Predicting ICU mortality in acute respiratory failure using echocardiographic parameters: a multicentre observational machine-learning study

Predicting ICU mortality in acute respiratory failure using echocardiographic parameters: a multicentre observational machine-learning study | 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 Research Article Predicting ICU mortality in acute respiratory failure using echocardiographic parameters: a multicentre observational machine-learning study Stephen Huang, Michelle Chew, philippe vignon, Armand Mekontso-Dessap, and 20 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9396592/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 6 You are reading this latest preprint version Abstract Background Acute respiratory failure (ARF) is associated with high ICU mortality. Early prediction of individual ICU outcome is helpful for personalizing management strategies to improve survival. However, individual prediction requires the consideration of a number of risk factors simultaneously, which may be challenging for the human brain. In this study, we explore if machine learning can predict individual ICU outcome based on early clinical and echocardiographic information. Methods Early clinical and the first echocardiographic data (features) from COVID-19 patients with ARF from two previous studies were combined. The importances of a collection of features (risk factors) were ranked and individual ICU outcome predictions based on these features were made using machine learning (XGBoost) algorithm. Machine learning prediction accuracy metrics were reported and were also compared to clinicians’ predictions on the same dataset. Results While age, PaO2/FiO2 (PF) ratio and MAP ranked the top three in the feature importance list, echocardiographic features (LVEF, RVEDA/LVEDA ratio and LVEDV) also contributed significantly to ICU mortality prediction but to a lesser degree. The relationships between each feature with ICU mortality risk were consistent with early studies but not in simple linear relationships. The performance of machine learning prediction of individual ICU outcome was reasonable with accuracy = 0.71 (ROC AUC = 0.76). While the specificity was high (93%), the sensitivity was poor (21%). Interestingly, the prediction performance was similar to clinicians’ predictions (sensitivity = 30%, specificity = 90%, accuracy = 0.69). Conclusions In patients with ARF from COVID, machine learning (XGBoost) shows promise in predicting individual ICU outcomes using a combination of early clinical and echocardiographic information. The prediction was a complex process and needed to take all the risk factors into consideration – a process which is similar to actual clinical practice. acute respiratory failure COVID-19 echocardiography intensive care machine learning mortality prediction XGBoost Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction Patients with severe SARS-CoV-2 viral infection (COVID-19) can develop acute respiratory failure (ARF) and progress to acute respiratory distress syndrome (ARDS) [1, 2]. ARF patients have high mortality rates, making the prediction of ICU outcomes a critical task for intensivists [3]. Accurate prediction models can assist clinicians in making informed decisions, optimizing resource allocation, and improving patient outcomes. Traditional methods for predicting ICU mortality have relied on clinical surrogates and scoring systems such as the APACHE II (Acute Physiology and Chronic Health Evaluation II), SAPS score, and SOFA (Sequential Organ Failure Assessment) scores [4, 5]. While these methods provide valuable insights, they are often limited by their reliance on static snapshots of patient data and their application at a population level, leading to potential ecological fallacies when predicting individual outcomes. Further, these scores do not include echocardiographic data or their interactions with other clinical variables, which are important because ICU patients with COVID-19 infection may develop ARDS [2] and septic shock, with variable effects on cardiac function [6, 7]. Several echocardiographic features were associated with higher mortality in this population [8, 9]. However, a majority of studies were relatively small and designed only to examine a single predictor for ICU mortality. In this study, we sought to develop a machine learning model to predict individual ICU outcome based on early clinical information and echocardiographic surrogates for ventricular functions as predictors in the model. Extreme gradient boosting (XGBoost), a machine learning algorithm, was used to predict ICU mortality in COVID-19 ARF patients. XGBoost has shown promise in leveraging large datasets and considering multiple interacting features simultaneously [10]. The study integrates early clinical and echocardiographic data from the ECHO-COVID study [11] and the Linköping ECHO dataset, aiming to provide a robust and useful predictive model to aid clinicians in discriminating patients at risk of death. The specific objectives were (1) to identify the most important features that can predict ICU mortality, (2) to examine the relationship between these features and actual ICU mortality, and (3) to use these features to predict individual ICU mortality. Methods Data source Acute respiratory failure patients were retrieved from data collected in the ECHO-COVID study (N = 689), a retrospective observational study involving 14 intensive care units of tertiary teaching hospitals in 8 countries [16], and from a prospective, observational COVID study from Linköping, Sweden (N = 157). Exclusion criteria was PaO2/FiO2 ratio >300 mmHg. Consecutive patients with confirmed SARS-CoV-2 infection between 1 February 2020 and 26 December 2022, admitted to an ICU and had at least one critical care echocardiography (CCE) in ICU, either a transthoracic (TTE) or transesophageal