Multimorbidity profile and the risk of major cardiovascular events in new antipsychotic users: a time-to-event prediction study using explainable machine learning

Multimorbidity profile and the risk of major cardiovascular events in new antipsychotic users: a time-to-event prediction study using explainable machine learning | 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 Multimorbidity profile and the risk of major cardiovascular events in new antipsychotic users: a time-to-event prediction study using explainable machine learning Qi Sun, Wenlong Liu, Cuiling Wei, Yuqi Hu, Lingyue Zhou, Boyan Liu, and 10 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5355838/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Introduction : Antipsychotic users are at an elevated risk of major adverse cardiovascular events (MACE) due to many interacting risk factors. However, specific antipsychotic agents, underlying multimorbidity, and chronic medication patterns in relation to MACE are little explored. Aims : To identify patients’ characteristics with increased risk of MACE in people with multimorbidity and using antipsychotics and to develop and evaluate a time-to-event prediction model. Methods : This retrospective cohort study utilized electronic health records from public healthcare facilities in Hong Kong. We included MACE-free patients aged 18-65 years with records of two or more chronic health conditions within three years prior to their first antipsychotic use. Baseline characteristics, such as age, sex, chronic disease history, antipsychotic usage history, and drug intake history over the previous year, were considered. The outcome was major adverse cardiovascular events (MACE), which included stroke, acute myocardial infarction (AMI), and cardiovascular-related death (CV death). The dataset was randomly divided into training and validation subsets in a 7:3 ratio based on the initial year of antipsychotic prescription. A Conditional Inference Survival Tree (CISTree) was employed to identify MACE risk groups. Ten machine learning models were trained using 5-fold cross-validation for hyperparameter optimization and validated on the validation set. We conducted time-dependent ROC curve analysis, calibration plots, and decision curve analysis plots to compare the models' discrimination capacity, calibration, and clinical application value, respectively. Time-dependent variable importance, partial dependence plots, and SHAP plots were used to explain the selected model. Results : A total of 26,274 patients were included in the study. The CISTree model identified older patients (>48 years) with chronic kidney disease (CKD), who were using antibacterial and antiplatelet drugs but not taking antidepressants, and without metastatic cancer, as having the highest MACE incidence rate (171.317 per 1,000 person-years; 95% CI: [130.088, 221.467]). The random survival model outperformed the other nine models, identifying age, antidepressant usage, and CKD as the top three significant predictors, consistent with the CISTree model. The survival C-statistics (ranging from 0 to 1, with higher values indicating better predictive precision) for 1-, 3-, and 5-year MACE predictions in the validation cohort were estimated at 0.841, 0.835, and 0.824, respectively. Conclusion : We identified specific high-risk MACE groups among individuals with multimorbidity who started using antipsychotics. Predictions based on these features demonstrated excellent accuracy and have the potential to aid clinical decision-making. Health sciences/Diseases/Cardiovascular diseases Health sciences/Diseases/Psychiatric disorders Health sciences/Medical research/Epidemiology Health sciences/Health care/Disease prevention Health sciences/Health care/Public health Biological sciences/Computational biology and bioinformatics/Machine learning Biological sciences/Computational biology and bioinformatics/Predictive medicine Multimorbidity antipsychotics major adverse cardiovascular events time-to-event prediction explainable machine learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 INTRODUCTION Major adverse cardiovascular events (MACE), including stroke, acute myocardial infarction (AMI), and cardiovascular-related death (CV death), represent a significant burden for individuals and society. 1 In 2021, MACE accounted for more than 224 deaths per 100,000 annually in the United States. 2 In addition, it incurs a substantial economic burden on the healthcare system with an estimated annual cost of $ 503.2 billion from direct and indirect expenses nationwide. 3 Of particular importance is the population of non-older-adults (aged 65 years or younger), for whom the incidence of early onset cardiovascular diseases (CVD) has increased substantially over recent decades, contributing to an increasing share of the total CVD burden and years of life lost. 4 , 5 A disproportionate share of the burden of MACE is borne by people with mental disorders. 6 In particular, long-term antipsychotic use has been shown to be moderately associated with MACE, 7 in addition to known adverse effects from behavioural and other prevalent risk factors among populations with mental health problems. Despite a good effectiveness in the treatment of a wide range of psychiatric disorders and symptoms with a generally tolerable safety profile, antipsychotics may be associated with cardiovascular side effects, e.g., QT interval prolongation, hypotension, metabolic side effects, etc. 8 – 10 The mechanisms underlying these associations operate either individually or in combination with other medications. 8 , 10 , 11 For instance, Marie et al. reported that Danish older people who used levomepromazine or haloperidol have a threefold higher incidence rate of MACE. 7 Wu et al. observed a similar effect where people using first-generation antipsychotics had 1.7-times higher odds of ventricular arrhythmia and/or sudden cardiac death, and second-generation antipsychotics users had 1.4-times higher such odds. 12 The current working hypothesis for this consistently observed elevated risk is largely based on the known metabolic abnormalities antipsychotics typically associate with. 13 A significant proportion of people with severe mental illness also live with multimorbidity, referred to as the co-existence of two or more chronic conditions, which constitutes additional increased risks of MACE. 14 Ross et al. estimated that there is a threefold higher hazard rate (HR) of MACE in rheumatoid arthritis (RA) patients with complex multimorbidity compared to those without it in the UK. 14 Previous research has also suggested certain specific multimorbidity patterns, such as hypertension-hyperlipidemia, to be more strongly associated with MACE or other adverse health outcomes. 15 Likewise, it is plausible that there are similar patterns that are predictive of the risk of MACE in antipsychotic users. Nevertheless, such patterns have yet to be investigated with population-representative data. Given the significant social and economic implications of MACE in antipsychotic users, it is beneficial to develop a robust, accurate, and interpretable MACE prediction model to inform clinical practices based on large longitudinal databases. Importantly, this will help clinicians prevent or mitigate the risk of MACE in younger patients on antipsychotics with multimorbidity. In this study, we aim to leverage a territory-wide, population-representative public healthcare database in Hong Kong to characterize the pre-existing multimorbidity profiles of new antipsychotic users at higher risk of MACE and to develop and validate an explainable machine learning-based prediction model. METHODS Study design and data source This retrospective machine learning cohort study utilized territory-wide electronic health records spanning from January 1, 2001, to August 31, 2022, maintained by the Hong Kong Hospital Authority (HA). The HA oversees all local public hospitals and the majority of public outpatient clinics, providing health services for over 7.5 million Hong Kong residents. The database documented every diagnosis made by registered clinicians based on the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) with accurate timestamps. Primary care (general outpatient clinic) records were coded according to the International Classification of Primary Care (ICPC) system. Death data were drawn from the city’s death registry, with causes of death coded according to the International Classification of Diseases, Tenth Revision, Clinical Modification (ICD-10-CM). The date of the first recorded antipsychotic drug initiation was used as the index date of the cohort. For the ascertainment of baseline characteristics and inclusion/exclusion criteria, we retrospectively collected chronic disease records ( Supplement Table 1 ) within three years prior to the index date and medication usage history ( Supplement Table 2 ) within one year. Individuals identified as having multimorbidity were followed up from the index date until the occurrence of the outcomes of interest (described in related sections below), death, or the end of data availability (August 31, 2022), whichever came first. This right censoring allowed the calculation of MACE-free periods from the index date to one of these observation endpoints. The Institutional Review Board of the University of Hong Kong/Hospital Authority Hong Kong West Cluster (HKU/HA HKW IRB, reference number: CIRB-2022-015-5) approved this study. Informed consent was not required, as the database was anonymized, and the data collection process complied with data privacy regulations. Inclusion and exclusion criteria Records from January 1, 2001, to August 31, 2022, were examined and patients with a record of at least two listed chronic diseases ( Supplement Table 1 ) within three years before the index date, i.e., antipsychotic initiation (British National Formulary (B.N.F.) 4.2.1 and 4.2.2, except lithium, were included. Patients who first used antipsychotics before January 1, 2004; had incorrect/ missing records (e.g., death time occurring before the diagnosis date); were not aged 18–65 on the index date; or had MACE before were excluded. As earliest relevant information was only available from 2001, we designated 2001–2003 as the wash-out period for the execution of the exclusion criteria. Outcomes, predictors, and data preparation MACE was adopted as the study outcome of interest and defined as a composite binary indicator of stroke (ICD-9-CM codes 430–438), acute myocardial infarction (AMI) (ICD-9 code 410), or cardiovascular-related death (CV death, ICD-10-CM codes I00 - I02, I05 - I16, I1A, I20 - I28, I30 - I52, I5A, I60 - I89, I95 - I99). Patients with any of these records were considered to have MACE. Baseline characteristics collected for analyses as independent variables included age, sex, chronic disease history, specific antipsychotic agent initiated, and one-year medication history ( Supplement Table 2 ). Diagnostic records were identified by ICD-9 or ICPC codes ( Supplement Table 1 ). For MACE events occurring at 0 days, we encoded as 0.5 days to avoid modelling errors. We performed random splitting to generate training and validation datasets at a 7:3 ratio stratified by index date. We set 1-year, 3-year, and 5-year time points after the index date as interesting time points in post-hoc model performance diagnostics. The validation dataset was only used for validating machine learning models. Risk group characterization Based on the wide range of predictors, we employed conditional inference survival tree (CISTree) to risk-classify MACE within the training dataset only. Renowned for its unbiased recursive partitioning, CISTree is less prone to overfitting compared to conventional decision tree models. 16 Development, validation and interpretation of prediction models We employed the Boruta algorithm, with 200 iterations, as the variable selection method on the training set to free machine learning models from correlated variables and unwanted noise in prediction. 17 We then trained ten machine learning models—Cox regression model (Cox), generalized linear models with least absolute shrinkage and selection operator regression (LASSO), 18 Generalized linear models with elastic net regularization (Elastic net), 18 survival tree, 19 random survival forest (RSF), 20 conditional random forest (Cforest), 21 survival gradient boosting machine (GBM), 22 survival neural networks (Deepsurv), 23 extreme gradient boosting survival learner (XGBoost), 24 and likelihood-based boosting survival Cox model (Coxboost) 25 —using 5-fold cross-validation for grid hyperparameter searching. (modelling and variable selection details are in ( Supplementary Table 3 ) The performance of each model was assessed internally in the training cohort using the "concordance statistics survival measure (survival C-index)" during each validation iteration. 26 All prediction algorithms were employed to develop and tune models in the training set and then we validated their performance in the validation set. The models’ performance was compared in three aspects—discrimination, calibration, and clinical value which will be reflected by receiver operating characteristic (ROC) curves, calibration plots, and DCA plots. 27 Advancements in machine learning interpretation have enabled us to explain these "black box" models. 