","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-1824283/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-1824283/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41598-023-27776-0","type":"published","date":"2023-01-19T18:24:43+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":24052561,"identity":"4ff15110-d898-4c43-ad36-5700abcb276a","added_by":"auto","created_at":"2022-07-19 17:22:57","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":61660,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTop predictors to explain at least 90% of variance in \u003cem\u003eTabNet\u003c/em\u003e model for \u003cem\u003ein-hospital\u003c/em\u003e MACE\u003c/strong\u003e. Legend: (\u003cem\u003ea\u003c/em\u003e) the global cohort and (\u003cem\u003eb\u003c/em\u003e) young individuals with ACS (≤ 55 years old)\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-1824283/v1/e57cab593ff1aa1ca8501ec2.png"},{"id":24052563,"identity":"b2248c10-8ba0-41ce-89f8-6c41f7ced890","added_by":"auto","created_at":"2022-07-19 17:22:57","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":135742,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eEstimated cumulative incidence functions (CIFs) for 12 random individuals with premature acute coronary syndrome (yACS) by using the \u003cem\u003eDeepHit\u003c/em\u003e algorithm in\u003cem\u003e the postdischarge\u003c/em\u003e horizon\u003c/strong\u003e.\u003c/p\u003e\u003cp\u003eLegend: zero (0) denotes CIF for noncardiovascular death, while one (1) means CIF for major cardiovascular adverse events (MACEs, cardiovascular deaths and recurrent ischemic events) occurring after discharge from index ACS hospitalization.\u003c/p\u003e\u003cp\u003e\u003cbr\u003e\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-1824283/v1/344aab504e64956aa7f24278.png"},{"id":24052560,"identity":"95370b9f-1b9d-4769-8150-684798546ecb","added_by":"auto","created_at":"2022-07-19 17:22:57","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":250949,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFeature importance masks (indicating feature selection at the \u003cem\u003ei\u003c/em\u003e\u003c/strong\u003e\u003csup\u003e\u003cstrong\u003e\u003cem\u003eth\u003c/em\u003e\u003c/strong\u003e\u003c/sup\u003e\u003cstrong\u003e step) and the aggregate feature importance mask (mask 0) showing the global instancewise feature selection\u003c/strong\u003e on the global cohort (\u003cem\u003ea\u003c/em\u003e) and young subjects with ACS (\u003cem\u003eb\u003c/em\u003e). Brighter colors show higher values.\u003c/p\u003e\u003cp\u003eLegend: In the global cohort, 28 variables are recruited, and 20 are responsible for 90% of the model variance. In the yACS cohort, 24 variables were recruited, and 20 were responsible for 90% of the model variance.\u003c/p\u003e\u003cp\u003e\u003cbr\u003e\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-1824283/v1/0d5b8a010eb97b13cd1b84dd.png"},{"id":44717354,"identity":"d02adc92-2965-44aa-b4aa-ebc9e9b5dce1","added_by":"auto","created_at":"2023-10-16 18:34:09","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":846499,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-1824283/v1/a1f913e4-31b4-45ea-a738-c3534cfc79d7.pdf"},{"id":24052691,"identity":"5a4f51f5-961c-4f8c-9789-6a377920eedb","added_by":"auto","created_at":"2022-07-19 17:27:57","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":39484,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementFile124062022.docx","url":"https://assets-eu.researchsquare.com/files/rs-1824283/v1/da9f4d639dfdaf7c272a2d9b.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"The Framing of time-dependent machine learning risk prediction models among young individuals with acute coronary syndromes","fulltext":[{"header":"Introduction","content":"\u003cp\u003eIn the last four decades, a major concern has arisen from the progressive increase in the incidence rates of acute coronary syndrome (ACS) among young individuals (yACS, i.e. before 55 years of age)\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e,\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e and the high recurrence rate of these events\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. In addition to amplifying ACS-related reduction in quality of life and life expectancy, yACS also carries a heavy economic burden by reducing work capacity in early adulthood\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e,\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e,\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eYoung individuals with ACS are more likely to be men, smokers, obese, sedentary or to present familial-combined hyperlipidemia, and they more frequently consume cocaine or androgenic anabolic steroids than older ACS patients\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e,\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. Furthermore, compared with older individuals, those with yACS have a higher proportion of traditional cardiovascular risk factors out of control\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. Despite this clear distinction in the recurrence of coronary events and the prevalence of risk factors, effective risk prediction tools specific to individuals with yACS remain an unmet need.\u003c/p\u003e \u003cp\u003eIn clinical research, risk prediction tools infrequently explore the role of the observation window, i.e., the moment when predictors are captured, and the forecast window, i.e., the period from which the event is surveyed or sampled. As reviewed recently, the match between the clinical problem and these temporal \u003cem\u003eframing structures\u003c/em\u003e is essential for high-quality predictive models\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e. This indicates that some \u003cem\u003e\u0026ldquo;acute phase\u0026rdquo;\u003c/em\u003e information captured during ACS hospitalization may be useful to predict short-term outcomes but may not be useful to predict long-term outcomes\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e,\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e. We therefore hypothesized that splitting predictive rules into two (short- and long-term) would allow more accurate risk prediction.\u003c/p\u003e \u003cp\u003eWhile for short-term outcomes it is commonly accepted that binary classification rules are reasonable, for instance \u003cem\u003emajor adverse cardiovascular events\u003c/em\u003e [MACE] vs non-MACE\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e, long-term predictive models need to consider time-to-event with competing events to minimize censoring bias\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e,\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. To develop and validate models, we used interpretable and \u003cem\u003estate-of-the-art\u003c/em\u003e algorithms for tabular data to predict \u003cem\u003ein-hospital\u003c/em\u003e outcomes\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e and used survival analysis with competing risks\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e to predict long-term clinical events in a large cohort of yACS individuals. Furthermore, we studied the differences in risk factors and optimal prediction rules for yACS compared to ACS in older subjects.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eThe study population had a mean age of 48\u0026plusmn;6 years, and 66% were male. Table 1 depicts\u0026nbsp;the\u0026nbsp;characteristics of yACS subjects (n=2242) compared to older individuals with ACS (n=4099). A total\u0026nbsp;of\u0026nbsp;170 deaths (11.4 per 1000 patients-years), 132 STEMI (8.8 per 1000 patients-years) and 421 NSTEMI (28.2 per 1000 patients-years) occurred after a median follow-up of 6.67 years (95% confidence interval [CI] of 5.59-7.24) among yACS individuals. As described in Supplemental Table S1, \u003cem\u003ein-hospital\u003c/em\u003e MACEs\u0026nbsp;occurred in 180 individuals,\u0026nbsp;and\u0026nbsp;postdischarge MACEs\u0026nbsp;occurred in 454 subjects with yACS. Among subjects older than 55 years\u0026nbsp;old, \u003cem\u003ein-hospital\u003c/em\u003e MACEs\u0026nbsp;occurred in 493 individuals,\u0026nbsp;and\u0026nbsp;postdischarge MACEs\u0026nbsp;occurred in 881.\u003c/p\u003e\n\u003cp\u003eYounger subjects were more frequently smokers and obese and had a more frequent family history of premature CAD and personal history of alcohol or cocaine use (Table 1). Although type 2 diabetes mellitus (T2DM) was less frequently found and\u0026nbsp;the\u0026nbsp;global mean for HbA1c was lower among yACS, among subjects with T2DM,\u0026nbsp;those with yACS presented a higher HbA1c (9.08\u0026plusmn;0.92%) than their older counterparts (8.12\u0026plusmn;1.09%; p\u0026lt;0.0001). Younger subjects were more frequently admitted due to STEMI,\u0026nbsp;but the severity was generally lower than\u0026nbsp;that in\u0026nbsp;older subjects with ACS, as cardiac arrest before admission and Killip scores III or IV were less frequent. As expected, the burden of coronary artery disease was also lower among yACS individuals (Table 1).\u003c/p\u003e\n\u003cp\u003eAs shown, the yACS subgroup shows a highly distinctive epidemiologic profile compared with older ACS subjects. Therefore, we explored which are the key risk factors for MACEs in the short- and long-term among the two subgroups and found different patterns.\u0026nbsp;\u003c/p\u003e\n\u003ch2\u003e\u003cem\u003eShort-term MACE\u003c/em\u003e\u003c/h2\u003e\n\u003cp\u003eSTWm has shown that there are significant differences in key predictors and model accuracy both by using stepwise LR (sLR) and more complex predictive algorithms. While\u0026nbsp;the\u0026nbsp;\u003cem\u003esLR\u003c/em\u003e model within the global cohort (training/validation with n=4439)\u0026nbsp;showed an\u0026nbsp;\u003cem\u003eaccuracy\u003c/em\u003e in the yACS test\u0026nbsp;set (n=673) of 0.82 (95% CI of 0.79-0.84) and\u0026nbsp;\u003cem\u003ea C-statistic\u003c/em\u003e of 0.79 (95% CI of 0.77-0.81),\u0026nbsp;an\u0026nbsp;sLR developed specifically within the yACS individuals showed\u0026nbsp;a\u0026nbsp;significantly higher \u003cem\u003eC-\u003c/em\u003e\u003cem\u003estatistic\u003c/em\u003e of 0.88 (95% CI of 0.85-0.90, \u003cem\u003ep\u003c/em\u003e for \u003cem\u003eC-\u003c/em\u003e\u003cem\u003estatistic\u003c/em\u003e comparison \u0026lt;0.001) in\u0026nbsp;the\u0026nbsp;yACS test\u0026nbsp;set (Table 2). Supplementary\u0026nbsp;Tables\u0026nbsp;S2 and S3 show that the most important predictor variables in \u003cem\u003esLR\u003c/em\u003e to explain at least 90% of model variance were different in yACS and the global cohort. In yACS subjects, the \u003cem\u003eodds ratios\u003c/em\u003e for MACE compared to the global cohort were higher for blood glycemia, prior chronic kidney disease (CKD), Killip class and syncope at ACS onset and lower for myocardial blush grade (MBG) and presence of dyskinesia (any wall). Increasing duration of catheterization (cath) was highly associated with\u0026nbsp;MACEs\u0026nbsp;in yACS and linked to intraprocedural complications such as coronary artery dissections (3.02% of yACS) and coronary rupture (0.1% of yACS). Late catheterization (12 h\u0026nbsp;after\u0026nbsp;symptom\u0026nbsp;onset for STEMI and\u0026nbsp;24 h\u0026nbsp;after\u0026nbsp;symptom\u0026nbsp;onset for UA/NSTEMI) was also an independent risk factor for MACE only in yACS.\u003c/p\u003e\n\u003cp\u003eThe \u003cem\u003esLR\u003c/em\u003e model trained in the yACS cohort performed as well as \u003cem\u003ethe random forest\u003c/em\u003e and \u003cem\u003eXGBoost\u003c/em\u003e algorithms (\u003cem\u003ep\u003c/em\u003e for \u003cem\u003eC-\u003c/em\u003e\u003cem\u003estatistic\u003c/em\u003e comparisons of 0.68 and 0.77, respectively), and \u003cem\u003esLR\u003c/em\u003e was superior to the GRACE score-based model (\u003cem\u003ep\u003c/em\u003e=0.031) (Table 2). However, with a \u003cem\u003eC-\u003c/em\u003e\u003cem\u003estatistic\u003c/em\u003e of 0.92 (95% CI 0.89-0.95), the \u003cem\u003eTabNet\u003c/em\u003e algorithm trained in the yACS cohort was superior to sLR (\u003cem\u003ep\u003c/em\u003e for \u003cem\u003eC-\u003c/em\u003e\u003cem\u003estatistic\u003c/em\u003e comparisons \u0026lt;0.001) and superior to \u003cem\u003eTabNet\u003c/em\u003e trained in the global cohort (C-statistic of 0.90 (95% CI 0.88-0.92), \u003cem\u003ep\u003c/em\u003e for \u003cem\u003eC-\u003c/em\u003e\u003cem\u003estatistic\u003c/em\u003e comparisons 0.011).\u003c/p\u003e\n\u003cp\u003eAs shown in Figure 3, 28 variables are included in the \u003cem\u003eTabNet\u003c/em\u003e algorithm for the global cohort, and 20 are responsible for 91% of the model variance. In the yACS cohort, 24 variables were recruited, and 20 were responsible for 93% of the model variance. Among the top predictor variables that explain at least 90% of the model variance, Figure 1 shows very different patterns for the \u003cem\u003eTabNet\u003c/em\u003e algorithm trained in yACS subjects and \u003cem\u003eTabNet\u003c/em\u003e trained in the global cohort. Risk models share three variables in common (blood glycemia, BMI and right ventricular akinesia), and the algorithm trained in yACS contains characteristics related to microvascular thrombosis and intraprocedural complications of catheterization.