echocardiography (TEE), were included. TTE or TEE were provided as part of routine care. A total of 29 patients were excluded due to lack of data and a further 17 patients were excluded as their PF ratio >300 mmHg. A total of 741 patients were included in this study (Supplementary Figure S1). The original study was registered in ClinicalTrials.gov (no. NCT 04414410) and were approved by the respective ethics committees. Main investigators were part of the European Diploma in Advanced Critical Care Echocardiography (EDEC) expert group of the European Society of Intensive Care Medicine (ESICM). Echocardiography Only the first CCE (TTE or TEE) examination after admission was used for the present study. All studies were recorded and interpreted off-line by international experts in the field, all of whom have EDEC or equivalent qualifications and experience. For the purpose of this study, the following echocardiographic parameters (features) were used as surrogates for left ventricular (LV) function, right ventricular (RV) function, and hemodynamic status: LV ejection fraction (LVEF), RV tricuspid annular plane systolic excursion (TAPSE) and RV-to-LV end-diastolic area ratio (RVEDA/LVEDA) and LV end-diastolic volume (LVEDV). Data analysis Features inclusion The primary outcome (target variable) was ICU mortality. The following features (variables) were selected in modelling: age, sex, body mass index (BMI), pre-existing diseases (cardiac failure, hypertension, lung disease, chronic renal disease (CRD), diabetes), mean arterial pressure (MAP), heart rate (HR)), LVEF, LVEDV, TAPSE, RVEDA/LVEDA and the ratio between partial arterial pressure of oxygen to the fraction of inspired oxygen [PaO2/FiO2 (PF ratio)]. All physiological data were collected during the first CCE study. Data handling and machine learning with XGBoost The training model was built using extra-gradient boosting (XGBoost), a tree-based ensemble supervised learning algorithm [17]. The dataset consisting of 741 patients and was randomly split into a Train set (n = 555) for model training and a Test set (n = 186) for model validation (in a 3:1 ratio). The objective function was binary logistic and the loss function used was classification error rate. Hyperparameters tuning was performed using grid search with 10-fold cross validation. To assist interpretability of XGBoost results, we used SHAP (SHapley Additive exPlanation) values to summarize the results. SHAP values are calculated by comparing a model’s prediction with and without a particular feature present [18,19]. As the target variable (ICU mortality) was coded as binary in this study, SHAP values can be interpreted as the -log[OR] of ICU survival (positive value favors survival and negative favors non-survival). The relative importance of the features in the prediction model were ranked using the mean(|SHAP|) value for each feature. Large mean(|SHAP|) corresponds to greater importance in [20] outcome prediction. The final model was validated using the Test set. Individual’s outcome prediction was based on the total SHAP score which was calculated by summing all the SHAP values for the features for that patient. Prediction accuracy was examined using a confusion matrix and receiver operating characteristic curve (ROC). Comparison to clinician predictions To compare XGBoost prediction to that of clinicians, 5 clinicians with critical care echocardiographic background were asked to predict ICU survival for patients randomly sampled from the same cohort (n = 189). These clinicians were not shown the results (and conclusion) of this study and were blinded from the actual outcomes. The average sensitivity, specificity and accuracy were compared to XGBoost predictions. Statistics Unless otherwise stated, data were reported as median [q1, q3] for continuous data and number (%) for categorical data. Correlations between the features were performed using Pearson’s, Cramer’s V and ANOVA R-squared. Since machine learning is an algorithm, it does not perform statistical hypothesis testing. The associations between the SHAP values and each of the features was fitted by generalized additive model. Statistical tests were 2-sided, and p-values were reported as appropriate. All analysis were performed using R (version 4.3.1), and the main packages used were tidymodels, caret, and shapviz. Results Patient characteristics A total of 741 patients were included from the combined datasets. The demography of the combined study populations is presented in Table 1. There was no difference between the Train and Test sets. Features included and their importance The set of features included in machine learning modeling is shown in Table 1, and their correlations are shown in Figure 1. Most of the features were only weakly correlated with each other (median r = 0.045 [0.005, 0.104]). Maximal positive and negative correlations were observed between TAPSE and LVEF (r = 0.292), and between HR and PF ratio (r = -0.283), respectively. Figure 2A ranks the importances of the features that contributed to the prediction in the Train dataset. When all the features were considered together, cardiac rhythm, history of chronic renal disease (CRD) and lung disease, were deemed unimportant by the model. The two features that contributed most to the model were age and PF ratio, followed by MAP. Echocardiographic parameters LVEF, RVEDA/LVEDA, and LVEDV contributed significantly and similarly, sex contributed the least. Model performance The performance of the model obtained using the Train set was used to predict the outcome in the Test set. The prediction accuracy (for non-survival) was 0.71, with sensitivity = 0.21 and specificity = 0.93 (Figure 2B). The receiver-operating characteristics curve indicates that the model performance is acceptable with AUC of 0.76 (Figure 2B). The effects of the main features with continuous data on ICU