28 We created time-dependent feature importance plots, partial dependence plots, and Shapley Additive explanations (SHAP) plots for all variables and for specific individual in the training set. 28 – 31 Computing programming packages and environment We conducted all analyses using R (version 4.4.0), employing the " partykit " package to establish the CISTree model, " Boruta " for variable selection, " mlr3proba ” to build ten machine learning models, " survex " for model explanations, and the " shiny " package to construct the prediction platform. RESULTS In total, 443,231 patients used antipsychotics (B.N.F. 4.2.1 and 4.2.2, except lithium) from January 1, 2004, to August 31, 2022. We included 26,274 patients with information on their age, sex, antipsychotic usage, additional drug use, and diagnosed chronic diseases from January 1, 2004, to August 31, 2022, as the original database. After random splitting, 18,400 individuals were allocated to the training set, and 7,874 were in the validation set (see Fig. 1 ). Baseline characteristics are tabulated as Table 1 . Totally, 1,260 out of 18,400 patients experienced MACE within a median follow-up period of 514 (Interquartile range (IQR) 33 to 2,850) days in the training cohort, and 524 out of 7,874 people in the validation cohort experienced MACE during a median follow-up time of 554 (IQR: 36 to 2,878) days. The MACE incidence rate (IR) in the training set was 16.153 (95% confidence interval (CI): [15.273, 17.070]) per 1,000 person-years. Table 1 Baseline characteristics Training cohort Validating cohort N. 18,400 7,874 Sex (%) Male 9,387 (51.0) 4,126 (52.4) Female 9,013 (49.0) 3,748 (47.6) Age (median [IQR]) 55 [47,60] 55 [47,61] 18–30 (No. (%)) 885 (4.8) 385 (4.9) 31–45 (No. (%)) 3,223 (17.5) 1,361 (17.3) 46–65 (No. (%)) 14,292 (77.7) 6,128 (77.8) Multimorbidity (%) Alcohol misuse 1,111 (6.0) 555 (7.0) Asthma 769 (4.2) 336 (4.3) Atrial fibrillation 301 (1.6) 162 (2.1) Cancer lymphoma 228 (1.2) 111 (1.4) Cancer metastatic 6,912 (37.6) 2,863 (36.4) Cancer non-metastatic 5,812 (31.6) 2,441 (31.0) Chronic kidney disease 1,713 (9.3) 752 (9.6) Chronic pain 5,755 (31.3) 2,466 (31.3) Chronic pulmonary disease 1,179 (6.4) 514 (6.5) Chronic viral hepatitis B 678 (3.7) 290 (3.7) Cirrhosis 1,224 (6.7) 509 (6.5) Dementia 197 (1.1) 76 (1.0) Depression 5,597 (30.4) 2,348 (29.8) Diabetes 3,196 (17.4) 1,396 (17.7) Epilepsy 425 (2.3) 212 (2.7) Hypertension 4,969 (27.0) 2,171 (27.6) Hypothyroidism 462 (2.5) 177 (2.2) Inflammatory bowel disease 6 (0.0) 1 (0.0) Irritable bowel syndrome 137 (0.7) 43 (0.5) Multiple sclerosis 54 (0.3) 35 (0.4) Parkinson disease 91 (0.5) 38 (0.5) Peptic ulcer disease 411 (2.2) 190 (2.4) Peripheral vascular disease 20 (0.1) 13 (0.2) Psoriasis 90 (0.5) 35 (0.4) Rheumatoid arthritis 192 (1.0) 80 (1.0) Schizophrenia 651 (3.5) 271 (3.4) Severe constipation 2,827 (15.4) 1,235 (15.7) Retinal vascular occlusion 53 (0.3) 18 (0.2) One-year medication history (%) Antibacterial drugs 10,120 (55.0) 4,302 (54.6) Drug used in diabetes 3,386 (18.4) 1,484 (18.8) Antidepressant drugs 7,228 (39.3) 3,095 (39.3) Corticosteroids 6,411 (34.8) 2,707 (34.4) Antiviral drugs 1,833 (10.0) 736 (9.3) β-adrenoceptor blocking drugs 3,343 (18.2) 1,445 (18.4) Antiplatelet drugs 1,312 (7.1) 547 (6.9) Anti-arrhythmic drugs 570 (3.1) 264 (3.4) Antianginal drugs 4,739 (25.8) 2,073 (26.3) Drugs used in hypertension and heart failure 2,569 (14.0) 1,136 (14.4) Diuretics 3,941 (21.4) 1,681 (21.3) Lipid regulating drugs 2,329 (12.7) 1,036 (13.2) Anticoagulants and protamine 364 (2.0) 137 (1.7) Drugs affecting the immune response 286 (1.6) 131 (1.7) Immunoglobulins 1 (0.0) 1 (0.0) Cough preparations 1 (0.0) 0 (0.0) Cytotoxic drugs 0 (0.0) 1 (0.0) Antipsychotic agent initiated (%) Amisulpride 118 (0.6) 52 (0.7) Aripiprazole 247 (1.3) 105 (1.3) Asenapine 0 (0.0) 0 (0.0) Brexpiprazole 14 (0.1) 8 (0.1) Chlorpromazine 1,897 (10.3) 834 (10.6) Clozapine 10 (0.1) 7 (0.1) Flupenthixol 182 (1.0) 88 (1.1) Fluphenazine 0 (0.0) 0 (0.0) Haloperidol 9,862 (53.6) 4,235 (53.8) Lurasidone 6 (0.0) 3 (0.0) Molindone 0 (0.0) 0 (0.0) Olanzapine 355 (1.9) 142 (1.8) Paliperidone 14 (0.1) 5 (0.1) Pericyazine 21 (0.1) 10 (0.1) Perphenazine 112 (0.6) 41 (0.5) Pimozide 2 (0.0) 1 (0.0) Quetiapine 3,370 (18.3) 1,435 (18.2) Risperidone 1,243 (6.8) 542 (6.9) Sertindole 2 (0.0) 0 (0.0) Sulpiride 785 (4.3) 346 (4.4) Thioridazine 28 (0.2) 8 (0.1) Thiothixene 0 (0.0) 1 (0.0) Trifluoperazine 377 (2.0) 154 (2.0) Ziprasidone 16 (0.1) 1 (0.0) Zuclopenthixol 13 (0.1) 6 (0.1) MACE distribution MACE cases (%) 1,260 (6.8) 524 (6.7) MACE happening time (days, median [IQR]) 514 [33, 2,850] 554 [36, 2878] Identification of MACE risk group There were 18,400 patients ranging across the 30 nodes generated by CISTree algorithm. In the Fig. 2 , node 55, which represents patients older than 48 years, living with chronic kidney disease (CKD), having used antibacterial and antiplatelet drugs but not antidepressants, and without metastatic cancer, has the highest IR, 171.317 (95% CI: [130.088, 221.467]) per 1,000 person-years. In node 57, people older than 48 years, diagnosed with CKD and metastatic cancer have the lowest MACE risk across all nodes, i.e., 0.000 (95% CI: [0.000, 32.939]) per 1,000 person-years. The CISTree suggests that age, antidepressant drugs, and CKD are the top three factors most related to MACE status. Time-to-MACE survival prediction A total of 24 variables were chosen for modelling, with the selection process depicted in Supplementary File . These covariates include age, atrial fibrillation, metastatic cancer, non-metastatic cancer, CKD, cirrhosis, diabetes, hypertension, schizophrenia, severe constipation, antidepressant drugs, corticosteroids, antibacterial drugs, antiviral drugs, drugs used in diabetes, β adrenoceptor blocking drugs, antiplatelet drugs, antianginal drugs, drugs used in hypertension and heart failure, diuretics, lipid regulating drugs, haloperidol, quetiapine, and risperidone. Internal validation diagnostics The hyperparameters tuning results and the chosen values are in Supplementary Table 3 . Supplementary Table 4 presents the survival C-index, survival calibration score, and right-censored log loss for each tuned machine learning model in training set. RSF and XGBoost demonstrate better overall internal performance among the ten models. Validation using the validating cohort We compared the discrimination capacity of ten tuned machine learning methods on the validation dataset. 32 The time-dependent ROC plots are presented in Fig. 3 (a to c) . According to these figures, RSF performs the best among the three time points with an AUC from 0.824 to 0.841, while Deepsurv performs the worst with an AUC below 0.800. The calibration performance of each model was gauged and visualized using survival Brier Scores and calibration plot, in Fig. 3 (d to f) , respectively. These figures show that RSF, has the best calibration ability since their calibration curves align with the diagonals better than other models, and its brier score is relatively smaller, at the three time points. Comparing the Decision Curve Analysis (DCA) curves among the ten models is necessary (see Fig. 3 -g to i ). 33 The higher the net benefit under certain threshold probability, which was used as the threshold to predict MACE happening, the better the model performs in practice. Here, RSF consistently outperforms compared to other models. (In case of avoiding missed diagnosed case, lowering false negative, the net proportion of true negative can serve as the net benefit, in Supplement Fig. 1 , where RSF performs best) 33 , 34 From the three aspects of model comparison - discrimination, calibration, and clinical practice, it can be concluded that the RSF model exhibits the most comprehensive performance (high survival C-index, low survival brier score, and high net benefit) in the validation set. Machine learning model interpretation We interpreted the RSF model by identifying variable importance to discover the most important predictors over time. Figure 4 demonstrates how the importance of different predictors varied over time, suggesting that age, CKD, and antidepressant drugs are the top three predictors. Furthermore, for predicting MACE status in the first year, antidepressant drug is the most important predictor, which is surpassed by CKD in the third and fifth years. Analogizing important predictors from RSF model to nodes of CISTree, all factors appearing in the CISTree model are included in RSF. Particularly, the top three factors of CISTree, age, CKD, and antidepressant drugs, are not only strongly associated with the time-to-event outcome of MACE, but also have significant prediction/ risk classification value. Additionally, Fig. 5 , the partial dependent plot (PDP), reveals how the top seven prediction-important features (the rest variables PDP are in Supplementary Fig. 2 ) values affects MACE-free predictive probability when other factors remain constant. It shows that with increasing age (Fig. 5 -a), a patient's MACE status worsens, while other predictors are held constant. Similar phenomena exist with the usage of antiplatelet drugs (Fig. 5 -f), haloperidol (Fig. 5 -g), and a CKD diagnosis (Fig. 5 -c). The difference in MACE-free predictive probability between population non-recorded and recorded CKD is 3.9% on the first follow-up year and 8.0% on the fifth year; the difference in MACE-free predictive probability between patients unused and used antiplatelet drug is 5.7% and 8.3% on the first and fifth year respectively. Constantly, antidepressants usage is associated with 1.5%-higher MACE-free probability. Results are similar with the Cox model ( Supplementary Table 5 ), not only the direction but the effect magnitudes are consistent, the more greatly the two lines diverge on PDP, the larger the coefficients in Cox. The global SHAP plot (Fig. 6 ) illustrates how feature values impact the predicted risk of MACE compared to the average predicted probability. Older age is associated with a higher estimated MACE risk, while younger age (< 43 years old) indicates a lower and consistent MACE risk. Additionally, chronic kidney disease (CKD) and haloperidol use are linked to an increased estimated MACE risk. (A SHAP plot for individual explanations is shown in Supplement Fig. 3 .) Based on the results from RSF model, we built a pilot online interactive platform, HKU Antipsychotics Multimorbidity MACE calculator ( Supplement Fig. 4 , HKUAPMMCal, https://hkupmmcal.shinyapps.io/HKUAPMMCal ) to potentially aid clinical practice. DISCUSSION We found that the cluster of new antipsychotic users with multimorbidity, aged 48 and older, living with chronic kidney disease (CKD), using antibacterial and antiplatelet medications (but not antidepressants), and without metastatic cancer, has the highest risk of MACE. Among various machine learning models evaluated for discrimination, calibration, and clinical utility, the Random Survival Forest (RSF) model outperformed the others, demonstrating the highest validation performance and potentially offering the greatest net benefit in aiding clinical management. Older age, CKD, and the absence of antidepressant use emerged as the top three crucial predictors, with older age and CKD indicating a substantially heightened risk of MACE. Additionally, users of haloperidol or antiplatelet drugs within this population were shown to have a significantly elevated risk of MACE. Consistent with our observation on specific predictors as interpreted from the model, a US cohort study reported a slight increase in mortality risk within seven days among patients with AMI starting haloperidol compared to other typical antipsychotics (hazard ratio (HR): 1.50, 95% CI 1.14 to 1.96). 35 Similarly, Mikkel et al. found that low-dose quetiapine (≤ 50 mg tablets prescribed for over 365 days) was associated with a 1.13 times higher risk of MACE in their intention-to-treat analysis (95% CI: [1.02, 1.24]) and a 1.52 times higher risk in as-treated analysis (95% CI: [1.35, 1.70]). 36 These findings are consistent with our conclusions as shown in the PDP (Fig. 5 -g and Supplement Fig. 2 -a). A cohort study in Wales demonstrated an increased risk of MACE among individuals with CKD (eGFR 30 mL/min/1.73m²): HR 3.43 (95% CI: [3.22, 3.64]) in the heart failure and atrial fibrillation multimorbidity cluster, and HR 4.18 (95% CI: [3.65, 4.78]) in the heart failure, peripheral vascular disease, and diabetes cluster, as observed in another cohort study in Sweden. 37 Briana et al. also noted that among patients with type 2 diabetes, those with at least four co-morbidities had a 2.68 times higher risk of cardiovascular death (95% CI: [2.52, 2.85]) compared to those without multimorbidity, 19 years after the diagnosis of type 2 diabetes. 38 To our knowledge, this study represents the first machine learning approach for predicting MACE risks among new antipsychotic users with pre-existing multimorbidity, employing a robust model with comprehensive comparisons. Machine learning algorithms are renowned for their flexibility and comprehensiveness in considering all available information specific to individuals. In clinical practice, it is particularly valuable for clinicians and health consultants to assess not only the likelihood of a patient experiencing MACE but also the approximate timing. This allows for more precise preventive measures, potentially conserving medical resources and alleviating the healthcare burden on communities and nations. Previous machine learning studies on MACE prediction have primarily focused on predicting whether MACE will occur within a certain period of time. For instance, Ayako et al. utilized a random forest model with MRI data to predict MACE among adults with repaired tetralogy of Fallot, achieving a C-index of 0.82 (95% CI: [0.74, 0.89]). 