\u0026nbsp;\u003c/p\u003e\n\u003ch2\u003e\u003cem\u003eLong-term MACE with competing risks\u003c/em\u003e\u003c/h2\u003e\n\u003cp\u003eHere, the clinical question is whether LTW\u003csub\u003em\u003c/sub\u003e would\u0026nbsp;be\u0026nbsp;better\u0026nbsp;suited\u0026nbsp;for \u003cem\u003eglobal\u003c/em\u003e \u003cem\u003efollow-up\u003c/em\u003e (observation window of\u0026nbsp;the\u0026nbsp;first 48 hours) or whether it would perform better for\u0026nbsp;\u003cem\u003epostdischarge\u003c/em\u003e (from index ACS) risk prediction (observation window including \u003cem\u003ein-hospital\u003c/em\u003e stay).\u0026nbsp;\u003cem\u003ePostdischarge\u003c/em\u003e\u003cem\u003e\u0026nbsp;\u003c/em\u003emodels had 47\u0026nbsp;noncardiovascular\u0026nbsp;deaths and 454\u0026nbsp;MACEs, while the \u003cem\u003eglobal follow-up\u003c/em\u003e models had 92\u0026nbsp;noncardiovascular\u0026nbsp;deaths and 631\u0026nbsp;MACEs\u0026nbsp;among individuals with yACS.\u003c/p\u003e\n\u003cp\u003eIn\u0026nbsp;\u003cem\u003ethe postdischarge\u003c/em\u003e models (available in Table 3), \u003cem\u003eCS-Cox\u003c/em\u003e and \u003cem\u003eFine-Gray\u003c/em\u003e yielded the lowest C\u003cem\u003e\u003csup\u003etd\u003c/sup\u003e\u003c/em\u003e indexes in\u0026nbsp;the\u0026nbsp;test set, 0.602 (95% CI 0.556-0.649) and 0.612 (95% CI 0.564-0.663), while \u003cem\u003eDMGP\u003c/em\u003e and \u003cem\u003eDeepHit\u003c/em\u003e reached 0.685 (95% CI 0.639-0.725) and 0.722 (95% CI 0.678-0.760), respectively. \u003cem\u003eGlobal follow-up\u003c/em\u003e models (Table 3) produced generally lower concordance indexes in\u0026nbsp;the\u0026nbsp;test set, 0.597 (95% CI 0.552-0.643), 0.601 (95% CI 0.559-0.660), 0.687 (95% CI 0.640-0.728) and 0.681 (95% CI 0.654-0.703) for \u003cem\u003eCS-Cox\u003c/em\u003e, \u003cem\u003eFine-Gray\u003c/em\u003e, \u003cem\u003eDMGP\u003c/em\u003e and \u003cem\u003eDeepHit,\u0026nbsp;\u003c/em\u003erespectively. \u003cem\u003eDeepHit\u003c/em\u003e in\u0026nbsp;\u003cem\u003ethe postdischarge\u003c/em\u003e horizon yielded the highest C\u003cem\u003e\u003csup\u003etd\u003c/sup\u003e\u003c/em\u003e index and the lowest IBS (0.0579), suggesting\u0026nbsp;the\u0026nbsp;highest accuracy. The\u0026nbsp;CIFs\u0026nbsp;of 12 random yACS individuals are depicted in Figure 2. C\u003cem\u003e\u003csup\u003etd\u003c/sup\u003e\u003c/em\u003e-indexes for the global cohort were similar to yACS both in\u0026nbsp;\u003cem\u003epostdischarge\u003c/em\u003e and \u003cem\u003eglobal follow-up\u003c/em\u003e horizons.\u003c/p\u003e\n\u003cp\u003eAmong the algorithms, only \u003cem\u003eCS-Cox\u003c/em\u003e is easily interpretable;\u0026nbsp;therefore,\u0026nbsp;it was used to acquire a glance at the risk components for long-term MACE. As seen in Supplementary\u0026nbsp;Tables\u0026nbsp;S3 and S4, most short- and long-term MACE predictors differ significantly. Only\u0026nbsp;the\u0026nbsp;Killip class and prior CKD stood as predictors in both long-term \u003cem\u003eCS-Cox\u003c/em\u003e and short-term \u003cem\u003esLR\u003c/em\u003e. The \u003cem\u003eCS-Cox\u003c/em\u003e model in yACS individuals showed that drugs prescribed at discharge from index ACS,\u0026nbsp;such as anticoagulants, furosemide and ticagrelor/prasugrel,\u0026nbsp;are independently associated with\u0026nbsp;MACEs. The atherosclerotic burden (Synthax score) and low ejection fraction were also linked to\u0026nbsp;MACEs, but in yACS individuals,\u0026nbsp;STEMI in index ACS showed reduced long-term risk compared to NSTEMI, and CABG as a treatment of index ACS was also associated with lower risk compared to PCI. Finally, the occurrence of non-fatal MACE during index ACS hospitalization was associated with\u0026nbsp;an\u0026nbsp;increased risk of long-term MACE. We did not observe differences in risk components for the global cohort and yACS individuals in\u0026nbsp;the\u0026nbsp;\u003cem\u003eCS-Cox\u003c/em\u003e model.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we found that individuals with yACS present different demographic characteristics and susceptibility to risk factors for MACEs compared to older subjects. We also identified that risk prediction models are optimized by using a compound strategy: (i) specific risk prediction rules for yACS individuals rather than targeted to the overall population; (ii) short-term predictions are highly efficient; and (iii) long-term prediction models should incorporate competing events and should be optimized by including \u003cem\u003ein-hospital\u003c/em\u003e clinical data in the observation window. Roughly, the best model using this compound strategy led to the detection of 80% of events, compared to 68% by using general rules.\u003c/p\u003e \u003cp\u003eAs mentioned, risk prediction rules are improved by the optimal selection of observation windows. This issue was recently reviewed by Lauritsen et al.\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e and suggested that temporal \u003cem\u003eframing structures\u003c/em\u003e are critical for successful risk prediction. In models for predicting sepsis, the authors suggested not only implementing optimal selection of observation/prediction windows but also including a sequential evaluation by using predictions made until the current timestep\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. Indeed, Wong et al. suggested that a hospitalization-level risk score for sepsis based on the entire trajectory of predictions may enable more realistic evaluations\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e. However, in clinical cardiology, risk scores are typically less dynamic and employ temporal \u003cem\u003eframing\u003c/em\u003e suboptimally. By setting a short-term endpoint, we could identify important predictors of \u003cem\u003ein-hospital\u003c/em\u003e MACE with a set of data gathered from the first two days of index ACS onset. In parallel, long-term risk prediction taking into account competing risks was optimized by including \u003cem\u003epredischarge\u003c/em\u003e information, including prescription at discharge and \u003cem\u003ein-hospital\u003c/em\u003e clinical events.\u003c/p\u003e \u003cp\u003eAnother argument in favor of splitting two prediction windows is that we showed large differences between key predictors of MACE in the short term and the long term. In short-term models, the most important predictors of MACEs are symptoms at ACS presentation, microvascular thrombosis and intraprocedural complications of catheterization. Instead, in long-term models, the top predictors of worse clinical outcomes are mostly related to \u003cem\u003ein-hospital\u003c/em\u003e outcomes, discharge medications, past medical history and severity of coronary artery lesions. In addition, splitting two prediction windows permits a flexible and dynamic way of dealing with clinical problems\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eIt is important to mention that binary classification can provide predictions for a predetermined duration (e.g., \u003cem\u003ein-hospital\u003c/em\u003e stay), useful for short-term outcomes where time to event is not an issue. If transported to long-term risk modeling, binary classification typically ignores time censoring and may increase the risk for false negatives. Hence, in clinical problems with a substantial amount of censoring, the use of survival models tends to be advantageous\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e. On the other hand, if censoring bias is not accounted for or the context can neutralize censoring, binary classification tends to maximize accuracy compared to survival models\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e. Therefore, the way to better fit a real-world scenario was to combine short-term classification with long-term survival.\u003c/p\u003e \u003cp\u003eFor \u003cem\u003ein-hospital\u003c/em\u003e MACE prediction, \u003cem\u003eTabNet\u003c/em\u003e yielded the best results. The algorithm has been recently described and couples a deep neural network architecture and gradient descent-based optimization designed specifically for tabular data\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e. Together with the great predictive capacity, it also enables interpretability. Although no causality can be attributed to top predictors, they are consistent with the most prevalent risk factors for MACEs among yACS\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e,\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e. As observed by others\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e, we observed that variables of interest for predicting MACEs in individuals with premature ACS differed from the top predictors among the global cohort and older subjects.\u003c/p\u003e \u003cp\u003eAmong the long-term models, \u003cem\u003eDeepHit\u003c/em\u003e was the most accurate. \u003cem\u003eDeepHit\u003c/em\u003e is a multitask network that makes no linear assumptions during the predictive process, allowing for the possibility that the relationship between covariates and risks changes over time\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e. Although such architecture improves predictive ability and flexibility to deal with competing risks compared to \u003cem\u003eCS-Cox\u003c/em\u003e and \u003cem\u003eFine-Gray\u003c/em\u003e models, it is not possible to interpret which variables are recruited at each step. However, among the long-term predictors of MACEs using \u003cem\u003eCS-Cox\u003c/em\u003e, we identified that yACS may be at higher risk when prescribed at discharge drugs such as ticagrelor or prasugrel than clopidogrel. These observations contradict the findings from major clinical trials such as PLATO\u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e and TRITON-TIMI-38\u003csup\u003e19\u003c/sup\u003e but should be explored in other real-world scenarios with appropriate techniques for neutralizing any potential selection bias.\u003c/p\u003e \u003cp\u003eThere are limitations in this study that should be acknowledged. First, the observational and retrospective design of this study limits any potential causal conclusions. Second, the definition of yACS is not consensus; while some consider an age threshold of 55 years old, others consider 50 or 45 years old\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e,\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e,\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. Third, our models were trained in a relatively small cohort. Although the B-CaRe:QCO yACS cohort is among the largest cohorts of yACS, some algorithms, such as \u003cem\u003eDeepHit\u003c/em\u003e, \u003cem\u003eDMGP\u003c/em\u003e, and \u003cem\u003eTabNet\u003c/em\u003e, were originally developed in datasets of \u0026gt;\u0026thinsp;10,000 individuals\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e,\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e,\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e. Our results suggest that these algorithms also perform well in smaller datasets, and we did our best to maximize external validity by using cross-validation and resampling techniques. The main advantage of our cohort is that we systematically included all subjects admitted due to ACS in public hospitals from Bras\u0026iacute;lia (Brazil) who underwent coronarography up to 48 hours after hospital admission between January 2011 and February 2020.\u003c/p\u003e \u003cp\u003eIn summary, we found that individuals with premature ACS share considerable morbidity and show unique epidemiologic features compared to those of older subjects. In this study, we also identified that risk prediction models are optimized by using specific risk prediction rules for yACS individuals in two windows: a short-term window and a long-term window that incorporate competing events and \u003cem\u003ein-hospital\u003c/em\u003e clinical data within the observation window. It is critical to better understand risk factors within this subgroup to allow public health initiatives that mitigate the economic burden aroused by yACS\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e,\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. Risk prediction-enhanced clinical care could turn into a framework for intensified clinical surveillance in individuals predicted to be high risk\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003eStudy design and participants\u003c/h2\u003e \u003cp\u003eThe set of individuals was selected from the \u003cem\u003eB-CaRe:QCO\u003c/em\u003e (\u003cem\u003eBrasilia Cardiovascular Registry for Quality of Care and Outcomes\u003c/em\u003e), a retrospective registry of 6341 subjects with ACS (n\u0026thinsp;=\u0026thinsp;2242 with yACS). The \u003cem\u003eB-CaRe:QCO\u003c/em\u003e study included consecutive individuals admitted to public hospitals in Bras\u0026iacute;lia (DF) with ACS who underwent coronarography up to 48 h after hospital admission from January 2011 to February 2020. At that time, all coronarographies were carried out in \u003cem\u003eHospital de Base\u003c/em\u003e (Bras\u0026iacute;lia-DF, Brazil) and \u003cem\u003eInstituto de Cardiologia\u003c/em\u003e (Bras\u0026iacute;lia-DF, Brazil). We excluded 17 individuals who died within the first 48 hours.\u003c/p\u003e \u003cp\u003eEnrolled subjects experienced therapies based on guidelines for the treatment of ACS\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e. Attending physicians made all therapeutic decisions and were blinded to the study evaluations. Most individuals admitted due to STEMI (n\u0026thinsp;=\u0026thinsp;1659 with premature ST-elevation myocardial infarction [STEMI]) were treated by primary percutaneous coronary intervention (pPCI) or pharmacoinvasive strategy.