survival, i.e. SHAP values, after taking the effects of other features into consideration in the Test dataset are shown in Figure 3. All features exhibited non-linear relationships with ICU outcome. Echocardiographic features displayed complex relationships with ICU survival. Of interest, most of these features had inflection points, around their corresponding “normal” ranges. For example, better survival was observed when MAP was between 80 to 100 mmHg, when LVEDV was around 100 ml, or when RVEDA/LVEDA was around 0.6. Higher LVEF (hyperdynamic LV) seemed to be associated with higher mortality and lower LVEF seemed to be associated with survival. Also displayed in Figure 3 are the data points for five randomly selected patients with age between 69 and 71. While they were approximately the same age, the mortality risks (SHAP) were different. Also, the individuals’ mortality risks for each of the features shown were also different: an individual could have higher mortality risk for one of the features, but lower risk for another. XGBoost predictions of individual outcomes: examples Figure 4 shows the waterfall plots explaining the prediction of individual outcome by XGBoost using the features included in the model. The figure shows the predicted outcome for 6 randomly selected patients: 3 from those who died and 3 who survived. Out of these 6 randomly chosen patients, XGBoost correctly predicted the outcome of 4 patients, and incorrectly predicted 2 patients who died as survived. The features contributing to the prediction differ from patient to patient, as do their relative importance. Comparison of XGBoost vs clinicians’ predictions Figure 5 compares the prediction results of 5 clinicians and XGBoost using 189 patients randomly selected from the same cohort. The prediction accuracy of XGBoost was 0.70 (sensitivity = 0.30 and specificity = 0.92). On the other hand, the mean accuracy of clinicians’ predictions was 0.69 (range = 0.67 to 0.70), mean sensitivity was 0.30 (range = 0.20 to 0.38), and specificity was 0.90 (range = 0.88 to 0.93). The agreement between the clinicians’ predictions was moderate (Fleiss kappa = 0.574 [0.468, 0.68]). Discussion In COVID-19 patients with ARF, we showed that machine learning (XGBoost) was able to predict individual ICU outcome using a combination of early clinical and echocardiographic features with acceptable accuracy, but with low sensitivity. The prediction performance was comparable to clinicians’ predictions. The most useful (important) echocardiographic features were LVEF, RVEDA/LVEDA ratio and LVEDV which contributed similarly and significantly in the prediction model. Relationship between each feature and mortality: mean effects Using XGBoost algorithm, we managed to tease out the relationships between different feature individually and their contributions to ICU mortality in ARF patients. To the best of our knowledge, this is the only study that described these relationships in one single analysis. Our results showed that echocardiographic features displayed different (non-linear) relationships and had complex interactions with other features in predicting ICU mortality. When considered in isolation, the average relationships between each feature on ICU mortality displayed complex non-linear relationships when other features were kept constant. For example, LV volume has long been reported to be associated with mortality, especially in sepsis, albeit with conflicting results [12, 13]. Our results showed that, on average, small LVEDV increased the risk of ICU mortality, consistent with the findings of Furian et al. [13]. In patients with chronic obstructive pulmonary diseases, small LV size has been shown to be an independent predictor of all-cause mortality [14]. This may due to increased pulmonary vascular resistance, hence reduced pulmonary blood flow, leading to underfilling of the LV [15]. In ARDS patients, a post-hoc analysis of the HEMOPRED study demonstrated that hypovolemia was associated with ICU mortality [16]. Since LVEDV is related to intravascular volume, hypovolemia may partly explain our observation. This notion is supported from the relationships seen between other related features with ICU mortality: high heart rate, hyperdynamic LV and low MAP all suggested hypovolemia (Figure 3). Although not directly shown in this study, the association between LV volume and mortality could also be due to severe RV dilatation or distributive shock [17]. LV systolic dysfunction is common in COVID-19 patients. For example, our previous COVID-ECHO study showed that approximately 22% of the patients exhibited LV systolic dysfunction [11]. Other studies report a reduced LVEF in 16% to 34.7% of patients [18, 19]. While LV longitudinal strain has been reported to be associated with higher mortality [20], no association was found between ICU or in-hospital mortality and LVEF in other studies [11, 18, 21]. Therefore, the relationship between reduced LV strain and mortality is inconsistent [22] despite it being more sensitive in detecting subclinical LV dysfunction [23]. We did not have data on LV strain in this study and LVEF was instead used as a global indicator of LV systolic function. According to our model, higher LVEFs had more negative SHAP values contributing to model‐predicted mortality (i.e. higher risk), while lower LVEFs had positive SHAP values (i.e. lower risk). Thus, in our model, ‘hyperdynamic’ LVEF was deleterious, while a low LVEF appeared to be protective. This seems contrary to prior belief that low LVEF is associated with increased mortality. The results in this study also suggest that patients with high LVEF (hyperdynamic LV) were more likely to die consistent with previous findings in COVID-19 [24] and septic patients [25]. In septic shock patients, hyperdynamic LV seems frequent, particularly in the early phase, and may be associated with LV outflow obstruction