39 Similarly, Jain et al. employed Gradient Boosting Machines (GBM) and XGBoost algorithms to predict MACE occurrence following orthotopic liver transplantation, achieving a C-index of 0.71 (95% CI: [0.63, 0.79]). 40 These studies, however, are limited in clinical applicability without sufficiently assessing the trade-offs between sensitivity, specificity, as well as burdens from false positives and false negatives, which we thoroughly investigated in a series of decision curve analyses. And they did not demonstrate a thoughtful selection process of the appropriate model through comparison by a wide range of diagnostics. More importantly, this territory-wide database enhances the robustness and prediction performance of the RSF model due to its large size, representativeness, and continuous updates. The selected predictors are readily accessible compared to more complex data sources like highly specific laboratory test results, MRI, or ultrasound scans, as all variables can be obtained through routine clinical assessments and patient records. This accessibility facilitates the swift computation of predicted risk for individual patients during consultations, enabling clinicians to easily explain risk prospects. Therefore, our study is motivated by clinical needs, reproducible, and applicable in clinical practice. However, it is important to note that despite its plausible local applicability and clinical value, this is a single-center study. Multiple independent external validation studies are essential to assess the model's generalizability across different populations and settings, should this prediction model prove to be needed elsewhere. Additionally, the observed relationships between antipsychotics, concurrent medications, multimorbidity patterns, and MACE risk warrant further investigation, given the predictive and observational nature of our study. CONCLUSION This study identified a high-risk group for MACE among non-older patients with multimorbidity receiving antipsychotics and established a validated machine learning time-to-event MACE prediction model. Further external validation is necessary to assess the model’s applicability across different settings. Additionally, future research should investigate potential causal relationships between the identified factors and MACE risk, as inferring causality is beyond the scope of this study. Declarations Author Contribution F.T.T.L. Conceptualization, project administration, writing – review and editing; I.C.K.W. Data curation, supervision, methodology, project administration, writing – review and editing; E.W.Y.C. Data curation, supervision, project administration, writing – review and editing; D.P.J.O. and D.S. Methodology, project administration, writing – review and editing; K.K.F.T and S.K.W.C. Data curation, project administration, writing – review and editing; Q.S. Conceptualization, formal analysis, visualization, writing – review and editing; W.L. and B.L. formal analysis, visualization, writing – review and editing; Y.H. Methodology, formal analysis, writing – review and editing; C.W. Methodology, formal analysis, writing – review and editing; L.Z. Data curation, writing – review and editing; R.Y.K.C. Data curation, writing – review and editing; S.S. Data curation, writing – review and editing; W.T. Resources, writing – review and editing Acknowledgement We gratefully acknowledge Dr. Man-Li Tse, Dr. Joey Shuk-Yan Leung, Dr. Lawrence Chi-Lun Lai, Dr. Jonathan Gabriel Sung, Prof. Sandra Sau Man Chan, Dr. Ivan Man-Ho Wong for their advice from clinical perspectives and support in funding acquisition. We also thank the Hospital Authority for providing data. Data Availability Data underlying the results presented in this article cannot be shared publicly because the raw data is confidential and not allowed for sharing in accordance with the prevailing policies of the Hospital Authority of Hong Kong. The data may be requested from Hong Kong Hospital Authority's Central Panel on Administrative Assessment of External Data Requests (https://www3.ha.org.hk/data/Provision/Submission). References Carlsson A, Irewall AL, Graipe A, Ulvenstam A, Mooe T, Ögren J. Long-term risk of major adverse cardiovascular events following ischemic stroke or TIA. Sci Rep. 2023;13(1):8333. Heart and Stroke Statistics. 2024; https://www.heart.org/en/about-us/heart-and-stroke-association-statistics . Lloyd-Jones D, Adams RJ, Brown TM, et al. Heart Disease and Stroke Statistics—2010 Update. Circulation. 2010;121(7):e46-e215. Andersson C, Vasan RS. Epidemiology of cardiovascular disease in young individuals. Nature Reviews Cardiology. 2018;15(4):230–240. Masaebi F, Salehi M, Kazemi M, Vahabi N, Azizmohammad Looha M, Zayeri F. Trend analysis of disability adjusted life years due to cardiovascular diseases: results from the global burden of disease study 2019. BMC Public Health. 2021;21(1):1268. Goldfarb M, De Hert M, Detraux J, et al. Severe Mental Illness and Cardiovascular Disease: JACC State-of-the-Art Review. Journal of the American College of Cardiology. 2022;80(9):918–933. Sahlberg M, Holm E, Gislason GH, Køber L, Torp-Pedersen C, Andersson C. Association of Selected Antipsychotic Agents With Major Adverse Cardiovascular Events and Noncardiovascular Mortality in Elderly Persons. J Am Heart Assoc. 2015;4(9):e001666. D'Errico S, Russa R, Maiese A, et al. Atypical antipsychotics and oxidative cardiotoxicity: review of literature and future perspectives to prevent sudden cardiac death. J Geriatr Cardiol. 2021;18(8):663–685. David M. Taylor AHY, Thomas R. E. Barnes. The Maudsley Prescribing Guidelines in Psychiatry, 14th Edition. Wiley-Blackwell; 2021. Siwek M, Woroń J, Gorostowicz A, Wordliczek J. Adverse effects of interactions between antipsychotics and medications used in the treatment of cardiovascular disorders. Pharmacological Reports. 2020;72(2):350–359. Schizophrenia and Related Psychoses. In: The Maudsley Prescribing Guidelines in Psychiatry. 1-224. Wu CS, Tsai YT, Tsai HJ. Antipsychotic Drugs and the Risk of Ventricular Arrhythmia and/or Sudden Cardiac Death: A Nation-wide Case‐Crossover Study. Journal of the American Heart Association. 2015;4(2):e001568. Friedrich ME, Winkler D, Konstantinidis A, et al. Cardiovascular Adverse Reactions During Antipsychotic Treatment: Results of AMSP, A Drug Surveillance Program Between 1993 and 2013. Int J Neuropsychopharmacol. 2020;23(2):67–75. McQueenie R, Nicholl BI, Jani BD, et al. Patterns of multimorbidity and their effects on adverse outcomes in rheumatoid arthritis: a study of 5658 UK Biobank participants. BMJ Open. 2020;10(11):e038829. Dai D, Fernandes J, Sun X, Lupton L, Payne VW, Berk A. Multimorbidity in Atherosclerotic Cardiovascular Disease and Its Associations With Adverse Cardiovascular Events and Healthcare Costs: A Real-World Evidence Study. J Health Econ Outcomes Res. 2024;11(1):75–85. Hothorn T, Hornik K, Zeileis A. ctree: Conditional inference trees. The comprehensive R archive network. 2015;8. Kursa MB, Rudnicki WR. Feature Selection with the Boruta Package. Journal of Statistical Software. 2010;36(11):1–13. Friedman J, Hastie T, Tibshirani R. Regularization Paths for Generalized Linear Models via Coordinate Descent. J Stat Softw. 2010;33(1):1–22. Breiman L, Friedman, J., Olshen, R.A., & Stone, C.J. Classification and Regression Trees . 1st ed. ed: Chapman and Hall/CRC; 1984. Ishwaran H, Kogalur UB, Blackstone EH, Lauer MS. Random survival forests. The Annals of Applied Statistics. 2008;2(3):841–860, 820. Hothorn T, Zeileis A. Partykit: a modular toolkit for recursive partytioning in R. J Mach Learn Res. 2015;16(1):3905–3909. Friedman JH. Stochastic gradient boosting. Computational Statistics & Data Analysis. 2002;38(4):367–378. Katzman JL, Shaham U, Cloninger A, Bates J, Jiang T, Kluger Y. DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network. BMC Medical Research Methodology. 2018;18(1):24. Chen T, Guestrin C. XGBoost: A Scalable Tree Boosting System. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining; 2016; San Francisco, California, USA. Binder H, Allignol A, Schumacher M, Beyersmann J. Boosting for high-dimensional time-to-event data with competing risks. Bioinformatics. 2009;25(7):890–896. Hastie T T, R., & Friedman, J. The Elements of Statistical Learning: Data Mining, Inference, and Prediction . 2nd ed. Stanford, CA: Stanford University; 2009. Steyerberg EW, Vergouwe Y. Towards better clinical prediction models: seven steps for development and an ABCD for validation. Eur Heart J. 2014;35(29):1925–1931. Hassija V, Chamola V, Mahapatra A, et al. Interpreting Black-Box Models: A Review on Explainable Artificial Intelligence. Cognitive Computation. 2024;16(1):45–74. Binder H, Allignol A, Schumacher M, Beyersmann J. Boosting for high-dimensional time-to-event data with competing risks. Bioinformatics. 2009;25(7):890–896. Lipovetsky S, Conklin M. Analysis of regression in game theory approach. Applied Stochastic Models in Business and Industry. 2001;17(4):319–330. Altmann A, Toloşi L, Sander O, Lengauer T. Permutation importance: a corrected feature importance measure. Bioinformatics. 2010;26(10):1340–1347. Baj G, Gandin I, Scagnetto A, et al. Comparison of discrimination and calibration performance of ECG-based machine learning models for prediction of new-onset atrial fibrillation. BMC Medical Research Methodology. 2023;23(1):169. Piovani D, Sokou R, Tsantes AG, Vitello AS, Bonovas S. Optimizing Clinical Decision Making with Decision Curve Analysis: Insights for Clinical Investigators. Healthcare. 2023;11(16):2244. Van Calster B, Wynants L, Verbeek JFM, et al. Reporting and Interpreting Decision Curve Analysis: A Guide for Investigators. Eur Urol. 2018;74(6):796–804. Park Y, Bateman BT, Kim DH, et al. Use of haloperidol versus atypical antipsychotics and risk of in-hospital death in patients with acute myocardial infarction: cohort study. BMJ. 2018;360:k1218. Højlund M, Andersen K, Ernst MT, Correll CU, Hallas J. Use of low-dose quetiapine increases the risk of major adverse cardiovascular events: results from a nationwide active comparator-controlled cohort study. World Psychiatry. 2022;21(3):444–451. Sullivan MK, Carrero JJ, Jani BD, et al. The presence and impact of multimorbidity clusters on adverse outcomes across the spectrum of kidney function. BMC Med. 2022;20(1):420. Coles B, Zaccardi F, Hvid C, Davies MJ, Khunti K. Cardiovascular events and mortality in people with type 2 diabetes and multimorbidity: A real-world study of patients followed for up to 19 years. Diabetes, Obesity and Metabolism. 2021;23(1):218–227. Ishikita A, McIntosh C, Hanneman K, et al. Machine Learning for Prediction of Adverse Cardiovascular Events in Adults With Repaired Tetralogy of Fallot Using Clinical and Cardiovascular Magnetic Resonance Imaging Variables. Circulation: Cardiovascular Imaging. 2023;16(6):e015205. Jain V, Bansal A, Radakovich N, et al. Machine Learning Models to Predict Major Adverse Cardiovascular Events After Orthotopic Liver Transplantation: A Cohort Study. J Cardiothorac Vasc Anesth. 2021;35(7):2063–2069. Additional Declarations No competing interests reported. Supplementary Files SupplementFiguresandTabs.docx supplementfilemlmodelling.docx Cite Share Download PDF Status: Posted Version 1 posted 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-5355838","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":376656536,"identity":"8b7dcff3-17f9-41d1-af23-5700266e71d8","order_by":0,"name":"Qi Sun","email":"","orcid":"","institution":"The University of Hong Kong","correspondingAuthor":false,"prefix":"","firstName":"Qi","middleName":"","lastName":"Sun","suffix":""},{"id":376656537,"identity":"02d60171-43df-4f42-afee-7a9f711d64fb","order_by":1,"name":"Wenlong Liu","email":"","orcid":"","institution":"The University of Hong Kong","correspondingAuthor":false,"prefix":"","firstName":"Wenlong","middleName":"","lastName":"Liu","suffix":""},{"id":376656538,"identity":"3c17a982-7272-4360-991a-6741f663b60b","order_by":2,"name":"Cuiling Wei","email":"","orcid":"","institution":"The University of Hong Kong","correspondingAuthor":false,"prefix":"","firstName":"Cuiling","middleName":"","lastName":"Wei","suffix":""},{"id":376656539,"identity":"a899a859-353d-46c2-af8e-e18c4a1e0d88","order_by":3,"name":"Yuqi Hu","email":"","orcid":"","institution":"The University of Hong Kong","correspondingAuthor":false,"prefix":"","firstName":"Yuqi","middleName":"","lastName":"Hu","suffix":""},{"id":376656540,"identity":"7127ab45-ffcb-480d-b244-00305bf79da1","order_by":4,"name":"Lingyue Zhou","email":"","orcid":"","institution":"The University of Hong Kong","correspondingAuthor":false,"prefix":"","firstName":"Lingyue","middleName":"","lastName":"Zhou","suffix":""},{"id":376656541,"identity":"95cbd2e8-91d0-4b21-9829-25a3978877e8","order_by":5,"name":"Boyan Liu","email":"","orcid":"","institution":"The University of Hong Kong","correspondingAuthor":false,"prefix":"","firstName":"Boyan","middleName":"","lastName":"Liu","suffix":""},{"id":376656542,"identity":"2ba59e10-ec83-4eb7-a919-963200c6d0cf","order_by":6,"name":"Rachel Yui Ki Chu","email":"","orcid":"","institution":"The University of Hong Kong","correspondingAuthor":false,"prefix":"","firstName":"Rachel","middleName":"Yui Ki","lastName":"Chu","suffix":""},{"id":376656543,"identity":"b1f7f2cb-cc08-4943-96ec-6ae930b6ab75","order_by":7,"name":"Song Song","email":"","orcid":"","institution":"Laboratory of Data Discovery for Health (China)","correspondingAuthor":false,"prefix":"","firstName":"Song","middleName":"","lastName":"Song","suffix":""},{"id":376656544,"identity":"d16c5350-f2ba-46de-bbe6-1afc0eef8ca7","order_by":8,"name":"Wenxin