\u003c/p\u003e \u003cp\u003eThe methods were performed in accordance with relevant guidelines and regulations, and approved by the Institutional Ethics Review Board from \u003cem\u003eInstituto de Gest\u0026atilde;o Estrat\u0026eacute;gica em Sa\u0026uacute;de do Distrito Federal\u003c/em\u003e (IGESDF) (study protocol approval number [CAAE] 28530919.0.1001.8153).\u003c/p\u003e \u003cp\u003eFor predicting \u003cem\u003ein-hospital\u003c/em\u003e MACE (defined as cardiovascular deaths or recurrent ACS) occurring 48 h after hospital admission, the observation window comprised the first 48 h after hospital admission. The yACS dataset was divided into a training/validation set (70%, n\u0026thinsp;=\u0026thinsp;1569) and a test set (30%, n\u0026thinsp;=\u0026thinsp;673). Short-term models (STW\u003csub\u003em\u003c/sub\u003e) were trained and validated in a 5-fold cross-validation framework with upsampling to mitigate outcome imbalance. STW\u003csub\u003em\u003c/sub\u003e was then evaluated in the test set.\u003c/p\u003e \u003cp\u003eTo predict long-term outcomes with competing risks (noncardiovascular deaths vs MACE), two contexts were evaluated: (i) \u003cem\u003epostdischarge\u003c/em\u003e, where an observation window included the whole period of index hospitalization (mean of 5\u0026thinsp;\u0026plusmn;\u0026thinsp;2 days) and the outcomes were observed from hospital discharge to the end of follow-up (median of 6.67 years); (ii) \u003cem\u003eglobal follow-up\u003c/em\u003e, where the observation window included only the first 48 h and the outcomes observation period began at 48 h and extended to the end of follow-up. A training/validation set (n\u0026thinsp;=\u0026thinsp;1513) and test set (n\u0026thinsp;=\u0026thinsp;648) included individuals alive at discharge and were used to train and validate long-term window models (LTW\u003csub\u003em\u003c/sub\u003e). LTW\u003csub\u003em\u003c/sub\u003e was repeated over five cross-validation folds and then assessed in the test set.\u003c/p\u003e \u003cp\u003eTo better understand model accuracy and differences in key predictors for short-term MACE between the yACS and older subjects, we also created models using the \u003cem\u003eglobal cohort\u003c/em\u003e (n\u0026thinsp;=\u0026thinsp;6341) by splitting a training/validation set (n\u0026thinsp;=\u0026thinsp;4439) and a test set including only 673 individuals in the yACS test set (remaining 1229 individuals older than 55 years were not included in the test set to prevent sampling imbalance). Again, we used 5-fold cross-validation with upsampling for STW\u003csub\u003em\u003c/sub\u003e and evaluated the model in the yACS test set (n\u0026thinsp;=\u0026thinsp;673).\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e \u003ch2\u003eClinical definitions and outcome assessment\u003c/h2\u003e \u003cp\u003eCurrent smokers were defined as those who had smoked at least 100 cigarettes during their lifetime and were smoking at least one year before ACS onset, according to the National Health Interview Survey (NHIS) definition\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e. Ex-smoking status was defined as smoking cessation for at least the last 6 months. Diabetes was defined as the use of antidiabetic medications, prior diagnosis of diabetes, or glycosylated hemoglobin (HbA1c)\u0026thinsp;\u0026ge;\u0026thinsp;6.5% at hospital admission. Patients were considered hypertensive if they were taking any antihypertensive medication or presented systolic blood pressure (SBP)\u0026thinsp;\u0026ge;\u0026thinsp;140 mm Hg or diastolic blood pressure (DBP)\u0026thinsp;\u0026ge;\u0026thinsp;90 mmHg. The anthropometric measurements obtained were body weight (kg), height (m), and waist circumference (cm). The Killip class and GRACE scores for \u003cem\u003ein-hospital\u003c/em\u003e MACEs were evaluated in all enrolled patients\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eClinical outcomes were assessed by checking electronic health records (EHRs). Information about the cause of death and clinical events was obtained from the death certificate or medical records. The following adverse cardiac events for both STW\u003csub\u003em\u003c/sub\u003e and LTW\u003csub\u003em\u003c/sub\u003e were considered: cardiovascular deaths and recurrent ACS (MACE). For STW\u003csub\u003em\u003c/sub\u003e, those who had any event during follow-up were marked as 1, and those who did not were coded as 0. For LTW\u003csub\u003em,\u003c/sub\u003e we considered a competing event approach in survival analyses, i.e., individuals were followed until their deaths, the occurrence of recurrent ischemic events or the end of follow-up (last visit to the outpatient clinic registered in EHRs). Reinfarction was defined as the occurrence of new ischemic symptoms during the first 28 days after index MI associated with a\u0026thinsp;\u0026gt;\u0026thinsp;20% increase in cTn levels after a 3-to-6-hour interval from symptoms\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e \u003ch2\u003eModels and variable selection\u003c/h2\u003e \u003cp\u003eA domain-knowledge-driven approach was first used to select variables. From 186 variables at baseline, we excluded variables with no potential causal link with the outcomes and included those proven as predictors in previous models, leaving the remaining 108 variables. Variables were included only if they were unambiguous in their interpretation and recorded in a structured (numeric/binary) format.\u003c/p\u003e \u003cp\u003eAfter this, a data-driven approach took place and consisted of an automated process based on actual data and the relevance of each variable to a specific outcome\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e. For most of the STW\u003csub\u003em\u003c/sub\u003e and LTW\u003csub\u003em\u003c/sub\u003e, we used a fully automated process incorporated into the algorithms. When selection could not be performed automatically, we followed guidelines as proposed by Belsley et al\u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e: in the case of high correlation between variables (partial R\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;\u0026ge;\u0026thinsp;0.5 in univariate regression with MACE[=\u0026thinsp;1] as the dependent variable or variance inflation factor [VIF]\u0026thinsp;\u0026gt;\u0026thinsp;10), we dropped the variables with lower R\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. Information-gain ranking was used to evaluate the worth of each variable by measuring the entropy gain with respect to the outcome, followed by ranking the attributes by their individual evaluations. Considering the \u003cem\u003etradeoffs\u003c/em\u003e between the cost of information and information gain, only attributes resulting in information gain higher than 0.01 were subsequently used in STW\u003csub\u003em\u003c/sub\u003e and LTW\u003csub\u003em\u003c/sub\u003e. Variable selection was performed in the training/validation dataset.\u003c/p\u003e \u003cp\u003eMissing values (MVs) were relatively rare (2.7% of B-CaRe:QCO data). We handled MVs with multiple imputations directly in the training/validation dataset by using boosted trees. Only a few variables showed MV frequencies\u0026thinsp;\u0026ge;\u0026thinsp;10% (plasma TSH, free T4 and urea). Imputation using boosted trees fills each column by treating it as a regression problem. We did not impute missing values for the outcomes.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003ePredictive algorithms\u003c/h2\u003e \u003cp\u003eFor predicting short-term outcomes, we used \u003cem\u003eXGBoost\u003c/em\u003e\u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e, \u003cem\u003erandom forests\u003c/em\u003e\u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e, and \u003cem\u003eTabNet\u003c/em\u003e\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e. \u003cem\u003eRandom forests\u003c/em\u003e, based on decision trees, rank variable importance on the selection frequency of the variable as a decision node and generally show good performance for classification problems in tabular data with a single outcome\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. \u003cem\u003eXGBoost\u003c/em\u003e is also based on decision trees and uses gradient descent-based optimization\u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e. \u003cem\u003eTabNet\u003c/em\u003e has an interpretable canonical deep tabular data learning architecture, merging both deep learning and gradient descent-based optimization. The observation window was considered the first 2 days upon hospital admission and encompassed past medical history, emergency room data and coronarography. We compared models with the benchmark GRACE score\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e, recalibrated using regression coefficients of risk factors derived from logistic regressions (LR) as described elsewhere\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e (details in below).\u003c/p\u003e \u003cp\u003eFor long-term outcomes, we used the following survival algorithms with competing risks: cause-specific Cox-proportional hazards model (\u003cem\u003eCS-Cox\u003c/em\u003e)\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e, Fine-Gray proportional subdistribution hazards model (\u003cem\u003eFine-Gray\u003c/em\u003e)\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e, deep multitask Gaussian process (\u003cem\u003eDMGP\u003c/em\u003e)\u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e, and \u003cem\u003eDeepHit\u003c/em\u003e\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e. \u003cem\u003eCS-Cox\u003c/em\u003e and \u003cem\u003eFine-Gray\u003c/em\u003e assume linear proportional hazards, \u003cem\u003eDMGP\u003c/em\u003e assumes the underlying stochastic process to follow the Gaussian process, and \u003cem\u003eDeepHit\u003c/em\u003e employs a network architecture that makes no assumptions about the relationship between predictors and outcomes.\u003c/p\u003e \u003cp\u003eEach model\u0026rsquo;s hyperparameters were determined using the grid search method\u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e and 5-fold cross-validation for STW\u003csub\u003em\u003c/sub\u003e and LTW\u003csub\u003em\u003c/sub\u003e. STW\u003csub\u003em\u003c/sub\u003e were generated with upsampling to mitigate outcome imbalance. Performance in the validation set is reported as the mean of 5-folds. A full description of variable selection, hyperparameters and model architectures can be found below.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section3\"\u003e \u003ch2\u003eModel Development Process\u003c/h2\u003e \u003cp\u003eTo develop the prognostic models, B-CaRe:QCO data were extracted into a labelled dataset containing the independent variables (using the patients\u0026rsquo; clinical records at their baseline dates or during index hospitalization) and all dependent variables (occurrence of a composite endpoint of death due to cardiovascular causes and recurrent ACS following the baseline date).\u003c/p\u003e \u003cp\u003eWe implemented a grid search for the hyperparameter optimization using the method reported by Bergstra and Bengio\u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e. This requires the operator to specify a range of values for each hyperparameter, and all possible combinations of the hyperparameters are investigated, with the combination corresponding to the highest cross-validation performance metric (in this case, maximization of the \u003cem\u003eC-statistics\u003c/em\u003e being chosen for the final model). The justification for selecting the hyperparameters that maximise the \u003cem\u003eC-statistics\u003c/em\u003e is that this is less affected when the labelled data are unbalanced compared to using accuracy as a metric. When the classes are unbalanced, it is also a common strategy to oversample the rare label data and undersample the common label data, as many machine learning models can be sensitive to unbalanced data. Below, we describe in further detail the algorithms used.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section3\"\u003e \u003ch2\u003eShort-term predictive algorithms for classification\u003c/h2\u003e \u003cp\u003e \u003cem\u003eRandom forests\u003c/em\u003e. For the hyperparameter grid search, we investigated ntree\u0026thinsp;=\u0026thinsp;50, 150, and 350; mtry from 5 up to the maximum number of variables in increments of 5; max depth\u0026thinsp;=\u0026thinsp;2, 4, 6, 8, and 10; and row samples of 90%, 95% and 100%. The chosen (optimal) random forest model had the following hyperparameters: ntree\u0026thinsp;=\u0026thinsp;350, mtry\u0026thinsp;=\u0026thinsp;25, max depth\u0026thinsp;=\u0026thinsp;5 (up to 5 variable interactions were used by the model) and row sample fraction of 0.95 (95% of the data points were used to train each tree).\u003c/p\u003e \u003cp\u003e \u003cem\u003eXGboost.\u003c/em\u003e The grid search for the hyperparameters investigated in our models were ntree\u0026thinsp;=\u0026thinsp;25, 50, 75 and 100; max depth\u0026thinsp;=\u0026thinsp;2, 3, 4, 6 and 8; and the minimum observations per node was 5, 10, 20, and 40. The gradient boosting machine model was chosen to have a Bernoulli distribution, and the chosen model had the following hyperparameters: ntree\u0026thinsp;=\u0026thinsp;50, max depth\u0026thinsp;=\u0026thinsp;3 (up to 3 variable interactions were used by the model), and the minimum number of observations per node was 10. \u003cem\u003eXGBoost\u003c/em\u003e was implemented in Python.\u003c/p\u003e \u003cp\u003e \u003cem\u003eTabNet.