and increased mortality [26] although the latter was not assessed in the present study. A hyperkinetic LV has previously been demonstrated in COVID-19 patients, either alone or in association with RV dilation and/or ACP [11]. The pending question is whether hyperdynamic LV itself increases mortality or if it is the association with RV dilation. Although XG boost takes into account the codependencies between echocardiographic variables, these findings require confirmation. LV hyperkinesia and outflow obstruction could be precipitated by hypovolemia. In this study, hypovolemia surrogates, namely low MAP, small LVEDV, tachycardia and high LVEF, were all associated with higher mortality (see above). Meanwhile, having RV dilation and LV hyperkinesia increases the mortality in these patients should be considered. One can hypothesize that the restriction of the LV by a dilated RV further enhance the deleterious consequence of uncoupling between LV and aorta in the setting of profound vasoplegia. Overall result is further reduction of stroke volume and potential tissue hypoperfusion. Another question remaining is: was low LVEF associated with higher mortality? From our results, we are unable to confirm or rule out this trend due to a lack of sample size in this range (LVEF < 30%). Similarly, our findings should be interpreted cautiously due to the low numbers of patients at the extremes of LVEF, and that isolated LVEF is not meaningful in the context of the current prediction algorithm that takes into account interactions with other features. As illustrated by examples in Figure 4, features contributing to outcome prediction differ from patient to patient, as do their relative importance. Therefore, ML models such as the present may provide a paradigm shift for prediction in critically ill patients with ARDS. Interestingly, clinicians’ predictions performed similarly to the ML model, raising the question of whether it is able to break down clinical gestalt and intuition into explainable features COVID-19 as well as ARDS patients with RV dysfunction or dilatation displayed higher mortality [27, 28]. In a previous analysis, we demonstrated that age and ACP were related to mortality in the COVID-ECHO cohort [7]. In another analysis of the same dataset, we found mortality seemed different for different RV phenotypes [29]. Along with other studies, this study confirmed that the presence of RV dysfunction, dilation or ACP were associated with higher mortality among patients with COVID-19 infection [29-31]. Prediction of individual outcome Assessing features or clusters of features in isolation does not provide individual prediction, as their relationships change and differ from patient to patient. To illustrate this, note that there was a range (vertical dispersion) of SHAP values for each value of a particular feature. For example, for age around 70, the SHAP value ranged from -0.332 to 0.496 (i.e. mortality differed for similar age). Since age did not completely account for all the variability (i.e. R2 = 0.95) observed, other features might have contributed to the variability. In other words, for a patient of the same age, the SHAP value (survival) also depends on the values of other features. This is illustrated in the randomly selected 5 patients in Figure 3. The results are consistent with heterogeneous findings at population level – if each feature is considered individually, LV size and function as well as RV size and function are associated with ICU mortality to different degrees [24, 29]. Traditional statistical prediction model, while useful in predicting group outcomes, it is limited incapabilty of predicting individual outcome. On the other hand, machine learning has the capacity to predict individual patient’s outcome based on a collection of features (predictors). In this study, we perform prediction on individual outcome using one of the machine learning algorithm - XGBoost. The prediction process takes all the features into account, and the magnitude and direction of any feature may not be the same depending on the values and combination of other features- a process which resembles to real life clinical practice. Despite its superior prediction performance in other areas [32], its prediction performance was only acceptable in this study. Although the specificity was high (0.93), the sensitivity was low (0.21). Interesting, and perhaps surprisingly, this prediction performance was comparable to the average predictions performance made by clinicians, in terms of accuracy, sensitivity and specificity. This is to be expected given the fact that the data used for prediction in this study were collected in early ICU admission stage. Disease progress, treatment options, complications and other subsequent events might change patients’ outcomes, and unless the patient was in the extreme (and clear) case, making early prediction about a particular patient would be a very challenging task. Further, the lack of other relevant data, either due to missingness or not collected, also rendered the prediction less accurate both for the clinicians and the model. Strengths and limitations Apart from the large sample size, one other strength of this study is the use of multiple features (predictors) and their interactions in predicting ICU outcome in ARF patients. In many association studies to date, outcome predictions are often made with one predictor, which is seldom practiced in real life. For example, one would not ignore the age of the patients, comorbidities, and the severity of diseases when making outcome predictions using LVEF. The strength of machine learning is that it takes all the features that are supplied to the model, ranked these features according to their importances then makes predictions, a process akin to real clinical practice. Another strength is the use of XGBoost as the prediction algorithm, a tree-based