Tian","email":"","orcid":"","institution":"The University of Hong Kong","correspondingAuthor":false,"prefix":"","firstName":"Wenxin","middleName":"","lastName":"Tian","suffix":""},{"id":376656545,"identity":"15ea0b92-df6a-49e2-a12c-f92dd73903cf","order_by":9,"name":"Esther Wai Yin Chan","email":"","orcid":"","institution":"The University of Hong Kong","correspondingAuthor":false,"prefix":"","firstName":"Esther","middleName":"Wai Yin","lastName":"Chan","suffix":""},{"id":376656546,"identity":"e39f3ca7-b4f5-45e3-9069-b8693fd07770","order_by":10,"name":"Sherry Kit Wa Chan","email":"","orcid":"","institution":"The University of Hong Kong","correspondingAuthor":false,"prefix":"","firstName":"Sherry","middleName":"Kit Wa","lastName":"Chan","suffix":""},{"id":376656547,"identity":"b72d43e9-cf05-4f85-819e-abbcb722de3f","order_by":11,"name":"Kelvin Kam Fai Tsoi","email":"","orcid":"","institution":"The Chinese University of Hong Kong","correspondingAuthor":false,"prefix":"","firstName":"Kelvin","middleName":"Kam Fai","lastName":"Tsoi","suffix":""},{"id":376656548,"identity":"15371648-3ccb-4311-b2ee-8c503741d563","order_by":12,"name":"Ian Chi Kei Wong","email":"","orcid":"","institution":"Aston University Birmingham","correspondingAuthor":false,"prefix":"","firstName":"Ian","middleName":"Chi Kei","lastName":"Wong","suffix":""},{"id":376656549,"identity":"9c3c84c8-99ff-48fc-bb8f-728a5f3c937e","order_by":13,"name":"David P.J. Osborn","email":"","orcid":"","institution":"University College London","correspondingAuthor":false,"prefix":"","firstName":"David","middleName":"P.J.","lastName":"Osborn","suffix":""},{"id":376656550,"identity":"b0c4eaa1-0619-46c2-ad4d-90845c815183","order_by":14,"name":"Daniel Smith","email":"","orcid":"","institution":"University of Edinburgh","correspondingAuthor":false,"prefix":"","firstName":"Daniel","middleName":"","lastName":"Smith","suffix":""},{"id":376656551,"identity":"75db9e7e-bac7-4ca7-9560-6f032da376d0","order_by":15,"name":"Francisco Tsz Tsun Lai","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAo0lEQVRIiWNgGAWjYBAC9gYGxgMMBjYJEgwMbIwNxGjhOcbMANSSRrIWhsOkaJHvP3DgR8H5PMkZCWwPZxClhY2Z4WCPwe1iaYkEdsMNxGixB2o5wGNwO3GeRAKb5AOibfljcI5ELYd5DA4kzgZpIcphPGzJBodlDJITZ/Y8bJMkzvvMBx8+fPPHLnHG8eRjkj3EaEECxMXKKBgFo2AUjAJiAAAumDFt1UjtZAAAAABJRU5ErkJggg==","orcid":"","institution":"The University of Hong Kong","correspondingAuthor":true,"prefix":"","firstName":"Francisco","middleName":"Tsz Tsun","lastName":"Lai","suffix":""}],"badges":[],"createdAt":"2024-10-29 16:08:13","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-5355838/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5355838/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":69444701,"identity":"ba4876b7-a1d8-49fd-9a8b-d11951e567ca","added_by":"auto","created_at":"2024-11-20 11:44:49","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":190783,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart of study sample selection and model’s development and validation\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-5355838/v1/e1f0ae91c30d533c6c397cb0.png"},{"id":69445005,"identity":"57da0049-4e53-4cf9-af76-bc3eec25a4e1","added_by":"auto","created_at":"2024-11-20 11:52:49","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":138163,"visible":true,"origin":"","legend":"\u003cp\u003eConditional inference survival tree for MACE with corresponding incidence rates and 95% confidence interval\u003c/p\u003e","description":"","filename":"image2.png","url":"https://assets-eu.researchsquare.com/files/rs-5355838/v1/4ed1ca3691351de8dfd3bde1.png"},{"id":69444704,"identity":"b565a85e-c361-47c2-acdc-b0580fc39c96","added_by":"auto","created_at":"2024-11-20 11:44:49","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":157746,"visible":true,"origin":"","legend":"\u003cp\u003eROC, Calibration, and DCA plot. a, b, c are three ROC plots at the 1\u003csup\u003est\u003c/sup\u003e, 3\u003csup\u003erd\u003c/sup\u003e, and 5\u003csup\u003eth\u003c/sup\u003e year after the initial date of antipsychotics use. d, e, f are three calibration plots at different time. g, h, i are DCA plots, where the odds of the threshold probability shows the ratio of the benefit achieved by curing a true positive case to the harm caused by unnecessary treatment to a false positive case, so the threshold should be determined by clinicians.\u003c/p\u003e","description":"","filename":"image4.png","url":"https://assets-eu.researchsquare.com/files/rs-5355838/v1/4c2dc2e3cb90697f40f5baa6.png"},{"id":69444702,"identity":"e69dd79d-08fd-47cf-8139-9f184cf2dccf","added_by":"auto","created_at":"2024-11-20 11:44:49","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":38169,"visible":true,"origin":"","legend":"\u003cp\u003eTime-dependent feature importance plot. The higher the brier score loss shown on Y-axis, the more importance the factor owns.\u003c/p\u003e","description":"","filename":"image5.png","url":"https://assets-eu.researchsquare.com/files/rs-5355838/v1/e3f0ed106fb42b1faebf71a1.png"},{"id":69444703,"identity":"93e7a957-a8c5-42d5-8d52-314b73c0d68a","added_by":"auto","created_at":"2024-11-20 11:44:49","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":54754,"visible":true,"origin":"","legend":"\u003cp\u003eTop seven features time-dependent partial dependence plot.\u003c/p\u003e","description":"","filename":"image6.png","url":"https://assets-eu.researchsquare.com/files/rs-5355838/v1/8a837a613edb212835a686c2.png"},{"id":69444698,"identity":"6ae58fab-25dd-40a7-bee0-d100b0eee735","added_by":"auto","created_at":"2024-11-20 11:44:49","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":165426,"visible":true,"origin":"","legend":"\u003cp\u003eGlobal SHAP plot for age and other factors. A positive SHAP value indicates that the feature value of certain sample has an increasing effect on the prediction value of MACE risk against the average predictive MACE probability in whole population. The color of the dots symbolizes the numerical value of the variables.\u003c/p\u003e","description":"","filename":"image7.png","url":"https://assets-eu.researchsquare.com/files/rs-5355838/v1/678114d85f99593081bd09ca.png"},{"id":78526721,"identity":"65dd7c89-9ba8-4658-913e-59b20522d57e","added_by":"auto","created_at":"2025-03-14 13:17:01","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1739375,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5355838/v1/b42261a3-76a4-47ab-92f4-5e65ce14e2da.pdf"},{"id":69445004,"identity":"1a39f59b-eead-4573-9e3b-7cd3204d2c8a","added_by":"auto","created_at":"2024-11-20 11:52:49","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":348165,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementFiguresandTabs.docx","url":"https://assets-eu.researchsquare.com/files/rs-5355838/v1/2a47893fc2e4ba2e2ded22d7.docx"},{"id":69444706,"identity":"5684f65e-fd55-4193-adce-ca1292b5e14b","added_by":"auto","created_at":"2024-11-20 11:44:49","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":24622,"visible":true,"origin":"","legend":"","description":"","filename":"supplementfilemlmodelling.docx","url":"https://assets-eu.researchsquare.com/files/rs-5355838/v1/9e81bab773669f50c3db5f68.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Multimorbidity profile and the risk of major cardiovascular events in new antipsychotic users: a time-to-event prediction study using explainable machine learning","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eMajor adverse cardiovascular events (MACE), including stroke, acute myocardial infarction (AMI), and cardiovascular-related death (CV death), represent a significant burden for individuals and society.\u003csup\u003e \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e \u003c/sup\u003e In 2021, MACE accounted for more than 224 deaths per 100,000 annually in the United States.\u003csup\u003e \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e \u003c/sup\u003e In addition, it incurs a substantial economic burden on the healthcare system with an estimated annual cost of \u003cspan\u003e$\u003c/span\u003e503.2\u0026nbsp;billion from direct and indirect expenses nationwide.\u003csup\u003e \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e \u003c/sup\u003e Of particular importance is the population of non-older-adults (aged 65 years or younger), for whom the incidence of early onset cardiovascular diseases (CVD) has increased substantially over recent decades, contributing to an increasing share of the total CVD burden and years of life lost.\u003csup\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\u003eA disproportionate share of the burden of MACE is borne by people with mental disorders.\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e In particular, long-term antipsychotic use has been shown to be moderately associated with MACE,\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e in addition to known adverse effects from behavioural and other prevalent risk factors among populations with mental health problems. Despite a good effectiveness in the treatment of a wide range of psychiatric disorders and symptoms with a generally tolerable safety profile, antipsychotics may be associated with cardiovascular side effects, e.g., QT interval prolongation, hypotension, metabolic side effects, etc.\u003csup\u003e\u003cspan additionalcitationids=\"CR9\" citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e The mechanisms underlying these associations operate either individually or in combination with other medications.\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e,\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e,\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e For instance, Marie et al. reported that Danish older people who used levomepromazine or haloperidol have a threefold higher incidence rate of MACE.\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e Wu et al. observed a similar effect where people using first-generation antipsychotics had 1.7-times higher odds of ventricular arrhythmia and/or sudden cardiac death, and second-generation antipsychotics users had 1.4-times higher such odds.\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e The current working hypothesis for this consistently observed elevated risk is largely based on the known metabolic abnormalities antipsychotics typically associate with.\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e A significant proportion of people with severe mental illness also live with multimorbidity, referred to as the co-existence of two or more chronic conditions, which constitutes additional increased risks of MACE.\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e Ross et al. estimated that there is a threefold higher hazard rate (HR) of MACE in rheumatoid arthritis (RA) patients with complex multimorbidity compared to those without it in the UK.\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e Previous research has also suggested certain specific multimorbidity patterns, such as hypertension-hyperlipidemia, to be more strongly associated with MACE or other adverse health outcomes.\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e Likewise, it is plausible that there are similar patterns that are predictive of the risk of MACE in antipsychotic users. Nevertheless, such patterns have yet to be investigated with population-representative data.\u003c/p\u003e \u003cp\u003eGiven the significant social and economic implications of MACE in antipsychotic users, it is beneficial to develop a robust, accurate, and interpretable MACE prediction model to inform clinical practices based on large longitudinal databases. Importantly, this will help clinicians prevent or mitigate the risk of MACE in younger patients on antipsychotics with multimorbidity. In this study, we aim to leverage a territory-wide, population-representative public healthcare database in Hong Kong to characterize the pre-existing multimorbidity profiles of new antipsychotic users at higher risk of MACE and to develop and validate an explainable machine learning-based prediction model.\u003c/p\u003e"},{"header":"METHODS","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eStudy design and data source\u003c/h2\u003e \u003cp\u003eThis retrospective machine learning cohort study utilized territory-wide electronic health records spanning from January 1, 2001, to August 31, 2022, maintained by the Hong Kong Hospital Authority (HA). The HA oversees all local public hospitals and the majority of public outpatient clinics, providing health services for over 7.5\u0026nbsp;million Hong Kong residents. The database documented every diagnosis made by registered clinicians based on the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) with accurate timestamps. Primary care (general outpatient clinic) records were coded according to the International Classification of Primary Care (ICPC) system. Death data were drawn from the city\u0026rsquo;s death registry, with causes of death coded according to the International Classification of Diseases, Tenth Revision, Clinical Modification (ICD-10-CM).\u003c/p\u003e \u003cp\u003eThe date of the first recorded antipsychotic drug initiation was used as the index date of the cohort. For the ascertainment of baseline characteristics and inclusion/exclusion criteria, we retrospectively collected chronic disease records (\u003cb\u003eSupplement\u003c/b\u003e Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) within three years prior to the index date and medication usage history (\u003cb\u003eSupplement Table\u0026nbsp;2\u003c/b\u003e) within one year. Individuals identified as having multimorbidity were followed up from the index date until the occurrence of the outcomes of interest (described in related sections below), death, or the end of data availability (August 31, 2022), whichever came first. This right censoring allowed the calculation of MACE-free periods from the index date to one of these observation endpoints.