\u003c/em\u003e We used a canonical deep neural network (DNN) architecture for tabular data described by Arik et al\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e. Briefly, \u003cem\u003eTabNet\u003c/em\u003e is trained using gradient descent-based optimization and uses sequential attention to choose which features to reason from at each decision step, enabling (i) interpretability, (ii) more accurate and faster learning and (iii) flexible integration into end-to-end learning. Through sparse and instancewise selection (\u003cem\u003esparsemax\u003c/em\u003e is used for normalization of the coefficients) of features with the highest impact on outcomes, the learning capacity of a decision step is not wasted on irrelevant ones, and thus the model becomes more parameter efficient. \u003cem\u003eTabNet\u003c/em\u003e also constructs a sequential multistep architecture, where each step contributes to a portion of the decision based on the selected features, improves the learning capacity via nonlinear processing of the selected features, and mimics ensembling via higher dimensions. The \u003cem\u003eTabNet\u003c/em\u003e encoder is composed of a feature transformer, an attentive transformer and feature masking. A split block divides the processed representation to be used by the attentive transformer of the subsequent step as well as for the overall output. For each step, the feature selection mask provides interpretable information about the model\u0026rsquo;s functionality, and the masks can be aggregated to obtain global feature important attributions. The \u003cem\u003eTabNet\u003c/em\u003e decoder is composed of a feature transformer block at each step. Each feature transformer block is composed of a 4-layer network, where 2 are shared across all decision steps and 2 are decision step-dependent. Each layer is composed of a fully connected (FC) layer, ghost batch normalization (BN) and gated linear unit (GLU) nonlinearity. We used standard classification (\u003cem\u003esoftmax\u003c/em\u003e cross entropy) loss functions, and we trained the model until convergence using unsupervised pretraining. The final \u003cem\u003eTabNet\u003c/em\u003e model was implemented in \u003cem\u003ea PyTorch\u003c/em\u003e environment and had the following configuration: \u003cem\u003eAdam\u003c/em\u003e optimizer with a learning rate of 0.02 and a decay rate of 0.9 every 10 interactions, \u003cem\u003eGlorot\u003c/em\u003e uniform initialization, batch size of 256, Max epoch 1000, workers at zero, momentum of 0.9, N\u003csub\u003e\u003cem\u003esteps\u003c/em\u003e\u003c/sub\u003e=8, γ\u0026thinsp;=\u0026thinsp;2.0, and weight at 1 (automated sampling).\u003c/p\u003e \u003cp\u003e \u003cem\u003eLogistic regression models\u003c/em\u003e. We built a series of stepwise logistic regression models to predict \u003cem\u003ein-hospital\u003c/em\u003e MACEs.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section3\"\u003e \u003ch2\u003eLong-term predictive models \u0026ndash; survival with competing risks\u003c/h2\u003e \u003cp\u003e \u003cem\u003eCause-specific Cox-proportional hazards model (Cox) and Fine-Gray proportional subdistribution hazards model (Fine-Gray).\u003c/em\u003e The Cox model relates the covariates to the hazard function of the outcome of interest and not directly to the survival times themselves. The covariates have a relative effect on the hazard function because of the use of the logarithmic transformation, and the regression coefficients are interpreted as log-hazard ratios. The hazard ratio is equal to the exponential of the associated regression coefficient\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e. Competing risks imply that a subject can experience one of a set of different events or outcomes. In this case, two different types of hazard functions are of interest: the cause-specific hazard function and the subdistribution hazard function. The cause-specific hazard function indicates the instantaneous rate of occurrence of the \u003cem\u003ek\u003c/em\u003e\u003csup\u003eth\u003c/sup\u003e event in subjects who are currently event free (i.e., in subjects who have not yet experienced any of the different types of events). Considering two types of events, death attributable to cardiovascular causes and death attributable to noncardiovascular causes, the cause-specific hazard of cardiovascular death denotes the instantaneous rate of cardiovascular death in subjects who are still alive. It denotes the instantaneous risk of failure from the \u003cem\u003ek\u003c/em\u003e\u003csup\u003e\u003cem\u003eth\u003c/em\u003e\u003c/sup\u003e event in subjects who have not yet experienced an event of type \u003cem\u003ek\u003c/em\u003e. There is a distinct cause-specific hazard function for each of the distinct types of events and a distinct subdistribution hazard function for each of the distinct types of events. In settings in which competing risks are present, two different hazard regression models are available: modeling the cause-specific hazard and modeling the subdistribution hazard function. The second model has also been described as a \u003cem\u003ecumulative incidence function\u003c/em\u003e (CIF) regression model, which means that the subdistribution hazard model allows one to estimate the effect of covariates on the cumulative incidence function for the event of interest. However, it is recommended to use the Fine-Gray (FG) subdistribution hazard model when the focus is on estimating incidence or predicting prognosis in the presence of competing risks, since this model generally shows better accuracy than the \u003cem\u003eCox\u003c/em\u003e model. The \u003cem\u003e(cause-specific) cumulative incidence function\u003c/em\u003e (CIF) expresses the probability that a particular event \u003cem\u003ek*\u003c/em\u003e occurs on or before time \u003cem\u003et\u0026lowast;\u003c/em\u003e conditional on covariates \u003cem\u003ex*\u003c/em\u003e. Since \u003cem\u003etrue\u003c/em\u003e CIF is not known, the model utilizes \u003cem\u003eestimated\u003c/em\u003e CIF to compare the risk of events occurring and to assess how models discriminate across cause-specific risks among patients. Model performance was calculated by using the time-dependent concordance index C\u003csup\u003e\u003cem\u003etd\u003c/em\u003e \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u003c/sup\u003e (C\u003csup\u003e\u003cem\u003etd\u003c/em\u003e\u003c/sup\u003e-index), which measures the extent to which the ordering of actual survival times of pairs agrees with the ordering of their predicted risk. \u003cem\u003eCox\u003c/em\u003e and \u003cem\u003eFG\u003c/em\u003e benchmarks were run using the R libraries \u003cem\u003esurvival\u003c/em\u003e and \u003cem\u003ecmprsk\u003c/em\u003e. We estimated the time-dependent \u003cem\u003eC\u003c/em\u003e\u003csup\u003e\u003cem\u003etd\u003c/em\u003e\u003c/sup\u003e \u003cem\u003eindex\u003c/em\u003e for the survival analysis methods under consideration using the function \u003cem\u003ecindex\u003c/em\u003e of the R package \u003cem\u003epec\u003c/em\u003e.\u003c/p\u003e \u003cp\u003e \u003cem\u003eA deep multitask Gaussian process (DMGP)\u003c/em\u003e \u003csup\u003e\u003cem\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/em\u003e\u003c/sup\u003e is a nonparametric Bayesian model for survival analysis that relies on a conception of the competing risks problem as a multitask learning problem; i.e., it models the cause-specific survival times as the outputs of a random vector-valued function, the inputs to which are the patients\u0026rsquo; covariates. This allows the model to learn a \u0026ldquo;shared representation\u0026rdquo; of survival times with respect to multiple related comorbidities. Inference of patient-specific posterior survival distribution is conducted via a variational Bayes algorithm. By using \u003cem\u003einducing variables\u003c/em\u003e to derive a variational lower bound on the marginal likelihood of the observed time-to-event data, which is maximized using the adaptive moment estimation algorithm (\u003cem\u003eAdam\u003c/em\u003e). Hyperparameters Θ\u003csub\u003eZ\u003c/sub\u003e and Θ\u003csub\u003eT\u003c/sub\u003e were tuned using the \u003cem\u003eoffline\u003c/em\u003e B-CaRe:QCO dataset, and for any out-of-sample patient with all covariates, \u003cem\u003eDMGP\u003c/em\u003e evaluates posterior probability density by direct Monte Carlo sampling. Hyperparameters were calibrated by maximizing the marginal likelihood of posterior probability density. \u003cem\u003eDMGP\u003c/em\u003e was implemented in Python.\u003c/p\u003e \u003cp\u003e \u003cem\u003eDeepHit\u003c/em\u003e trains a neural network to learn the estimated joint distribution of survival time and event while capturing the right-censored nature inherent in survival data \u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e. The network is trained by using a loss function that exploits both survival times and relative risks. \u003cem\u003eDeepHit\u003c/em\u003e makes no assumptions about the underlying stochastic process and allows for the possibility that the relationship between covariates and risks changes over time. \u003cem\u003eDeepHit\u003c/em\u003e is a multitask network that consists of a shared subnetwork and \u003cem\u003eK\u003c/em\u003e cause-specific subnetworks, differing from that of a conventional multitask network in two ways: (i) it utilizes a single \u003cem\u003esoftmax\u003c/em\u003e layer as the output layer of \u003cem\u003eDeepHit\u003c/em\u003e to ensure that the network learns the joint distribution of \u003cem\u003eK\u003c/em\u003e competing events, not the marginal distributions of each event; (ii) it keeps a residual connection from the input covariates into the input of each cause-specific subnetwork. To train \u003cem\u003eDeepHit\u003c/em\u003e, a total loss function \u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003eTotal\u003c/em\u003e\u003c/sub\u003e is specifically designed to handle censored data. This loss function is the sum of two \u003cem\u003eterms L\u003c/em\u003e\u003csub\u003e\u003cem\u003eTotal\u003c/em\u003e\u003c/sub\u003e \u003cem\u003e= L\u003c/em\u003e\u003csub\u003e\u003cem\u003e1\u003c/em\u003e\u003c/sub\u003e \u003cem\u003e+ L\u003c/em\u003e\u003csub\u003e\u003cem\u003e2\u003c/em\u003e\u003c/sub\u003e; \u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003e1\u003c/em\u003e\u003c/sub\u003e is the log-likelihood of the joint distribution of the first hitting time and event; \u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003e2\u003c/em\u003e\u003c/sub\u003e incorporates a combination of cause-specific ranking loss functions that adapts the idea of concordance. The hyperparameters for \u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003eTotal\u003c/em\u003e\u003c/sub\u003e were selected based on the discriminative performance on the validation set. Early stopping was performed based on the total loss. \u003cem\u003eDeepHit\u003c/em\u003e is a 4-layer network consisting of 1 fully connected layer for the shared subnetwork and 2 fully connected layers for each cause-specific subnetwork and a \u003cem\u003esoftmax\u003c/em\u003e layer as the output layer. For hidden layers, the number of nodes was set as 3, 5, and 3 times the covariate dimension for layers 1, 2, and 3, respectively, with the \u003cem\u003eReLu\u003c/em\u003e activation function. The network was trained by backpropagation via the \u003cem\u003eAdam\u003c/em\u003e optimizer with a batch size of 50 and a learning rate of 0.0001. A dropout probability of 0.6 and Xavier initialization were applied for all layers. \u003cem\u003eDeepHit\u003c/em\u003e was implemented in a \u003cem\u003eTensorFlow\u003c/em\u003e environment in Python.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eStatistical analysis\u003c/h2\u003e \u003cp\u003eSTW\u003csub\u003em\u003c/sub\u003e were compared using \u003cem\u003eaccuracy\u003c/em\u003e and \u003cem\u003eC-statistics\u003c/em\u003e for their performance on the test and validation datasets. We calculated the median performance and 95% confidence intervals (CIs) for the \u003cem\u003eC-statistics\u003c/em\u003e for each algorithm. We built models with the training/validation set and finally evaluated the model on the test set to estimate performance. STW\u003csub\u003em\u003c/sub\u003e was compared to the \u003cem\u003eC-statistics\u003c/em\u003e obtained by the recalibrated GRACE score\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e. LTW\u003csub\u003em\u003c/sub\u003e evaluates each individual\u0026rsquo;s cumulative incidence function (CIF), also known as the \u003cem\u003esubdistribution function\u003c/em\u003e. CIF is commonly used in settings with competing risks and refers to the probability of a particular event during follow-up. CIFs are used to evaluate the case-specific concordance, and this concept is used to derive a performance metric to compare LTW\u003csub\u003em\u003c/sub\u003e, the time-dependent concordance index C\u003csup\u003e\u003cem\u003etd\u003c/em\u003e \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u003c/sup\u003e. The C\u003csup\u003e\u003cem\u003etd\u003c/em\u003e\u003c/sup\u003e-index measures the extent to which the ordering of actual survival times of pairs agrees with the ordering of their predicted risk (further information is available in Supplemental Methods). A confidence interval for the C\u003csup\u003e\u003cem\u003etd\u003c/em\u003e\u003c/sup\u003e index is derived using the jackknife method on correlated one-sample U-statistics. The integrated Brier score (IBS) was also used as an LTW\u003csub\u003em\u003c/sub\u003e evaluation measure. Normally distributed data are presented as the mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD, and skewed data are presented as the median [interquartile range (IQR)]. Normality of distribution and variances were checked using histograms, Kolmogorov-Smirnoff test, normal probability plots and residual scatter plots. Chi-square or two-tailed \u003cem\u003et\u003c/em\u003e-tests were used for comparison of baseline data. P-values\u0026thinsp;\u0026lt;\u0026thinsp;0.05 were considered significant. Analyses were carried out using R[v4.0.1] and Python[v3.8], and the packages used are described in the Supplemental Methods.