model which is closer to our thinking process. In many statistical analyses, e.g. t-test, ANOVA and regression method, normal distribution is assumed and in many cases this requirement may not be satisfied. XGBoost on the other hand, makes no assumption about distribution and can handle missing data and outliers, making it one of the most robust algorithms. However, as in many machine learning algorithms, interpretation of model and results can be challenging due to its complexity. Combining two large datasets, COVID-ECHO and Linköping ECHO, offered another benefit: better generalization of the results. That said, our cohort only comprised of COVID-19 ARF patients, while we believe the results can be generalized to non-COVID ARF patients, this may not be true and needs to be confirmed. The biggest limitation of this study is that we only used the first echocardiography study data early clinical data for outcome prediction. Some other relevant outcome-related data were not taken into account, e.g. medications and subsequent complications, and that complete longitudinal data were not available to monitor disease progress – a limitation of retrospective study. This might have affected the accuracy and sensitivity of the predictions. Future research should incorporate prospective validation, include strain imaging, use external multicenter cohorts, and explore whether real-time incorporation of evolving variables (e.g. daily echo or hemodynamic changes) improves mortality prediction. Conclusion We demonstrated that a machine learning model using early echocardiographic and clinical markers can yield individualized risk predictions for ICU mortality in COVID-19 ARF patients, with performance comparable to clinician judgment. Echocardiographic features including LVEDV, LVEF, and RVEDA/LVEDA interact in complex, patient-specific ways to influence risk. Our findings suggest that bedside echocardiography, when combined with interpretable ML models, may enhance early risk stratification for induvial patients. Abbreviations Abbreviation Definition ACP acute cor pulmonale ARDS acute respiratory distress syndrome ARF acute respiratory failure BMI body mass index CCE critical care echocardiography CRD chronic renal disease HR heart rate ICU intensive care unit LV left ventricle/left ventricular LVEDV left ventricular end-diastolic volume LVEF left ventricular ejection fraction MAP mean arterial pressure PF ratio PaO2/FiO2 ratio PSM paradoxical septal motion ROC receiver operating characteristic RV right ventricle/right ventricular RVEDA/LVEDA right-to-left ventricular end-diastolic area ratio SHAP SHapley Additive exPlanation TAPSE tricuspid annular plane systolic excursion TEE transesophageal echocardiography TTE transthoracic echocardiography Declarations Ethics approval and consent to participate The original studies were approved by the respective ethics committees of the participating centers. The ECHO-COVID study was registered in ClinicalTrials.gov (NCT04414410). Requirement for informed consent was handled according to local regulations and ethics approvals. Consent for publication Not applicable. Availability of data and materials The datasets analysed during the current study are available from the corresponding author on reasonable request. Competing interests The authors declare that they have no competing interests Funding No funding Authors' contributions Study design: MS Chew, P Vignon, M Slama Patient inclusion: S Tran, G Prat, M Balik, F Sanfilippo, G Banauch, F Clau-Terre, A Morelli, D De Backer, B Cholley, C Charron, M Goudelin, F Bagate, P Bailly, PB Blixt, P Masi, B Evrard, S Orde, P Mayo, A Vieillard-Baron, M Slama Statistical analysis: S Huang Manuscript drafting: M Slama, MS Chew, S Huang Critical revision and approval of the final manuscript: All authors Acknowledgements Collaborators are listed below: Anne-Marie Welsh (Nepean Hospital, Sydney, Australia), H Didriksson (Linköping University hospital, Sweden), Yoann Zerbib (University hospital of Amiens, Amiens, France), Clément Brault (University hospital of Amiens, Amiens, France), Laetitia Bodénes (CHU La Cavale Blanche, Brest, France), Nicolas Ferrière (CHU La Cavale Blanche, Brest, France), Gabor Zilahi (St George’s University hospital, London, UK), Sue Wright (St George’s University hospital, London, UK), S Clavier (Hôpital Européen Georges Pompidou, AP-HP and Université de Paris, Paris, France), I Ma (Hôpital Européen Georges Pompidou, AP-HP and Université de Paris, Paris, France), JB Rius (Vall d’Hebron University hospital, Barcelona, Spain), JR Palomares (Vall d’Hebron University hospital, Barcelona, Spain), Fernando Piscioneri (Department of Clinical Internal, Anesthesiological and Cardiovascular Sciences, University of Rome, "La Sapienza", Policlinico Umberto Primo), S Giglioli (CHIREC Hospitals, Université Libre de Bruxelles, Brussels, Belgium), Marine Goudelin (University hospital of Limoges, France), Bruno Evrard (University hospital of Limoges, France) References Berlin DA, Gulick RM, Martinez FJ. Severe Covid-19. N Engl J Med. 2020;383:2451-2460. Yang X, Yu Y, Xu J, Shu H, Xia J, Liu H, et al. Clinical course and outcomes of critically ill patients with SARS-CoV-2 pneumonia in Wuhan, China: a single-centered, retrospective, observational study. Lancet Respir Med. 2020;8:475-481. GBD Causes of Death Collaborators. Global burden of 288 causes of death and life expectancy decomposition in 204 countries and territories and 811 subnational locations, 1990-2021: a systematic analysis for the Global Burden of Disease Study 2021. Lancet. 