\u003c/p\u003e \u003cp\u003e The Institutional Review Board of the University of Hong Kong/Hospital Authority Hong Kong West Cluster (HKU/HA HKW IRB, reference number: CIRB-2022-015-5) approved this study. Informed consent was not required, as the database was anonymized, and the data collection process complied with data privacy regulations.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eInclusion and exclusion criteria\u003c/h3\u003e\n\u003cp\u003eRecords from January 1, 2001, to August 31, 2022, were examined and patients with a record of at least two listed chronic diseases (\u003cb\u003eSupplement\u003c/b\u003e Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) within three years before the index date, i.e., antipsychotic initiation (British National Formulary (B.N.F.) 4.2.1 and 4.2.2, except lithium, were included. Patients who first used antipsychotics before January 1, 2004; had incorrect/ missing records (e.g., death time occurring before the diagnosis date); were not aged 18\u0026ndash;65 on the index date; or had MACE before were excluded. As earliest relevant information was only available from 2001, we designated 2001\u0026ndash;2003 as the wash-out period for the execution of the exclusion criteria.\u003c/p\u003e\n\u003ch3\u003eOutcomes, predictors, and data preparation\u003c/h3\u003e\n\u003cp\u003eMACE was adopted as the study outcome of interest and defined as a composite binary indicator of stroke (ICD-9-CM codes 430\u0026ndash;438), acute myocardial infarction (AMI) (ICD-9 code 410), or cardiovascular-related death (CV death, ICD-10-CM codes I00 - I02, I05 - I16, I1A, I20 - I28, I30 - I52, I5A, I60 - I89, I95 - I99). Patients with any of these records were considered to have MACE.\u003c/p\u003e \u003cp\u003eBaseline characteristics collected for analyses as independent variables included age, sex, chronic disease history, specific antipsychotic agent initiated, and one-year medication history (\u003cb\u003eSupplement Table\u0026nbsp;2\u003c/b\u003e). Diagnostic records were identified by ICD-9 or ICPC codes (\u003cb\u003eSupplement\u003c/b\u003e Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eFor MACE events occurring at 0 days, we encoded as 0.5 days to avoid modelling errors. We performed random splitting to generate training and validation datasets at a 7:3 ratio stratified by index date. We set 1-year, 3-year, and 5-year time points after the index date as interesting time points in post-hoc model performance diagnostics. The validation dataset was only used for validating machine learning models.\u003c/p\u003e\n\u003ch3\u003eRisk group characterization\u003c/h3\u003e\n\u003cp\u003eBased on the wide range of predictors, we employed conditional inference survival tree (CISTree) to risk-classify MACE within the training dataset only. Renowned for its unbiased recursive partitioning, CISTree is less prone to overfitting compared to conventional decision tree models.\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n\u003ch3\u003eDevelopment, validation and interpretation of prediction models\u003c/h3\u003e\n\u003cp\u003eWe employed the Boruta algorithm, with 200 iterations, as the variable selection method on the training set to free machine learning models from correlated variables and unwanted noise in prediction.\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eWe then trained ten machine learning models\u0026mdash;Cox regression model (Cox), generalized linear models with least absolute shrinkage and selection operator regression (LASSO),\u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e Generalized linear models with elastic net regularization (Elastic net),\u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e survival tree,\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e random survival forest (RSF),\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e conditional random forest (Cforest),\u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e survival gradient boosting machine (GBM),\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e survival neural networks (Deepsurv),\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e extreme gradient boosting survival learner (XGBoost),\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e and likelihood-based boosting survival Cox model (Coxboost)\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e\u0026mdash;using 5-fold cross-validation for grid hyperparameter searching. (modelling and variable selection details are in (\u003cb\u003eSupplementary Table\u0026nbsp;3\u003c/b\u003e) The performance of each model was assessed internally in the training cohort using the \"concordance statistics survival measure (survival C-index)\" during each validation iteration.\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eAll prediction algorithms were employed to develop and tune models in the training set and then we validated their performance in the validation set. The models\u0026rsquo; performance was compared in three aspects\u0026mdash;discrimination, calibration, and clinical value which will be reflected by receiver operating characteristic (ROC) curves, calibration plots, and DCA plots.\u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eAdvancements in machine learning interpretation have enabled us to explain these \"black box\" models.\u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e We created time-dependent feature importance plots, partial dependence plots, and Shapley Additive explanations (SHAP) plots for all variables and for specific individual in the training set.\u003csup\u003e\u003cspan additionalcitationids=\"CR29 CR30\" citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eComputing programming packages and environment\u003c/h2\u003e \u003cp\u003eWe conducted all analyses using \u003cem\u003eR\u003c/em\u003e (version 4.4.0), employing the \"\u003cem\u003epartykit\u003c/em\u003e\" package to establish the CISTree model, \"\u003cem\u003eBoruta\u003c/em\u003e\" for variable selection, \"\u003cem\u003emlr3proba\u003c/em\u003e\u0026rdquo; to build ten machine learning models, \"\u003cem\u003esurvex\u003c/em\u003e\" for model explanations, and the \"\u003cem\u003eshiny\u003c/em\u003e\" package to construct the prediction platform.\u003c/p\u003e \u003c/div\u003e"},{"header":"RESULTS","content":"\u003cp\u003eIn total, 443,231 patients used antipsychotics (B.N.F. 4.2.1 and 4.2.2, except lithium) from January 1, 2004, to August 31, 2022. We included 26,274 patients with information on their age, sex, antipsychotic usage, additional drug use, and diagnosed chronic diseases from January 1, 2004, to August 31, 2022, as the original database. After random splitting, 18,400 individuals were allocated to the training set, and 7,874 were in the validation set (see Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Baseline characteristics are tabulated as Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Totally, 1,260 out of 18,400 patients experienced MACE within a median follow-up period of 514 (Interquartile range (IQR) 33 to 2,850) days in the training cohort, and 524 out of 7,874 people in the validation cohort experienced MACE during a median follow-up time of 554 (IQR: 36 to 2,878) days. The MACE incidence rate (IR) in the training set was 16.153 (95% confidence interval (CI): [15.273, 17.070]) per 1,000 person-years.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eBaseline characteristics\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTraining cohort\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eValidating cohort\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eN.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e18,400\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e7,874\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSex (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e9,387 (51.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4,126 (52.4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFemale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e9,013 (49.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3,748 (47.6)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge (median [IQR])\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e55 [47,60]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e55 [47,61]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e18\u0026ndash;30 (No. (%))\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e885 (4.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e385 (4.9)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e31\u0026ndash;45 (No. (%))\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3,223 (17.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1,361 (17.3)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e46\u0026ndash;65 (No. (%))\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e14,292 (77.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e6,128 (77.8)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c3\" namest=\"c1\"\u003e \u003cp\u003eMultimorbidity (%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAlcohol misuse\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1,111 (6.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e555 (7.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAsthma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e769 (4.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e336 (4.3)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAtrial fibrillation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e301 (1.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e162 (2.1)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCancer lymphoma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e228 (1.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e111 (1.4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCancer metastatic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6,912 (37.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2,863 (36.4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCancer non-metastatic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5,812 (31.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2,441 (31.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChronic kidney disease\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1,713 (9.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e752 (9.6)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChronic pain\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5,755 (31.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2,466 (31.3)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChronic pulmonary disease\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1,179 (6.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e514 (6.5)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChronic viral hepatitis B\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e678 (3.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e290 (3.7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCirrhosis\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1,224 (6.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e509 (6.5)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDementia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e197 (1.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e76 (1.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDepression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5,597 (30.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2,348 (29.8)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiabetes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3,196 (17.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1,396 (17.7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEpilepsy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e425 (2.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e212 (2.7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHypertension\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4,969 (27.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2,171 (27.6)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHypothyroidism\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e462 (2.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e177 (2.2)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInflammatory bowel disease\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6 (0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1 (0.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIrritable bowel syndrome\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e137 (0.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e43 (0.5)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMultiple sclerosis\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e54 (0.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e35 (0.4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eParkinson disease\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e91 (0.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e38 (0.5)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePeptic ulcer disease\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e411 (2.