\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eConflict of Interest Disclosures:\u003c/strong\u003e There are no conflicts of interest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding/Support:\u003c/strong\u003e This work was supported by grant 2019/09068-3 from S\u0026atilde;o Paulo Research Foundation (FAPESP) and grants 437413/2018-7 and 310718/2021-0 from the Brazilian National Research Council (CNPq).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eRole of the Funder/Sponsor:\u003c/strong\u003e The funder had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDATA AVAILABILITY: \u003c/strong\u003eCodes are available at https://github.com/lsergiocarvalho/openwindowACS. All requests for raw and analyzed data and related materials, excluding programming codes, will be reviewed by the Clarity Healthcare Intelligence legal department to verify whether the request is subject to any intellectual property or confidentiality obligations. Requests for patient-related data can be considered upon request. Any data and materials that can be shared will be released via a Material Transfer Agreement.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDECLARATIONS: \u003c/strong\u003eThe authors had full access to all of the data (including statistical reports and tables) in the study and can take responsibility for the integrity of the data and the accuracy of the data analysis. They consent to the submission of the manuscript as it is. There are no financial and nonfinancial competing interests for all authors. The authors declare that they do not have a conflict of interest regarding the study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding/Support: \u003c/strong\u003eThis work was supported by grants 310718/2021-0 from the Brazilian National Research Council (CNPq), 371/2021 from FAPDF and 2019/09068-3 from FAPESP.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eRole of the Funder/Sponsor: \u003c/strong\u003eThe funder had \u003cu\u003eno role\u003c/u\u003e in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication. \u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eIRB Approval and Patient Consent\u003c/strong\u003e: The study proceedings are in accordance with the Helsinki Declaration and the study was approved by the Institutional Ethics Review Board (IRB) from \u003cem\u003eInstituto de Gest\u0026atilde;o Estrat\u0026eacute;gica do Distrito Federal\u003c/em\u003e (IGESDF) (study protocol approval number [CAAE] 28530919.0.1001.8153). Since this is a retrospective study, the IRB approved the waiver of participants informed consent as long as data is captured anonymously.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAUTHOR CONTRIBUTIONS\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eConcept and design\u003c/em\u003e: Sposito, Carvalho, Fernandez, Avila, Nogueira, Alexim, Rezende\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eAcquisition of data\u003c/em\u003e: Carvalho, Alexim, Nogueira\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eAnalysis and interpretations of data\u003c/em\u003e: Carvalho, Sposito, Fernandez, Alexim, Rezende, Reis, Rezende\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eDrafting of the manuscript\u003c/em\u003e: Carvalho\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eCritical revision of the paper for important intellectual content\u003c/em\u003e: Sposito, Avila, Fernandez, Alexim, Rezende, Nogueira, Soares, Reis\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eStatistical analysis\u003c/em\u003e: Carvalho, Reis\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eProvision of study materials or patients\u003c/em\u003e: Carvalho, Alexim, Nogueira\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eObtaining funding\u003c/em\u003e: Carvalho, Sposito\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eAdministrative, technical, or logistic support\u003c/em\u003e: Carvalho, Sposito, Avila, Alexim, Nogueira\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eSupervision\u003c/em\u003e: Sposito, Avila, Carvalho\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eArora S, Stouffer GA, Kucharska-Newton AM, et al. Twenty Year Trends and Sex Differences in Young Adults Hospitalized With Acute Myocardial Infarction. \u003cem\u003eCirculation. \u003c/em\u003e2019;139(8):1047-1056.\u003c/li\u003e\n\u003cli\u003eGupta A, Wang Y, Spertus JA, et al. Trends in acute myocardial infarction in young patients and differences by sex and race, 2001 to 2010. \u003cem\u003eJ Am Coll Cardiol. \u003c/em\u003e2014;64(4):337-345.\u003c/li\u003e\n\u003cli\u003eZeitouni M, Clare RM, Chiswell K, et al. Risk Factor Burden and Long-Term Prognosis of Patients With Premature Coronary Artery Disease. \u003cem\u003eJ Am Heart Assoc. \u003c/em\u003e2020;9(24):e017712.\u003c/li\u003e\n\u003cli\u003eDreyer RP, Raparelli V, Tsang SW, et al. Development and Validation of a Risk Prediction Model for 1-Year Readmission Among Young Adults Hospitalized for Acute Myocardial Infarction. \u003cem\u003eJ Am Heart Assoc. \u003c/em\u003e2021;10(18):e021047.\u003c/li\u003e\n\u003cli\u003ede Carvalho LSF, Gioppato S, Fernandez MD, et al. Machine Learning Improves the Identification of Individuals With Higher Morbidity and Avoidable Health Costs After Acute Coronary Syndromes. \u003cem\u003eValue Health. \u003c/em\u003e2020;23(12):1570-1579.\u003c/li\u003e\n\u003cli\u003eSagris M, Antonopoulos AS, Theofilis P, et al. Risk factors profile of young and older patients with Myocardial Infarction. \u003cem\u003eCardiovasc Res. \u003c/em\u003e2021.\u003c/li\u003e\n\u003cli\u003eYanase T, Sakakura K, Taniguchi Y, et al. Comparison of Clinical Characteristics of Acute Myocardial Infarction Between Young (\u0026lt; 55 Years) and Older (55 to \u0026lt; 70 Years) Patients. \u003cem\u003eInt Heart J. \u003c/em\u003e2021;62(1):33-41.\u003c/li\u003e\n\u003cli\u003eLauritsen SM, Thiesson B, Jorgensen MJ, et al. The Framing of machine learning risk prediction models illustrated by evaluation of sepsis in general wards. \u003cem\u003eNPJ Digit Med. \u003c/em\u003e2021;4(1):158.\u003c/li\u003e\n\u003cli\u003eKuo RN, Dong YH, Liu JP, Chang CH, Shau WY, Lai MS. Predicting healthcare utilization using a pharmacy-based metric with the WHO\u0026apos;s Anatomic Therapeutic Chemical algorithm. \u003cem\u003eMedical care. \u003c/em\u003e2011;49(11):1031-1039.\u003c/li\u003e\n\u003cli\u003eLauffenburger JC, Mahesri M, Choudhry NK. Use of Data-Driven Methods to Predict Long-term Patterns of Health Care Spending for Medicare Patients. \u003cem\u003eJAMA Netw Open. \u003c/em\u003e2020;3(10):e2020291.\u003c/li\u003e\n\u003cli\u003eLee C, Zame WR, Yoon J, Schaar Mvd. DeepHit: A Deep Learning Approach to Survival Analysis With Competing Risks. Paper presented at: XXXII Association for the Advancement of Artificial Intelligence (AAAI) Conference2018.\u003c/li\u003e\n\u003cli\u003eMcCaw ZR, Claggett BL, Tian L, et al. Practical Recommendations on Quantifying and Interpreting Treatment Effects in the Presence of Terminal Competing Risks: A Review. \u003cem\u003eJAMA Cardiol. \u003c/em\u003e2021.\u003c/li\u003e\n\u003cli\u003eArik SO, Pfister T. TabNet: Attentive Interpretable Tabular Learning. Association for the Advancement of Artificial Intelligence; 2020.\u003c/li\u003e\n\u003cli\u003eLauritsen SM, Kalor ME, Kongsgaard EL, et al. Early detection of sepsis utilizing deep learning on electronic health record event sequences. \u003cem\u003eArtif Intell Med. \u003c/em\u003e2020;104:101820.\u003c/li\u003e\n\u003cli\u003eWong A, Otles E, Donnelly JP, et al. External Validation of a Widely Implemented Proprietary Sepsis Prediction Model in Hospitalized Patients. \u003cem\u003eJAMA Intern Med. \u003c/em\u003e2021;181(8):1065-1070.\u003c/li\u003e\n\u003cli\u003eKvamme H, Borgan \u0026Oslash;, Scheel I. Time-to-Event Prediction with Neural Networks and Cox Regression. \u003cem\u003eJournal of Machine Learning Research \u003c/em\u003e2019;20:1-30.\u003c/li\u003e\n\u003cli\u003eLei L, Bin Z. Risk Factor Differences in Acute Myocardial Infarction between Young and Older People: A Systematic Review and Meta-Analysis.\u003cem\u003e Int J Cardiovasc Sci. \u003c/em\u003e2019;32(2).\u003c/li\u003e\n\u003cli\u003eWallentin L, Becker RC, Budaj A, et al. Ticagrelor versus clopidogrel in patients with acute coronary syndromes. \u003cem\u003eN Engl J Med. \u003c/em\u003e2009;361(11):1045-1057.\u003c/li\u003e\n\u003cli\u003eWiviott SD, Braunwald E, McCabe CH, et al. Prasugrel versus clopidogrel in patients with acute coronary syndromes. \u003cem\u003eN Engl J Med. \u003c/em\u003e2007;357(20):2001-2015.\u003c/li\u003e\n\u003cli\u003eDivakaran S, Singh A, Biery D, et al. Diabetes Is Associated With Worse Long-term Outcomes in Young Adults After Myocardial Infarction: The Partners YOUNG-MI Registry. \u003cem\u003eDiabetes Care. \u003c/em\u003e2020;43(8):1843-1850.\u003c/li\u003e\n\u003cli\u003eAlaa AM, van der Schaar M. Deep multi-task gaussian processes for survival analysis with competing risks. 30th Conference on Neural Information Processing Systems; 2017.\u003c/li\u003e\n\u003cli\u003eIbanez B, James S, Agewall S, et al. 2017 ESC Guidelines for the management of acute myocardial infarction in patients presenting with ST-segment elevation: The Task Force for the management of acute myocardial infarction in patients presenting with ST-segment elevation of the European Society of Cardiology (ESC). \u003cem\u003eEur Heart J. \u003c/em\u003e2018;39(2):119-177.\u003c/li\u003e\n\u003cli\u003eRyan H, Trosclair A, Gfroerer J. Adult current smoking: differences in definitions and prevalence estimates--NHIS and NSDUH, 2008. \u003cem\u003eJ Environ Public Health. \u003c/em\u003e2012;2012:918368.\u003c/li\u003e\n\u003cli\u003eFox KA, Dabbous OH, Goldberg RJ, et al. Prediction of risk of death and myocardial infarction in the six months after presentation with acute coronary syndrome: prospective multinational observational study (GRACE). \u003cem\u003eBMJ. \u003c/em\u003e2006;333(7578):1091.\u003c/li\u003e\n\u003cli\u003eThygesen K, Alpert JS, Jaffe AS, et al. Fourth Universal Definition of Myocardial Infarction (2018). \u003cem\u003eCirculation. \u003c/em\u003e2018;138(20):e618-e651.\u003c/li\u003e\n\u003cli\u003eHeinze G, Wallisch C, Dunkler D. Variable selection - A review and recommendations for the practicing statistician. \u003cem\u003eBiom J. \u003c/em\u003e2018;60(3):431-449.\u003c/li\u003e\n\u003cli\u003eBelsley; DA, Kuh; E, Welsch RE. \u003cem\u003eRegression Diagnostics: Identifying Influential Data and Sources of Collinearity.\u003c/em\u003e 2nd edition ed: Wiley Interscience; 2013.\u003c/li\u003e\n\u003cli\u003eChen T, Guestrin C. XGBoost: A Scalable Tree Boosting System. . Paper presented at: KDD \u0026apos;16: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining2016; San Francisco.\u003c/li\u003e\n\u003cli\u003eBreiman L. Random forests. \u003cem\u003eMachine Learning. \u003c/em\u003e2001;45(1):5-32.\u003c/li\u003e\n\u003cli\u003eGranger CB, Goldberg RJ, Dabbous O, et al. Predictors of hospital mortality in the global registry of acute coronary events. \u003cem\u003eArch Intern Med. \u003c/em\u003e2003;163(19):2345-2353.\u003c/li\u003e\n\u003cli\u003eAustin PC, Lee DS, Fine JP. Introduction to the Analysis of Survival Data in the Presence of Competing Risks. \u003cem\u003eCirculation. \u003c/em\u003e2016;133(6):601-609.\u003c/li\u003e\n\u003cli\u003eFine JP, Gray RJ. A Proportional Hazards Model for the Subdistribution of a Competing Risk. \u003cem\u003eJournal of the American Statistical Association. \u003c/em\u003e1999;94 (446):496\u0026ndash;509.\u003c/li\u003e\n\u003cli\u003eBergstra J, Bengio Y. Random search for hyper-parameter optimization. \u003cem\u003eJournal of Machine Learning Research. \u003c/em\u003e2012;13:281\u0026ndash;305.\u003c/li\u003e\n\u003cli\u003eAntolini L, Boracchi P, Biganzoli E. A time-dependent discrimination index for survival data. . \u003cem\u003eStatistics in Medicine \u003c/em\u003e2005;24:3927\u0026ndash;3944.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp style=\"text-align: center;\"\u003e\u003cstrong\u003eTable 1\u003c/strong\u003e\u003c/p\u003e\n\u003cp style=\"text-align: center;\"\u003eBaseline characteristics and clinical outcomes of individuals with premature acute coronary syndrome (ACS, \u0026le;55 years old) and older subjects with ACS (\u0026gt;55 years old)\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cdiv align=\"center\"\u003e\n \u003ctable border=\"0\" cellpadding=\"0\" cellspacing=\"0\" width=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"42.96724470134875%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd width=\"24.277456647398843%\"\u003e\n \u003cp\u003e\u003cstrong\u003eACS\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026gt;55 years-old\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.69942196531792%\"\u003e\n \u003cp\u003e\u003cstrong\u003eACS\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u0026le;\u003cstrong\u003e55 years-old\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"9.05587668593449%\"\u003e\n \u003cp\u003ep\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003en\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e4099\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e2242\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eAge (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e68.47 (8.11)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e47.63 (5.84)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eMale gender (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e60.