2024;403:2100-2132. Salluh JI, Soares M. ICU severity of illness scores: APACHE, SAPS and MPM. Curr Opin Crit Care. 2014;20:557-565. Vincent JL, de Mendonca A, Cantraine F, Moreno R, Takala J, Suter PM, et al. Use of the SOFA score to assess the incidence of organ dysfunction/failure in intensive care units: results of a multicenter, prospective study. Crit Care Med. 1998;26:1793-1800. Castiello T, Georgiopoulos G, Finocchiaro G, Claudia M, Gianatti A, Delialis D, et al. COVID-19 and myocarditis: a systematic review and overview of current challenges. Heart Fail Rev. 2022;27:251-261. Gul M, Ozyilmaz S, Bastug Gul Z, Kacmaz C, Satilmisoglu MH. Evaluation of cardiac injury with biomarkers and echocardiography after COVID-19 infection. J Physiol Pharmacol. 2022;73. Mosallami Aghili SM, Khoshfetrat M, Asgari A, Arefizadeh R, Mohsenizadeh A, Mousavi SH. Association of echocardiographic findings with in-hospital mortality of COVID-19 patients and their changes in one-month follow-up: a cohort study. Arch Acad Emerg Med. 2022;10:e85. Beyls C, Martin N, Booz T, Viart C, Boisgard S, Daumin C, et al. Prognostic value of acute cor pulmonale in COVID-19-related pneumonia: a prospective study. Front Med (Lausanne). 2022;9:824994. Chen T, Guestrin C. XGBoost: a scalable tree boosting system. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 2016. p. 785-794. Huang S, Vignon P, Mekontso-Dessap A, Tran S, Prat G, Chew M, et al. Echocardiography findings in COVID-19 patients admitted to intensive care units: a multinational observational study (the ECHO-COVID study). Intensive Care Med. 2022;48:667-678. Zanotti Cavazzoni SL, Guglielmi M, Parrillo JE, Walker T, Dellinger RP, Hollenberg SM. Ventricular dilation is associated with improved cardiovascular performance and survival in sepsis. Chest. 2010;138:848-855. Furian T, Aguiar C, Prado K, Ribeiro RV, Becker L, Martinelli N, et al. Ventricular dysfunction and dilation in severe sepsis and septic shock: relation to endothelial function and mortality. J Crit Care. 2012;27:319.e9-319.e15. Abdo M, Watz H, Alter P, Kahnert K, Trudzinski F, Groth EE, et al. Characterization and mortality risk of impaired left ventricular filling in chronic obstructive pulmonary disease. Am J Respir Crit Care Med. 2025;211:477-485. Alter P, Watz H, Kahnert K, Pfeifer M, Randerath WJ, Andreas S, et al. Airway obstruction and lung hyperinflation in COPD are linked to an impaired left ventricular diastolic filling. Respir Med. 2018;137:14-22. Joseph A, Evrard B, Petit M, Goudelin M, Prat G, Slama M, et al. Fluid responsiveness in acute respiratory distress syndrome patients: a post hoc analysis of the HEMOPRED study. Intensive Care Med. 2024;50:1850-1860. Mercado P, Maizel J, Kontar L, Nalos M, Huang S, Orde S, et al. Moderate and severe acute respiratory distress syndrome: hemodynamic and cardiac effects of an open lung strategy with recruitment maneuver analyzed using echocardiography. Crit Care Med. 2018;46:1608-1616. Krishna H, Ryu AJ, Scott CG, Mandale DR, Naqvi TZ, Pellikka PA. Cardiac abnormalities in COVID-19 and relationship to outcome. Mayo Clin Proc. 2021;96:932-942. Jain SS, Liu Q, Raikhelkar J, Fried J, Elias P, Poterucha TJ, et al. Indications for and findings on transthoracic echocardiography in COVID-19. J Am Soc Echocardiogr. 2020;33:1278-1284. Wibowo A, Pranata R, Astuti A, Tiksnadi BB, Martanto E, Martha JW, et al. Left and right ventricular longitudinal strains are associated with poor outcome in COVID-19: a systematic review and meta-analysis. J Intensive Care. 2021;9:9. Park J, Kim Y, Pereira J, Hennessey KC, Faridi KF, McNamara RL, et al. Understanding the role of left and right ventricular strain assessment in patients hospitalized with COVID-19. Am Heart J Plus. 2021;6:100018. Croft LB, Krishnamoorthy P, Ro R, Anastasius M, Zhao W, Buckley S, et al. Abnormal left ventricular global longitudinal strain by speckle tracking echocardiography in COVID-19 patients. Future Cardiol. 2021;17:655-661. Shmueli H, Shah M, Ebinger JE, Nguyen LC, Chernomordik F, Flint N, et al. Left ventricular global longitudinal strain in identifying subclinical myocardial dysfunction among patients hospitalized with COVID-19. Int J Cardiol Heart Vasc. 2021;32:100719. Jansson S, Blixt PJ, Didriksson H, Jonsson C, Andersson H, Hedstrom C, et al. Incidence of acute myocardial injury and its association with left and right ventricular systolic dysfunction in critically ill COVID-19 patients. Ann Intensive Care. 2022;12:56. Sato R, Sanfilippo F, Hasegawa D, Prasitlumkum N, Duggal A, Dugar S. Prevalence and prognosis of hyperdynamic left ventricular systolic function in septic patients: a systematic review and meta-analysis. Ann Intensive Care. 2024;14:22. Chauvet JL, El-Dash S, Delastre O, Bouffandeau B, Jusserand D, Michot JB, et al. Early dynamic left intraventricular obstruction is associated with hypovolemia and high mortality in septic shock patients. Crit Care. 2015;19:262. Paternoster G, Bertini P, Innelli P, Trambaiolo P, Landoni G, Franchi F, et al. Right ventricular dysfunction in patients with COVID-19: a systematic review and meta-analysis. J Cardiothorac Vasc Anesth. 2021;35:3319-3324. Dong D, Zong Y, Li Z, Wang Y, Jing C. Mortality of right ventricular dysfunction in patients with acute respiratory distress syndrome subjected to lung protective ventilation: a systematic review and meta-analysis. Heart Lung. 2021;50:730-735. Huang S, Vieillard-Baron A, Evrard B, Prat G, Chew MS, Balik M, et al. Echocardiography phenotypes of right ventricular involvement in COVID-19 ARDS patients and ICU mortality: post hoc exploratory analysis of repeated data from the ECHO-COVID study. Intensive Care Med. 2023;49:946-956. Li YL, Zheng JB, Jin Y, Tang R, Li M, Xiu CH, et al. Acute right ventricular dysfunction in severe COVID-19 pneumonia. Rev Cardiovasc Med. 2020;21:635-641. Sanchez PA, O'Donnell CT, Francisco N, Santana EJ, Moore AR, Pacheco-Navarro A, et al. Right ventricular dysfunction patterns among patients with COVID-19 in the intensive care unit: a retrospective cohort analysis. Ann Am Thorac Soc. 2023;20:1465-1474. Moore A, Bell M. XGBoost, a novel explainable AI technique, in the prediction of myocardial infarction: a UK Biobank cohort study. Clin Med Insights Cardiol. 