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e190 (2.4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePeripheral vascular disease\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e20 (0.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e13 (0.2)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePsoriasis\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e90 (0.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e35 (0.4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRheumatoid arthritis\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e192 (1.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e80 (1.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSchizophrenia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e651 (3.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e271 (3.4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSevere constipation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2,827 (15.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1,235 (15.7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRetinal vascular occlusion\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e53 (0.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e18 (0.2)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c3\" namest=\"c1\"\u003e \u003cp\u003eOne-year medication history (%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAntibacterial drugs\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e10,120 (55.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4,302 (54.6)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDrug used in diabetes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3,386 (18.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1,484 (18.8)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAntidepressant drugs\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e7,228 (39.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3,095 (39.3)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCorticosteroids\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6,411 (34.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2,707 (34.4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAntiviral drugs\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1,833 (10.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e736 (9.3)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eβ-adrenoceptor blocking drugs\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3,343 (18.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1,445 (18.4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAntiplatelet drugs\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1,312 (7.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e547 (6.9)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAnti-arrhythmic drugs\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e570 (3.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e264 (3.4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAntianginal drugs\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4,739 (25.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2,073 (26.3)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDrugs used in hypertension and heart failure\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2,569 (14.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1,136 (14.4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiuretics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3,941 (21.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1,681 (21.3)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLipid regulating drugs\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2,329 (12.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1,036 (13.2)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAnticoagulants and protamine\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e364 (2.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e137 (1.7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDrugs affecting the immune response\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e286 (1.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e131 (1.7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eImmunoglobulins\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1 (0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1 (0.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCough preparations\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1 (0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0 (0.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCytotoxic drugs\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0 (0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1 (0.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c3\" namest=\"c1\"\u003e \u003cp\u003eAntipsychotic agent initiated (%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAmisulpride\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e118 (0.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e52 (0.7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAripiprazole\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e247 (1.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e105 (1.3)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAsenapine\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0 (0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0 (0.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBrexpiprazole\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e14 (0.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8 (0.1)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChlorpromazine\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1,897 (10.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e834 (10.6)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClozapine\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e10 (0.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e7 (0.1)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFlupenthixol\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e182 (1.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e88 (1.1)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFluphenazine\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0 (0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0 (0.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHaloperidol\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e9,862 (53.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4,235 (53.8)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLurasidone\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6 (0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3 (0.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMolindone\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0 (0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0 (0.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOlanzapine\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e355 (1.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e142 (1.8)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePaliperidone\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e14 (0.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5 (0.1)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePericyazine\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e21 (0.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e10 (0.1)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePerphenazine\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e112 (0.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e41 (0.5)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePimozide\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2 (0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1 (0.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eQuetiapine\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3,370 (18.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1,435 (18.2)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRisperidone\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1,243 (6.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e542 (6.9)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSertindole\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2 (0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0 (0.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSulpiride\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e785 (4.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e346 (4.4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eThioridazine\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e28 (0.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8 (0.1)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eThiothixene\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0 (0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1 (0.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTrifluoperazine\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e377 (2.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e154 (2.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eZiprasidone\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e16 (0.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1 (0.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eZuclopenthixol\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e13 (0.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e6 (0.1)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c3\" namest=\"c1\"\u003e \u003cp\u003eMACE distribution\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMACE cases (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1,260 (6.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e524 (6.7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMACE happening time (days, median [IQR])\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e514 [33, 2,850]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e554 [36, 2878]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003eIdentification of MACE risk group\u003c/h3\u003e\n\u003cp\u003eThere were 18,400 patients ranging across the 30 nodes generated by CISTree algorithm. In the Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, node 55, which represents patients older than 48 years, living with chronic kidney disease (CKD), having used antibacterial and antiplatelet drugs but not antidepressants, and without metastatic cancer, has the highest IR, 171.317 (95% CI: [130.088, 221.467]) per 1,000 person-years. In node 57, people older than 48 years, diagnosed with CKD and metastatic cancer have the lowest MACE risk across all nodes, i.e., 0.000 (95% CI: [0.000, 32.939]) per 1,000 person-years. The CISTree suggests that age, antidepressant drugs, and CKD are the top three factors most related to MACE status.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eTime-to-MACE survival prediction\u003c/h2\u003e \u003cp\u003eA total of 24 variables were chosen for modelling, with the selection process depicted in \u003cb\u003eSupplementary File\u003c/b\u003e. These covariates include age, atrial fibrillation, metastatic cancer, non-metastatic cancer, CKD, cirrhosis, diabetes, hypertension, schizophrenia, severe constipation, antidepressant drugs, corticosteroids, antibacterial drugs, antiviral drugs, drugs used in diabetes, β adrenoceptor blocking drugs, antiplatelet drugs, antianginal drugs, drugs used in hypertension and heart failure, diuretics, lipid regulating drugs, haloperidol, quetiapine, and risperidone.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eInternal validation diagnostics\u003c/h2\u003e \u003cp\u003eThe hyperparameters tuning results and the chosen values are in \u003cb\u003eSupplementary Table\u0026nbsp;3\u003c/b\u003e. \u003cb\u003eSupplementary Table\u0026nbsp;4\u003c/b\u003e presents the survival C-index, survival calibration score, and right-censored log loss for each tuned machine learning model in training set. RSF and XGBoost demonstrate better overall internal performance among the ten models.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eValidation using the validating cohort\u003c/h2\u003e \u003cp\u003eWe compared the discrimination capacity of ten tuned machine learning methods on the validation dataset.\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e The time-dependent ROC plots are presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e \u003cb\u003e(a to c)\u003c/b\u003e. According to these figures, RSF performs the best among the three time points with an AUC from 0.824 to 0.841, while Deepsurv performs the worst with an AUC below 0.800.\u003c/p\u003e \u003cp\u003eThe calibration performance of each model was gauged and visualized using survival Brier Scores and calibration plot, in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e\u003cb\u003e(d to f)\u003c/b\u003e, respectively. These figures show that RSF, has the best calibration ability since their calibration curves align with the diagonals better than other models, and its brier score is relatively smaller, at the three time points.\u003c/p\u003e \u003cp\u003eComparing the Decision Curve Analysis (DCA) curves among the ten models is necessary (see Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e-g \u003cb\u003eto i\u003c/b\u003e).