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e66.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eDiagnoses\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eIndex diagnosis (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eSTEMI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e60.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e74.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eNSTEMI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e20.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e15.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eUA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e19.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e10.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eT2DM (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e34.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e27.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eT2DM on insulin (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e9.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e8.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.177\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eSmokers (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e34.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e41.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eDyslipidemia (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e20.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e19.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.550\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eHypertension (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e74.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e66.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eObesity (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e5.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e7.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eFamily history of premature CAD (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e9.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e16.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003ePrior ethylic habit (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e9.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e14.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003ePrior drug abuse (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e1.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e6.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003ePrior AMI (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e8.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e7.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003ePrior stroke (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e4.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e2.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.003\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003ePrior PAD (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e4.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e2.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.025\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003ePrior CKD (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e8.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e2.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003ePrior PCI (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e8.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e6.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.002\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003ePrior CABG (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e5.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e3.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.007\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003ePrior cocaine abuse (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e0.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e1.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003ePrior marijuana abuse (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e0.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e0.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.027\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eAtrial fibrillation (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e3.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e0.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eDrugs prescribed at discharge\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eNitrate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e45.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e42.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.123\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eStatin (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e83.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e81.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.385\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eBetablockers (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e64.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e66.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.293\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eARB or ACEi (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e58.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e61.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eCCB (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e20.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e18.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.278\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eASA (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e91.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e89.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.153\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eClopidogrel (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e62.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e64.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.244\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003ePrasugrel (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e21.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e22.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.941\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eTicagrelor (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e3.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e4.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.422\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eAnticoagulant (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e3.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e3.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.816\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eSpironolactone (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e11.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e7.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eFurosemide (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e15.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e9.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eCoronary artery stenoses (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eLCA % (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e5.50 (17.66)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e1.97 (10.40)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eProximal LAD % (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e30.07 (39.15)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e29.18 (45.19)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.079\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eLAD % (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e36.77 (39.87)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e27.93 (38.85)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eD1 % (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e20.79 (40.38)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e15.28 (30.34)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eIntermedium % (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e4.56 (18.38)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e4.42 (18.19)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.077\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eCx % (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e35.27 (39.66)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e24.17 (36.35)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eM1 % (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e18.08 (33.17)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e11.12 (27.64)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eRCA % (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e50.77 (40.71)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e37.84 (41.10)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003ePDA % (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e6.94 (21.88)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e5.19 (18.70)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eRMA % (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e5.44 (19.88)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e4.12 (17.51)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eSevere coronary artery lesions\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003e1-vessel with proximal LAD (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e4.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e8.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003e1-vessel with LAD (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e8.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e9.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003e1-vessel with RCA (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e10.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e12.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003e2-vessels without LAD (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e13.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e11.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.133\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003e3-vessels with LAD (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e13.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e8.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003e3-vessels without LAD (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e16.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e10.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003e3-vessels with LCA and LAD (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e1.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e0.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003e3-vessels with LCA / without LAD (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e1.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e0.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.007\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003ePCI - index coronarography\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eLAD (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e13.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e18.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eMINOCA (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e2.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e1.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.223\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eNumber of new stents (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e1.38 (0.72)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e1.44 (0.70)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eEchocardiography\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eApical dyskinesia\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e3.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e1.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eApical akinesia\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e29.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e23.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eLV function at 3\u003csup\u003erd\u003c/sup\u003e to 5\u003csup\u003eth\u003c/sup\u003e day (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.005\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003e\u0026gt;45%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e22.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e24.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003e\u0026lt;45%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e77.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e75.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"2\" valign=\"bottom\" width=\"67.3076923076923%\"\u003e\n \u003cp\u003eAdmission metrics (STEMI individuals only; n=4042)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.653846153846153%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.038461538461538%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eCardiac arrest before admission (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e2.