2022;16:1-16. Table Table 1. Patient characteristics Characteristic Overall N = 7411 Train N = 5551 Test N = 1861 p-value2 q-value3 Patient Characteristics Patient Characteristics Patient Characteristics Patient Characteristics Patient Characteristics Patient Characteristics Sex 0.3 >0.9 M 518 (70%) 382 (69%) 136 (73%) F 223 (30%) 173 (31%) 50 (27%) Age 65 (56, 73) 66 (56, 73) 64 (56, 71) 0.2 >0.9 BMI 29.0 (25.6, 33.5) 28.7 (25.6, 33.4) 29.4 (25.5, 33.6) 0.6 >0.9 Comorbidities Comorbidities Comorbidities Comorbidities Comorbidities Comorbidities Cardiac failure 100 (15%) 75 (15%) 25 (15%) 0.9 >0.9 Hypertension 411 (56%) 303 (55%) 108 (59%) 0.4 >0.9 Lung disease 145 (20%) 106 (19%) 39 (21%) 0.5 >0.9 Diabetes 218 (30%) 160 (29%) 58 (31%) 0.6 >0.9 Chronic renal disease 60 (8.2%) 48 (8.7%) 12 (6.5%) 0.3 >0.9 Clinical information Clinical information Clinical information Clinical information Clinical information Clinical information HR 85 (71, 100) 85 (70, 100) 88 (72, 100) 0.4 >0.9 Cardiac rhythm 0.7 >0.9 Sinus 660 (92%) 492 (91%) 168 (92%) Atrial fibrillation 61 (8.5%) 47 (8.7%) 14 (7.7%) MAP 82 (72, 94) 83 (72, 95) 80 (72, 92) 0.12 >0.9 PF ratio 122 (89, 170) 121 (86, 167) 130 (96, 175) 0.11 >0.9 Echocardiographic information Echocardiographic information Echocardiographic information Echocardiographic information Echocardiographic information Echocardiographic information LVEF 58 (50, 65) 58 (49, 65) 59 (50, 65) 0.9 >0.9 LVEDV 91 (74, 118) 91 (75, 119) 92 (70, 115) 0.3 >0.9 RVEDA/LVEDA 0.60 (0.50, 0.78) 0.60 (0.50, 0.78) 0.60 (0.50, 0.75) 0.7 >0.9 TAPSE 20.0 (17.0, 24.0) 20.0 (17.0, 24.0) 20.4 (18.0, 24.0) 0.5 >0.9 PSM 229 (32%) 171 (32%) 58 (32%) >0.9 >0.9 ICU outcome ICU outcome ICU outcome ICU outcome ICU outcome ICU outcome ICU length of stay 14 (7, 26) 13 (7, 25) 16 (7, 29) 0.084 >0.9 ICU outome >0.9 >0.9 Died 225 (30%) 168 (30%) 57 (31%) Survived 516 (70%) 387 (70%) 129 (69%) 1n (%); Median (Q1, Q3) 1n (%); Median (Q1, Q3) 1n (%); Median (Q1, Q3) 1n (%); Median (Q1, Q3) 1n (%); Median (Q1, Q3) 1n (%); Median (Q1, Q3) 2Pearson's Chi-squared test; Wilcoxon rank sum test 2Pearson's Chi-squared test; Wilcoxon rank sum test 2Pearson's Chi-squared test; Wilcoxon rank sum test 2Pearson's Chi-squared test; Wilcoxon rank sum test 2Pearson's Chi-squared test; Wilcoxon rank sum test 2Pearson's Chi-squared test; Wilcoxon rank sum test 3Bonferroni correction for multiple testing 3Bonferroni correction for multiple testing 3Bonferroni correction for multiple testing 3Bonferroni correction for multiple testing 3Bonferroni correction for multiple testing 3Bonferroni correction for multiple testing Data are presented as n (%) or median (Q1, Q3). P values were obtained using Pearson's chi-squared test or the Wilcoxon rank-sum test. q values correspond to Bonferroni correction for multiple testing. Abbreviations: BMI, body mass index; HR, heart rate; MAP, mean arterial pressure; PF ratio, PaO2/FiO2 ratio; LVEF, left ventricular ejection fraction; LVEDV, left ventricular end-diastolic volume; RVEDA/LVEDA, right-to-left ventricular end-diastolic area ratio; TAPSE, tricuspid annular plane systolic excursion; PSM, paradoxical septal motion. Additional Declarations No competing interests reported. Supplementary Files FigureS1.jpg Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 28 Apr, 2026 Reviewers agreed at journal 23 Apr, 2026 Reviewers invited by journal 22 Apr, 2026 Editor assigned by journal 15 Apr, 2026 Submission checks completed at journal 15 Apr, 2026 First submitted to journal 12 Apr, 2026 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-9396592","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":631351739,"identity":"a3987587-c405-48ba-9961-c640d2147c35","order_by":0,"name":"Stephen Huang","email":"","orcid":"","institution":"The University of Sydney","correspondingAuthor":false,"prefix":"","firstName":"Stephen","middleName":"","lastName":"Huang","suffix":""},{"id":631351740,"identity":"22667573-29b1-4320-b468-f71d0f914f60","order_by":1,"name":"Michelle Chew","email":"","orcid":"","institution":"Linköping University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Michelle","middleName":"","lastName":"Chew","suffix":""},{"id":631351741,"identity":"1b787eb1-1509-489f-9045-becf13dfdec8","order_by":2,"name":"philippe vignon","email":"","orcid":"","institution":"Centre Hospitalier Universitaire de Limoges","correspondingAuthor":false,"prefix":"","firstName":"philippe","middleName":"","lastName":"vignon","suffix":""},{"id":631351742,"identity":"06f46321-164f-4271-9b89-ca8b4b42aa71","order_by":3,"name":"Armand Mekontso-Dessap","email":"","orcid":"","institution":"Hôpitaux Universitaires Henri-Mondor","correspondingAuthor":false,"prefix":"","firstName":"Armand","middleName":"","lastName":"Mekontso-Dessap","suffix":""},{"id":631351743,"identity":"99571ed3-fb60-4d91-8423-ccab163c82ed","order_by":4,"name":"s tran","email":"","orcid":"","institution":"Hôpital Ambroise-Paré","correspondingAuthor":false,"prefix":"","firstName":"s","middleName":"","lastName":"tran","suffix":""},{"id":631351744,"identity":"7fb14f64-b9ee-43a1-9a3a-487a9f2ba506","order_by":5,"name":"Gwanael Prat","email":"","orcid":"","institution":"Centre Hospitalier Régional Universitaire de Brest","correspondingAuthor":false,"prefix":"","firstName":"Gwanael","middleName":"","lastName":"Prat","suffix":""},{"id":631351745,"identity":"757f83a5-c78d-4fcc-b8ff-b448f0072b3a","order_by":6,"name":"Martin Balik","email":"","orcid":"","institution":"Charles University","correspondingAuthor":false,"prefix":"","firstName":"Martin","middleName":"","lastName":"Balik","suffix":""},{"id":631351747,"identity":"8785652b-beb2-403f-8c74-c98ee30a6a0f","order_by":7,"name":"filippo Sanfilippo","email":"","orcid":"","institution":"Policlinico Universitario di