\u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e The higher the net benefit under certain threshold probability, which was used as the threshold to predict MACE happening, the better the model performs in practice. Here, RSF consistently outperforms compared to other models. (In case of avoiding missed diagnosed case, lowering false negative, the net proportion of true negative can serve as the net benefit, in \u003cb\u003eSupplement\u003c/b\u003e Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, where RSF performs best)\u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e,\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eFrom the three aspects of model comparison - discrimination, calibration, and clinical practice, it can be concluded that the RSF model exhibits the most comprehensive performance (high survival C-index, low survival brier score, and high net benefit) in the validation set.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eMachine learning model interpretation\u003c/h2\u003e \u003cp\u003eWe interpreted the RSF model by identifying variable importance to discover the most important predictors over time. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e demonstrates how the importance of different predictors varied over time, suggesting that age, CKD, and antidepressant drugs are the top three predictors. Furthermore, for predicting MACE status in the first year, antidepressant drug is the most important predictor, which is surpassed by CKD in the third and fifth years. Analogizing important predictors from RSF model to nodes of CISTree, all factors appearing in the CISTree model are included in RSF. Particularly, the top three factors of CISTree, age, CKD, and antidepressant drugs, are not only strongly associated with the time-to-event outcome of MACE, but also have significant prediction/ risk classification value.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAdditionally, Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, the partial dependent plot (PDP), reveals how the top seven prediction-important features (the rest variables PDP are in \u003cb\u003eSupplementary Fig.\u0026nbsp;2\u003c/b\u003e) values affects MACE-free predictive probability when other factors remain constant. It shows that with increasing age (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e-a), a patient's MACE status worsens, while other predictors are held constant. Similar phenomena exist with the usage of antiplatelet drugs (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e-f), haloperidol (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e-g), and a CKD diagnosis (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e-c). The difference in MACE-free predictive probability between population non-recorded and recorded CKD is 3.9% on the first follow-up year and 8.0% on the fifth year; the difference in MACE-free predictive probability between patients unused and used antiplatelet drug is 5.7% and 8.3% on the first and fifth year respectively. Constantly, antidepressants usage is associated with 1.5%-higher MACE-free probability. Results are similar with the Cox model (\u003cb\u003eSupplementary Table\u0026nbsp;5\u003c/b\u003e), not only the direction but the effect magnitudes are consistent, the more greatly the two lines diverge on PDP, the larger the coefficients in Cox.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe global SHAP plot (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e) illustrates how feature values impact the predicted risk of MACE compared to the average predicted probability. Older age is associated with a higher estimated MACE risk, while younger age (\u0026lt;\u0026thinsp;43 years old) indicates a lower and consistent MACE risk. Additionally, chronic kidney disease (CKD) and haloperidol use are linked to an increased estimated MACE risk. (A SHAP plot for individual explanations is shown in Supplement Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.)\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eBased on the results from RSF model, we built a pilot online interactive platform, HKU Antipsychotics Multimorbidity MACE calculator (\u003cb\u003eSupplement\u003c/b\u003e Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, HKUAPMMCal, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://hkupmmcal.shinyapps.io/HKUAPMMCal\u003c/span\u003e\u003cspan address=\"https://hkupmmcal.shinyapps.io/HKUAPMMCal\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) to potentially aid clinical practice.\u003c/p\u003e \u003c/div\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003eWe found that the cluster of new antipsychotic users with multimorbidity, aged 48 and older, living with chronic kidney disease (CKD), using antibacterial and antiplatelet medications (but not antidepressants), and without metastatic cancer, has the highest risk of MACE. Among various machine learning models evaluated for discrimination, calibration, and clinical utility, the Random Survival Forest (RSF) model outperformed the others, demonstrating the highest validation performance and potentially offering the greatest net benefit in aiding clinical management. Older age, CKD, and the absence of antidepressant use emerged as the top three crucial predictors, with older age and CKD indicating a substantially heightened risk of MACE. Additionally, users of haloperidol or antiplatelet drugs within this population were shown to have a significantly elevated risk of MACE.\u003c/p\u003e \u003cp\u003eConsistent with our observation on specific predictors as interpreted from the model, a US cohort study reported a slight increase in mortality risk within seven days among patients with AMI starting haloperidol compared to other typical antipsychotics (hazard ratio (HR): 1.50, 95% CI 1.14 to 1.96).\u003csup\u003e\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/sup\u003e Similarly, Mikkel et al. found that low-dose quetiapine (\u0026le;\u0026thinsp;50 mg tablets prescribed for over 365 days) was associated with a 1.13 times higher risk of MACE in their intention-to-treat analysis (95% CI: [1.02, 1.24]) and a 1.52 times higher risk in as-treated analysis (95% CI: [1.35, 1.70]).\u003csup\u003e\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u003c/sup\u003e These findings are consistent with our conclusions as shown in the PDP (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e-g and Supplement Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e-a). A cohort study in Wales demonstrated an increased risk of MACE among individuals with CKD (eGFR 30 mL/min/1.73m\u0026sup2;): HR 3.43 (95% CI: [3.22, 3.64]) in the heart failure and atrial fibrillation multimorbidity cluster, and HR 4.18 (95% CI: [3.65, 4.78]) in the heart failure, peripheral vascular disease, and diabetes cluster, as observed in another cohort study in Sweden.\u003csup\u003e\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e Briana et al. also noted that among patients with type 2 diabetes, those with at least four co-morbidities had a 2.68 times higher risk of cardiovascular death (95% CI: [2.52, 2.85]) compared to those without multimorbidity, 19 years after the diagnosis of type 2 diabetes.\u003csup\u003e\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eTo our knowledge, this study represents the first machine learning approach for predicting MACE risks among new antipsychotic users with pre-existing multimorbidity, employing a robust model with comprehensive comparisons. Machine learning algorithms are renowned for their flexibility and comprehensiveness in considering all available information specific to individuals. In clinical practice, it is particularly valuable for clinicians and health consultants to assess not only the likelihood of a patient experiencing MACE but also the approximate timing. This allows for more precise preventive measures, potentially conserving medical resources and alleviating the healthcare burden on communities and nations.\u003c/p\u003e \u003cp\u003ePrevious machine learning studies on MACE prediction have primarily focused on predicting whether MACE will occur within a certain period of time. For instance, Ayako et al. utilized a random forest model with MRI data to predict MACE among adults with repaired tetralogy of Fallot, achieving a C-index of 0.82 (95% CI: [0.74, 0.89]).\u003csup\u003e\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e\u003c/sup\u003e Similarly, Jain et al. employed Gradient Boosting Machines (GBM) and XGBoost algorithms to predict MACE occurrence following orthotopic liver transplantation, achieving a C-index of 0.71 (95% CI: [0.63, 0.79]).\u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e These studies, however, are limited in clinical applicability without sufficiently assessing the trade-offs between sensitivity, specificity, as well as burdens from false positives and false negatives, which we thoroughly investigated in a series of decision curve analyses. And they did not demonstrate a thoughtful selection process of the appropriate model through comparison by a wide range of diagnostics.\u003c/p\u003e \u003cp\u003eMore importantly, this territory-wide database enhances the robustness and prediction performance of the RSF model due to its large size, representativeness, and continuous updates. The selected predictors are readily accessible compared to more complex data sources like highly specific laboratory test results, MRI, or ultrasound scans, as all variables can be obtained through routine clinical assessments and patient records. This accessibility facilitates the swift computation of predicted risk for individual patients during consultations, enabling clinicians to easily explain risk prospects. Therefore, our study is motivated by clinical needs, reproducible, and applicable in clinical practice.\u003c/p\u003e \u003cp\u003eHowever, it is important to note that despite its plausible local applicability and clinical value, this is a single-center study. Multiple independent external validation studies are essential to assess the model's generalizability across different populations and settings, should this prediction model prove to be needed elsewhere. Additionally, the observed relationships between antipsychotics, concurrent medications, multimorbidity patterns, and MACE risk warrant further investigation, given the predictive and observational nature of our study.\u003c/p\u003e"},{"header":"CONCLUSION","content":"\u003cp\u003eThis study identified a high-risk group for MACE among non-older patients with multimorbidity receiving antipsychotics and established a validated machine learning time-to-event MACE prediction model. Further external validation is necessary to assess the model\u0026rsquo;s applicability across different settings. Additionally, future research should investigate potential causal relationships between the identified factors and MACE risk, as inferring causality is beyond the scope of this study.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eF.T.T.L. Conceptualization, project administration, writing \u0026ndash; review and editing; I.C.K.W. Data curation, supervision, methodology, project administration, writing \u0026ndash; review and editing; E.W.Y.C. Data curation, supervision, project administration, writing \u0026ndash; review and editing; D.P.J.O. and D.S. Methodology, project administration, writing \u0026ndash; review and editing; K.K.F.T and S.K.W.C. Data curation, project administration, writing \u0026ndash; review and editing; Q.S. Conceptualization, formal analysis, visualization, writing \u0026ndash; review and editing; W.L. and B.L. formal analysis, visualization, writing \u0026ndash; review and editing; Y.H. Methodology, formal analysis, writing \u0026ndash; review and editing; C.W. Methodology, formal analysis, writing \u0026ndash; review and editing; L.Z. Data curation, writing \u0026ndash; review and editing; R.Y.K.C. Data curation, writing \u0026ndash; review and editing; S.S. Data curation, writing \u0026ndash; review and editing; W.T. Resources, writing \u0026ndash; review and editing\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eWe gratefully acknowledge Dr. Man-Li Tse, Dr. Joey Shuk-Yan Leung, Dr. Lawrence Chi-Lun Lai, Dr. Jonathan Gabriel Sung, Prof. Sandra Sau Man Chan, Dr. Ivan Man-Ho Wong for their advice from clinical perspectives and support in funding acquisition. We also thank the Hospital Authority for providing data.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eData underlying the results presented in this article cannot be shared publicly because the raw data is confidential and not allowed for sharing in accordance with the prevailing policies of the Hospital Authority of Hong Kong. The data may be requested from Hong Kong Hospital Authority's Central Panel on Administrative Assessment of External Data Requests (https://www3.ha.org.hk/data/Provision/Submission).\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eCarlsson A, Irewall AL, Graipe A, Ulvenstam A, Mooe T, \u0026Ouml;gren J. Long-term risk of major adverse cardiovascular events following ischemic stroke or TIA. Sci Rep. 2023;13(1):8333.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHeart and Stroke Statistics. 2024; \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.heart.org/en/about-us/heart-and-stroke-association-statistics\u003c/span\u003e\u003cspan address=\"https://www.heart.org/en/about-us/heart-and-stroke-association-statistics\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLloyd-Jones D, Adams RJ, Brown TM, et al. Heart Disease and Stroke Statistics\u0026mdash;2010 Update. Circulation. 2010;121(7):e46-e215.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAndersson C, Vasan RS. Epidemiology of cardiovascular disease in young individuals. Nature Reviews Cardiology. 