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e0.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eTime pain-primary hospital, minutes (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e164.51 (142.39)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e153.66 (131.44)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.043\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eTime door-needle, minutes (median [IQR])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e70.00 [43.00, 120.00]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e68.00 [43.00, 110.50]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.269\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eTime pain-needle, minutes (median [IQR])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e225.00 [150.00, 335.00]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e210.00 [140.00, 315.00]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.005\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eTime tnk-coronarography, minutes (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e1195.82 (1269.01)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e1270.69 (1137.35)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.114\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eCoronarography duration, minutes (median [IQR])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e55.00 [40.00, 75.00]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e50.00 [40.00, 65.00]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003ePharmacoinvasive strategy (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e88.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e90.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.698\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003ePrimary PCI (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e11.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e9.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.452\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eSBP at admission, mmHg (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e121.86 (26.01)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e143.43 (25.22)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eDBP at admission, mmHg (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e75.46 (17.19)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e88.18 (17.21)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eHR at admission, beats/minute (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e81.00 (19.47)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e77.40 (15.44)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eKillip score (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e43.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e84.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eII\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e25.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e14.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eIII\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e16.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eIV\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e14.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eTIMI flow pre-PCI (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e2.05 (1.22)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e2.20 (1.18)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.003\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eTIMI flow post-PCI (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e2.58 (0.81)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e2.69 (0.75)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.002\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eMBG pre-PCI (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e1.61 (1.45)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e1.84 (1.42)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eMBG post-PCI (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e1.86 (1.35)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e2.18 (1.22)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eClinical scores\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eTIMI score (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e5.15 (2.35)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e2.46 (1.56)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eGRACE in-hospital death (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e150.80 (32.75)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e89.47 (18.19)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eGRACE score (6 months) (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e141.49 (24.28)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e91.61 (15.96)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eCRUSADE (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e35.75 (13.87)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e19.44 (11.29)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eLaboratory exams\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eTroponin (peak) (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e9999 (10399)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e6838 (6465)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eGlycemia, mg/dL (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e159.51 (81.38)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e137.86 (62.80)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eHbA1c, % (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e6.87 (2.03)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e6.63 (2.01)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.014\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eTotal cholesterol, mg/dL (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e194.74 (48.58)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e204.71 (46.58)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eHDL-cholesterol, mg/dL (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e42.01 (12.26)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e41.02 (12.68)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.069\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eLDL-cholesterol, mg/dL (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e122.82 (39.59)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e132.19 (40.23)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eAST, mg/dL (median [IQR])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e185.00 [80.00, 334.00]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e148.00 [75.00, 272.75]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eALT, mg/dL (median [IQR])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e50.00 [29.00, 83.75]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e44.00 [29.00, 69.00]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eTriglycerides, mg/dL (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e152.54 (133.41)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e166.06 (118.62)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.009\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eCreatinine, mg/dL (median [IQR])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e0.98 [0.81, 1.23]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e0.84 [0.71, 1.00]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eCreatinine clearance, ml/min/1.73m2 (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e74.52 (28.94)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e107.10 (33.94)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eBMI, kg/m2 (mean (SD))\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e26.65 (4.57)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e27.15 (4.51)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.005\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"42.96724470134875%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"2\" valign=\"bottom\" width=\"67.3076923076923%\"\u003e\n \u003cp\u003eClinical outcomes (considering competing risks)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.653846153846153%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.038461538461538%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eIn-hospital deaths (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e5,7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e3,9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003ePost-discharge deaths (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e9,6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eGlobal deaths (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e15.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e7.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eGlobal CV deaths (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e5.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e3.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.018\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eGlobal non-CV deaths (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e10.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e4.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eMI during follow-up (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eSTEMI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e12.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e5.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eNSTEMI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e15.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e18.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eFollow-up time, days (median [IQR])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e2331 [2030, 2625]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e2436 [2039, 2644]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eCABG (in-hospital) (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e4.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e2.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.017\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" width=\"42.96724470134875%\"\u003e\n \u003cp\u003eCABG during long-term follow-up (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"24.277456647398843%\"\u003e\n \u003cp\u003e13.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"23.69942196531792%\"\u003e\n \u003cp\u003e10.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" width=\"9.05587668593449%\"\u003e\n \u003cp\u003e0.026\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eCABG: coronary artery bypass graft; CKD: chronic kidney disease; GFR: glomerular filtration rate (CKD-EPI); GPIIbIIIa: glycoprotein IIbIIIa; HbA1c: glycosylated hemoglobin; LDL-C: low-density lipoprotein cholesterol; LV: left ventricle; BMI: body mass index; NSTEMI: non-ST-elevation myocardial infarction; PCI: percutaneous coronary intervention; STEMI: ST-elevation myocardial infarction; SBP: systolic blood pressure; DBP: diastolic blood pressure; HR: heart rate; CV: cardiovascular.\u003cbr\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp style=\"text-align: center;\"\u003e\u003cstrong\u003eTable 2\u003c/strong\u003e\u003c/p\u003e\n\u003cp style=\"text-align: center;\"\u003e \u003cem\u003eAccuracy\u003c/em\u003e and \u003cem\u003eC-statistics\u003c/em\u003e for short-term window models in predicting \u003cem\u003ein-hospital\u003c/em\u003e death or recurrent ischemic events in 2,242 individuals with premature ACS (55 years old or younger), total number of events = 180 (\u003cem\u003ein-hospital\u003c/em\u003e CV deaths = 39, and MI = 141)\u003c/p\u003e\n\u003ctable align=\"\" border=\"1\" cellpadding=\"0\" cellspacing=\"0\" style=\"border-collapse: collapse; margin: 0px auto;\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"45.21452145214521%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.887788778877887%\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;(95% confidence interval)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"26.897689768976896%\"\u003e\n \u003cp\u003eC-statistics\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;(95% confidence interval)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"45.21452145214521%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.887788778877887%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"26.897689768976896%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"45.21452145214521%\"\u003e\n \u003cp\u003e\u003cem\u003eLogistic regression\u003c/em\u003e with GRACE score risk factors\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.887788778877887%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"26.897689768976896%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"45.21452145214521%\"\u003e\n \u003cp\u003eValidation set (mean of 5-folds)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.887788778877887%\"\u003e\n \u003cp\u003e0.841 (0.819 - 0.866)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"26.897689768976896%\"\u003e\n \u003cp\u003e0.834 (0.819 - 0.866)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"45.21452145214521%\"\u003e\n \u003cp\u003eTest set (n=673, 75 events)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.887788778877887%\"\u003e\n \u003cp\u003e0.820 (0.784 - 0.856)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"26.897689768976896%\"\u003e\n \u003cp\u003e0.819 (0.782 - 