2018;15(4):230\u0026ndash;240.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMasaebi F, Salehi M, Kazemi M, Vahabi N, Azizmohammad Looha M, Zayeri F. Trend analysis of disability adjusted life years due to cardiovascular diseases: results from the global burden of disease study 2019. BMC Public Health. 2021;21(1):1268.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGoldfarb M, De Hert M, Detraux J, et al. Severe Mental Illness and Cardiovascular Disease: JACC State-of-the-Art Review. Journal of the American College of Cardiology. 2022;80(9):918\u0026ndash;933.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSahlberg M, Holm E, Gislason GH, K\u0026oslash;ber L, Torp-Pedersen C, Andersson C. Association of Selected Antipsychotic Agents With Major Adverse Cardiovascular Events and Noncardiovascular Mortality in Elderly Persons. J Am Heart Assoc. 2015;4(9):e001666.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eD'Errico S, Russa R, Maiese A, et al. Atypical antipsychotics and oxidative cardiotoxicity: review of literature and future perspectives to prevent sudden cardiac death. J Geriatr Cardiol. 2021;18(8):663\u0026ndash;685.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDavid M. Taylor AHY, Thomas R. E. Barnes. \u003cem\u003eThe Maudsley Prescribing Guidelines in Psychiatry, 14th Edition.\u003c/em\u003e Wiley-Blackwell; 2021.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSiwek M, Woroń J, Gorostowicz A, Wordliczek J. Adverse effects of interactions between antipsychotics and medications used in the treatment of cardiovascular disorders. Pharmacological Reports. 2020;72(2):350\u0026ndash;359.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchizophrenia and Related Psychoses. In: \u003cem\u003eThe Maudsley Prescribing Guidelines in Psychiatry.\u003c/em\u003e1-224.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWu CS, Tsai YT, Tsai HJ. Antipsychotic Drugs and the Risk of Ventricular Arrhythmia and/or Sudden Cardiac Death: A Nation-wide Case‐Crossover Study. Journal of the American Heart Association. 2015;4(2):e001568.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFriedrich ME, Winkler D, Konstantinidis A, et al. Cardiovascular Adverse Reactions During Antipsychotic Treatment: Results of AMSP, A Drug Surveillance Program Between 1993 and 2013. Int J Neuropsychopharmacol. 2020;23(2):67\u0026ndash;75.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcQueenie R, Nicholl BI, Jani BD, et al. Patterns of multimorbidity and their effects on adverse outcomes in rheumatoid arthritis: a study of 5658 UK Biobank participants. BMJ Open. 2020;10(11):e038829.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDai D, Fernandes J, Sun X, Lupton L, Payne VW, Berk A. Multimorbidity in Atherosclerotic Cardiovascular Disease and Its Associations With Adverse Cardiovascular Events and Healthcare Costs: A Real-World Evidence Study. J Health Econ Outcomes Res. 2024;11(1):75\u0026ndash;85.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHothorn T, Hornik K, Zeileis A. ctree: Conditional inference trees. The comprehensive R archive network. 2015;8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKursa MB, Rudnicki WR. Feature Selection with the Boruta Package. Journal of Statistical Software. 2010;36(11):1\u0026ndash;13.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFriedman J, Hastie T, Tibshirani R. Regularization Paths for Generalized Linear Models via Coordinate Descent. J Stat Softw. 2010;33(1):1\u0026ndash;22.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBreiman L, Friedman, J., Olshen, R.A., \u0026amp; Stone, C.J. \u003cem\u003eClassification and Regression Trees\u003c/em\u003e. 1st ed. ed: Chapman and Hall/CRC; 1984.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIshwaran H, Kogalur UB, Blackstone EH, Lauer MS. Random survival forests. The Annals of Applied Statistics. 2008;2(3):841\u0026ndash;860, 820.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHothorn T, Zeileis A. Partykit: a modular toolkit for recursive partytioning in R. J Mach Learn Res. 2015;16(1):3905\u0026ndash;3909.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFriedman JH. Stochastic gradient boosting. Computational Statistics \u0026amp; Data Analysis. 2002;38(4):367\u0026ndash;378.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKatzman JL, Shaham U, Cloninger A, Bates J, Jiang T, Kluger Y. DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network. BMC Medical Research Methodology. 2018;18(1):24.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen T, Guestrin C. XGBoost: A Scalable Tree Boosting System. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining; 2016; San Francisco, California, USA.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBinder H, Allignol A, Schumacher M, Beyersmann J. Boosting for high-dimensional time-to-event data with competing risks. Bioinformatics. 2009;25(7):890\u0026ndash;896.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHastie T T, R., \u0026amp; Friedman, J. \u003cem\u003eThe Elements of Statistical Learning: Data Mining, Inference, and Prediction\u003c/em\u003e. 2nd ed. Stanford, CA: Stanford University; 2009.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSteyerberg EW, Vergouwe Y. Towards better clinical prediction models: seven steps for development and an ABCD for validation. Eur Heart J. 2014;35(29):1925\u0026ndash;1931.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHassija V, Chamola V, Mahapatra A, et al. Interpreting Black-Box Models: A Review on Explainable Artificial Intelligence. Cognitive Computation. 2024;16(1):45\u0026ndash;74.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBinder H, Allignol A, Schumacher M, Beyersmann J. Boosting for high-dimensional time-to-event data with competing risks. Bioinformatics. 2009;25(7):890\u0026ndash;896.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLipovetsky S, Conklin M. Analysis of regression in game theory approach. Applied Stochastic Models in Business and Industry. 2001;17(4):319\u0026ndash;330.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAltmann A, Toloşi L, Sander O, Lengauer T. Permutation importance: a corrected feature importance measure. Bioinformatics. 2010;26(10):1340\u0026ndash;1347.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBaj G, Gandin I, Scagnetto A, et al. Comparison of discrimination and calibration performance of ECG-based machine learning models for prediction of new-onset atrial fibrillation. BMC Medical Research Methodology. 2023;23(1):169.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePiovani D, Sokou R, Tsantes AG, Vitello AS, Bonovas S. Optimizing Clinical Decision Making with Decision Curve Analysis: Insights for Clinical Investigators. Healthcare. 2023;11(16):2244.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVan Calster B, Wynants L, Verbeek JFM, et al. Reporting and Interpreting Decision Curve Analysis: A Guide for Investigators. Eur Urol. 2018;74(6):796\u0026ndash;804.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePark Y, Bateman BT, Kim DH, et al. Use of haloperidol versus atypical antipsychotics and risk of in-hospital death in patients with acute myocardial infarction: cohort study. BMJ. 2018;360:k1218.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eH\u0026oslash;jlund M, Andersen K, Ernst MT, Correll CU, Hallas J. Use of low-dose quetiapine increases the risk of major adverse cardiovascular events: results from a nationwide active comparator-controlled cohort study. World Psychiatry. 2022;21(3):444\u0026ndash;451.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSullivan MK, Carrero JJ, Jani BD, et al. The presence and impact of multimorbidity clusters on adverse outcomes across the spectrum of kidney function. BMC Med. 2022;20(1):420.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eColes B, Zaccardi F, Hvid C, Davies MJ, Khunti K. Cardiovascular events and mortality in people with type 2 diabetes and multimorbidity: A real-world study of patients followed for up to 19 years. Diabetes, Obesity and Metabolism. 2021;23(1):218\u0026ndash;227.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIshikita A, McIntosh C, Hanneman K, et al. Machine Learning for Prediction of Adverse Cardiovascular Events in Adults With Repaired Tetralogy of Fallot Using Clinical and Cardiovascular Magnetic Resonance Imaging Variables. Circulation: Cardiovascular Imaging. 2023;16(6):e015205.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJain V, Bansal A, Radakovich N, et al. Machine Learning Models to Predict Major Adverse Cardiovascular Events After Orthotopic Liver Transplantation: A Cohort Study. J Cardiothorac Vasc Anesth. 2021;35(7):2063\u0026ndash;2069.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Multimorbidity, antipsychotics, major adverse cardiovascular events, time-to-event prediction, explainable machine learning","lastPublishedDoi":"10.21203/rs.3.rs-5355838/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5355838/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eIntroduction\u003c/strong\u003e:\u003c/p\u003e\n\u003cp\u003eAntipsychotic users are at an elevated risk of major adverse cardiovascular events (MACE) due to many interacting risk factors. However, specific antipsychotic agents, underlying multimorbidity, and chronic medication patterns in relation to MACE are little explored.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAims\u003c/strong\u003e:\u003c/p\u003e\n\u003cp\u003eTo identify patients’ characteristics with increased risk of MACE in people with multimorbidity and using antipsychotics and to develop and evaluate a time-to-event prediction model.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods\u003c/strong\u003e:\u003c/p\u003e\n\u003cp\u003eThis retrospective cohort study utilized electronic health records from public healthcare facilities in Hong Kong. We included MACE-free patients aged 18-65 years with records of two or more chronic health conditions within three years prior to their first antipsychotic use. Baseline characteristics, such as age, sex, chronic disease history, antipsychotic usage history, and drug intake history over the previous year, were considered. The outcome was major adverse cardiovascular events (MACE), which included stroke, acute myocardial infarction (AMI), and cardiovascular-related death (CV death). The dataset was randomly divided into training and validation subsets in a 7:3 ratio based on the initial year of antipsychotic prescription. A Conditional Inference Survival Tree (CISTree) was employed to identify MACE risk groups. Ten machine learning models were trained using 5-fold cross-validation for hyperparameter optimization and validated on the validation set. We conducted time-dependent ROC curve analysis, calibration plots, and decision curve analysis plots to compare the models' discrimination capacity, calibration, and clinical application value, respectively. Time-dependent variable importance, partial dependence plots, and SHAP plots were used to explain the selected model.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e:\u003c/p\u003e\n\u003cp\u003eA total of 26,274 patients were included in the study. The CISTree model identified older patients (\u0026gt;48 years) with chronic kidney disease (CKD), who were using antibacterial and antiplatelet drugs but not taking antidepressants, and without metastatic cancer, as having the highest MACE incidence rate (171.317 per 1,000 person-years; 95% CI: [130.088, 221.467]). The random survival model outperformed the other nine models, identifying age, antidepressant usage, and CKD as the top three significant predictors, consistent with the CISTree model. The survival C-statistics (ranging from 0 to 1, with higher values indicating better predictive precision) for 1-, 3-, and 5-year MACE predictions in the validation cohort were estimated at 0.841, 0.835, and 0.824, respectively.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion\u003c/strong\u003e:\u003c/p\u003e\n\u003cp\u003eWe identified specific high-risk MACE groups among individuals with multimorbidity who started using antipsychotics. Predictions based on these features demonstrated excellent accuracy and have the potential to aid clinical decision-making.\u003c/p\u003e","manuscriptTitle":"Multimorbidity profile and the risk of major cardiovascular events in new antipsychotic users: a time-to-event prediction study using explainable machine learning","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-11-20 11:44:44","doi":"10.21203/rs.3.rs-5355838/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"a57036cb-91ed-449c-96c4-206ba07f5aa0","owner":[],"postedDate":"November 20th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":40093561,"name":"Health sciences/Diseases/Cardiovascular diseases"},{"id":40093562,"name":"Health sciences/Diseases/Psychiatric disorders"},{"id":40093563,"name":"Health sciences/Medical research/Epidemiology"},{"id":40093564,"name":"Health sciences/Health care/Disease prevention"},{"id":40093565,"name":"Health sciences/Health care/Public health"},{"id":40093566,"name":"Biological sciences/Computational biology and bioinformatics/Machine learning"},{"id":40093567,"name":"Biological sciences/Computational biology and bioinformatics/Predictive medicine"}],"tags":[],"updatedAt":"2025-03-14T13:08:53+00:00","versionOfRecord":[],"versionCreatedAt":"2024-11-20 11:44:44","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-5355838","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5355838","identity":"rs-5355838","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}