0.853)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"45.21452145214521%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.887788778877887%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"26.897689768976896%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"45.21452145214521%\"\u003e\n \u003cp\u003e\u003cem\u003eLogistic regression\u003c/em\u003e (top 20 predictors)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.887788778877887%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"26.897689768976896%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"45.21452145214521%\"\u003e\n \u003cp\u003eValidation set (mean of 5-folds)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.887788778877887%\"\u003e\n \u003cp\u003e0.889 (0.867 - 0.904)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"26.897689768976896%\"\u003e\n \u003cp\u003e0.883 (0.867 - 0.904)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"45.21452145214521%\"\u003e\n \u003cp\u003eTest set (n=673, 75 events)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.887788778877887%\"\u003e\n \u003cp\u003e0.880 (0.854 - 0.896)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"26.897689768976896%\"\u003e\n \u003cp\u003e0.872 (0.851 - 0.894)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"45.21452145214521%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.887788778877887%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"26.897689768976896%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"45.21452145214521%\"\u003e\n \u003cp\u003e\u003cem\u003eRandom Forests\u003c/em\u003e (top 20 predictors)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.887788778877887%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"26.897689768976896%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"45.21452145214521%\"\u003e\n \u003cp\u003eValidation set (mean of 5-folds)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.887788778877887%\"\u003e\n \u003cp\u003e0.901 (0.870 - 0.928)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"26.897689768976896%\"\u003e\n \u003cp\u003e0.908 (0.879 - 0.933)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"45.21452145214521%\"\u003e\n \u003cp\u003eTest set (n=673, 75 events)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.887788778877887%\"\u003e\n \u003cp\u003e0.892 (0.852 - 0.930)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"26.897689768976896%\"\u003e\n \u003cp\u003e0.888 (0.846 - 0.927)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"45.21452145214521%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.887788778877887%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"26.897689768976896%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"45.21452145214521%\"\u003e\n \u003cp\u003e\u003cem\u003eXGBoost\u003c/em\u003e (top 20 predictors)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.887788778877887%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"26.897689768976896%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"45.21452145214521%\"\u003e\n \u003cp\u003eValidation set (mean of 5-folds)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.887788778877887%\"\u003e\n \u003cp\u003e0.894 (0.870 - 0.929)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"26.897689768976896%\"\u003e\n \u003cp\u003e0.898 (0.872 - 0.931)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"45.21452145214521%\"\u003e\n \u003cp\u003eTest set (n=673, 75 events)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.887788778877887%\"\u003e\n \u003cp\u003e0.876 (0.859 - 0.903)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"26.897689768976896%\"\u003e\n \u003cp\u003e0.861 (0.835 - 0.891)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"45.21452145214521%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.887788778877887%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"26.897689768976896%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"45.21452145214521%\"\u003e\n \u003cp\u003e\u003cem\u003eTabNet\u003c/em\u003e (top 20 predictors)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.887788778877887%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"26.897689768976896%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"45.21452145214521%\"\u003e\n \u003cp\u003eValidation set (mean of 5-folds)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.887788778877887%\"\u003e\n \u003cp\u003e0.951 (0.924 - 0.979)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"26.897689768976896%\"\u003e\n \u003cp\u003e0.936 (0.904 - 0.955)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"45.21452145214521%\"\u003e\n \u003cp\u003eTest set (n=673, 75 events)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.887788778877887%\"\u003e\n \u003cp\u003e0.946 (0.917 - 0.975)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"26.897689768976896%\"\u003e\n \u003cp\u003e0.921 (0.889 - 0.953)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp style=\"text-align: center;\"\u003e\u003cstrong\u003eTable 3\u003c/strong\u003e\u003c/p\u003e\n\u003cp style=\"text-align: center;\"\u003e\u0026nbsp;Time-dependent C-statistics (C\u003cem\u003e\u003csup\u003etd\u003c/sup\u003e\u003c/em\u003e-index) for predicting long-term noncardiovascular deaths or MACEs (cardiovascular deaths and recurrent ischemic events) with competing risks occurring (i) after discharge in 2,161 individuals with premature ACS (55 years old or younger), total number of \u003cem\u003epostdischarge\u003c/em\u003e\u003cem\u003e\u0026nbsp;\u003c/em\u003eevents: 47\u0026nbsp;noncardiovascular\u0026nbsp;deaths and 454\u0026nbsp;MACEs; (ii)\u0026nbsp;48 h\u0026nbsp;after index ACS hospital admission (\u003cem\u003eglobal follow-up\u003c/em\u003e models), total number of events: 92 noncardiovascular deaths and 631 MACEs.\u0026nbsp;\u003c/p\u003e\n\u003ctable align=\"left\" border=\"1\" cellpadding=\"0\" cellspacing=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"1.4285714285714286%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"55.91836734693877%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"38.97959183673469%\"\u003e\n \u003cp\u003eC\u003cem\u003e\u003csup\u003etd\u003c/sup\u003e\u003c/em\u003e-index\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;(95% confidence interval)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"3.673469387755102%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"1.4285714285714286%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"55.91836734693877%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"38.97959183673469%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"3.673469387755102%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"1.4285714285714286%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"55.91836734693877%\"\u003e\n \u003cp\u003e\u003cem\u003ePost-discharge\u003c/em\u003e horizon\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"38.97959183673469%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"3.673469387755102%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"2\" width=\"9.775967413441956%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"55.80448065173116%\"\u003e\n \u003cp\u003e\u003cem\u003eCS-Cox\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"34.419551934826885%\"\u003e\n \u003cp\u003e0.602 (95% CI 0.556-0.649)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"2\" width=\"9.775967413441956%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"55.80448065173116%\"\u003e\n \u003cp\u003e\u003cem\u003eFine-Gray\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"34.419551934826885%\"\u003e\n \u003cp\u003e0.612 (95% CI 0.564-0.663)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"2\" width=\"9.775967413441956%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"55.80448065173116%\"\u003e\n \u003cp\u003e\u003cem\u003eDMGP\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"34.419551934826885%\"\u003e\n \u003cp\u003e0.685 (95% CI 0.639-0.725)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"2\" width=\"9.775967413441956%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"55.80448065173116%\"\u003e\n \u003cp\u003e\u003cem\u003eDeepHit\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"34.419551934826885%\"\u003e\n \u003cp\u003e0.722 (95% CI 0.678-0.760)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"2\" width=\"9.775967413441956%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"55.80448065173116%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"34.419551934826885%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"55.80448065173116%\"\u003e\n \u003cp\u003e\u003cem\u003eGlobal follow-up\u003c/em\u003e horizon\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"34.419551934826885%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" width=\"9.775967413441956%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"2\" width=\"9.775967413441956%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"55.80448065173116%\"\u003e\n \u003cp\u003e\u003cem\u003eCS-Cox\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"34.419551934826885%\"\u003e\n \u003cp\u003e0.597 (95% CI 0.552-0.643)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"2\" width=\"9.775967413441956%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"55.80448065173116%\"\u003e\n \u003cp\u003e\u003cem\u003eFine-Gray\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"34.419551934826885%\"\u003e\n \u003cp\u003e0.601 (95% CI 0.559-0.660)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"2\" width=\"9.775967413441956%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"55.80448065173116%\"\u003e\n \u003cp\u003e\u003cem\u003eDMGP\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"34.419551934826885%\"\u003e\n \u003cp\u003e0.687 (95% CI 0.640-0.728)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"2\" width=\"9.775967413441956%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"55.80448065173116%\"\u003e\n \u003cp\u003e\u003cem\u003eDeepHit\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" valign=\"top\" width=\"34.419551934826885%\"\u003e\n \u003cp\u003e0.681 (95% CI 0.654-0.703)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Acute coronary syndromes, Risk prediction, Clinical intelligence, Population health management","lastPublishedDoi":"10.21203/rs.3.rs-1824283/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-1824283/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eAcute coronary syndrome (ACS) is a common cause of death in individuals older than 55 years. Although younger individuals are less frequently seen with ACS, this clinical event has increasing incidence trends and triggers considerable economic burden. Young individuals with ACS (yACS) are usually underrepresented and show idiosyncratic epidemiologic features compared to older subjects. These differences may justify why available risk prediction models usually penalize yACS with higher false positive rates compared to older subjects. We hypothesized that exploring temporal \u003cem\u003eframing structures\u003c/em\u003e such as prediction time, observation windows and window shifts, coupled with subgroup-specific prediction, could improve time-dependent prediction metrics. In a large cohort of ACS individuals (n\u003csub\u003e\u003cem\u003eglobal_cohort\u003c/em\u003e\u003c/sub\u003e=6341 and n\u003csub\u003eyACS\u003c/sub\u003e=2242), the predictive accuracy for adverse clinical events in ACS was optimized by using specific rules for yACS and splitting short-term and long-term prediction windows, leading to the detection of 80% of events, compared to 68% by using a rule designed for the global cohort.\u003c/p\u003e","manuscriptTitle":"The Framing of time-dependent machine learning risk prediction models among young individuals with acute coronary syndromes","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2022-07-19 17:22:55","doi":"10.21203/rs.3.rs-1824283/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Major revision","date":"2022-11-25T10:04:18+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2022-11-17T19:10:52+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"f905db3e-f575-46ed-b8aa-b6964fae29c3","date":"2022-11-02T03:15:19+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2022-08-08T06:00:52+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"f702a230-828f-4358-a8d3-123743f6ce7a","date":"2022-07-18T14:46:57+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2022-07-12T17:04:02+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2022-07-12T16:31:59+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2022-07-12T03:24:53+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2022-07-11T18:10:30+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2022-07-04T14:43:58+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"4c58c4db-96cf-4707-be6c-86793a948704","owner":[],"postedDate":"July 19th, 2022","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2023-10-16T18:31:01+00:00","versionOfRecord":{"articleIdentity":"rs-1824283","link":"https://doi.org/10.1038/s41598-023-27776-0","journal":{"identity":"scientific-reports","isVorOnly":false,"title":"Scientific Reports"},"publishedOn":"2023-01-19 18:24:43","publishedOnDateReadable":"January 19th, 2023"},"versionCreatedAt":"2022-07-19 17:22:55","video":"","vorDoi":"10.1038/s41598-023-27776-0","vorDoiUrl":"https://doi.org/10.1038/s41598-023-27776-0","workflowStages":[]},"version":"v1","identity":"rs-1824283","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-1824283","identity":"rs-1824283","version":["v1"]},"buildId":"_2-kVJe1T_tPrBINL-cwx","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}