Catania","correspondingAuthor":false,"prefix":"","firstName":"filippo","middleName":"","lastName":"Sanfilippo","suffix":""},{"id":631351749,"identity":"9428b74b-ab34-4e24-bfd2-6d3f87ac1aee","order_by":8,"name":"G Banauch","email":"","orcid":"","institution":"Massachusetts General Hospital","correspondingAuthor":false,"prefix":"","firstName":"G","middleName":"","lastName":"Banauch","suffix":""},{"id":631351750,"identity":"a6101875-5af0-45eb-80c0-94bfc0149619","order_by":9,"name":"Fernando Clau-Terre","email":"","orcid":"","institution":"Vall d'Hebron Hospital Universitari","correspondingAuthor":false,"prefix":"","firstName":"Fernando","middleName":"","lastName":"Clau-Terre","suffix":""},{"id":631351751,"identity":"e1bb91e9-27ff-4125-90a5-d8664ba26dd4","order_by":10,"name":"Andrea Morelli","email":"","orcid":"","institution":"Sapienza University of Rome","correspondingAuthor":false,"prefix":"","firstName":"Andrea","middleName":"","lastName":"Morelli","suffix":""},{"id":631351752,"identity":"d07c9189-38b7-4bdc-8602-9fba47240a65","order_by":11,"name":"Daniel De Backer","email":"","orcid":"","institution":"Université Libre de Bruxelles","correspondingAuthor":false,"prefix":"","firstName":"Daniel","middleName":"","lastName":"De Backer","suffix":""},{"id":631351753,"identity":"9e76eef0-ac2f-4a54-9cf4-4573f962955c","order_by":12,"name":"Bernard Cholley","email":"","orcid":"","institution":"Hôpital Européen Georges-Pompidou","correspondingAuthor":false,"prefix":"","firstName":"Bernard","middleName":"","lastName":"Cholley","suffix":""},{"id":631351754,"identity":"621c8072-136b-47d9-95e1-f8ecb3dd8b54","order_by":13,"name":"Cyril Charron","email":"","orcid":"","institution":"Hôpital Ambroise-Paré","correspondingAuthor":false,"prefix":"","firstName":"Cyril","middleName":"","lastName":"Charron","suffix":""},{"id":631351755,"identity":"a11bad24-5d94-4f9f-8b9d-802eed72a99e","order_by":14,"name":"M Goudelin","email":"","orcid":"","institution":"Linköping University Hospital","correspondingAuthor":false,"prefix":"","firstName":"M","middleName":"","lastName":"Goudelin","suffix":""},{"id":631351756,"identity":"73ab4848-b422-4575-98b7-cb6f7026304c","order_by":15,"name":"F Bagate","email":"","orcid":"","institution":"The University of Sydney","correspondingAuthor":false,"prefix":"","firstName":"F","middleName":"","lastName":"Bagate","suffix":""},{"id":631351757,"identity":"2e8a9a8e-0b55-40f3-a815-564b9e35b5d2","order_by":16,"name":"P Bailly","email":"","orcid":"","institution":"Hôpital Ambroise-Paré","correspondingAuthor":false,"prefix":"","firstName":"P","middleName":"","lastName":"Bailly","suffix":""},{"id":631351758,"identity":"b27c3c8f-1549-463f-888f-0cb988ec17b5","order_by":17,"name":"PB Blixt","email":"","orcid":"","institution":"Centre Hospitalier Régional Universitaire de Brest","correspondingAuthor":false,"prefix":"","firstName":"PB","middleName":"","lastName":"Blixt","suffix":""},{"id":631351759,"identity":"1916b583-9e73-4f0b-999f-4d220e2bd992","order_by":18,"name":"Patrick Masi","email":"","orcid":"","institution":"Centre Hospitalier Universitaire de Limoges","correspondingAuthor":false,"prefix":"","firstName":"Patrick","middleName":"","lastName":"Masi","suffix":""},{"id":631351760,"identity":"b386ed82-6a29-4eb0-99bb-e296a1b9fa95","order_by":19,"name":"Bruno evrard","email":"","orcid":"","institution":"Centre Hospitalier Universitaire de Limoges","correspondingAuthor":false,"prefix":"","firstName":"Bruno","middleName":"","lastName":"evrard","suffix":""},{"id":631351761,"identity":"8a48bc67-7496-4b81-8d9f-b4d4b8ab6208","order_by":20,"name":"Sam Orde","email":"","orcid":"","institution":"The University of Sydney","correspondingAuthor":false,"prefix":"","firstName":"Sam","middleName":"","lastName":"Orde","suffix":""},{"id":631351762,"identity":"ec2593b0-39f5-4f8a-be01-f486c2b522f1","order_by":21,"name":"Paul Mayo","email":"","orcid":"","institution":"Massachusetts General Hospital","correspondingAuthor":false,"prefix":"","firstName":"Paul","middleName":"","lastName":"Mayo","suffix":""},{"id":631351763,"identity":"c435064e-b9de-4bbe-9017-338a9b929f68","order_by":22,"name":"antoine vieillard-Baron","email":"","orcid":"","institution":"Hôpital Ambroise-Paré","correspondingAuthor":false,"prefix":"","firstName":"antoine","middleName":"","lastName":"vieillard-Baron","suffix":""},{"id":631351764,"identity":"9f5dca9b-92be-4211-bfe1-987caf37a3aa","order_by":23,"name":"Michel Slama","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABE0lEQVRIie3PMUvDQBTA8RcCZnma9cBqv8KBEBCEfpUGwanBTRxKOShcl7pX/BIHQnHzwoMs+gEa0kEHnSzERSpY6DUV7JDL7HD/IeFCfrx3AC7XfyzYvhAYeLoEXZ2qZyAsxP8jkE52CepmAhtCCDu/2Uh76L+pJUxb4d1Q01l/HovgKdafDwPAw3rDaS/Kx1Agm2dd6mXvscBEpbfPBHjQrSc+RDM0hLMep0RQ/KgTRftSQwdtiwVf+U9FLks6NUSEC0UraRazEHPlqPidAuRtCDNTPOlbCSe8Klrc3GV2wdNxRieCLVR6IwltpD0aTfOP66ITTs5fX5Z9OhJhcl9+y8GxdbHtrJpvjcDlcrlcza0B0nhkMiFz2qgAAAAASUVORK5CYII=","orcid":"","institution":"Centre Hospitalier Universitaire Amiens-Picardie","correspondingAuthor":true,"prefix":"","firstName":"Michel","middleName":"","lastName":"Slama","suffix":""}],"badges":[],"createdAt":"2026-04-12 19:53:46","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9396592/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9396592/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108401614,"identity":"e0e4b2cc-c032-455d-92bb-f534a0f7005e","added_by":"auto","created_at":"2026-05-04 09:06:22","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":176018,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eCorrelation network plot for the included features\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eBlue color indicates positive correlation, and red color indicates negative correlation. Darker color corresponds to higher correlation. Note that none of the features showed strong correlation to another feature.\u003c/p\u003e","description":"","filename":"Figure1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9396592/v1/c27ba22efec4c40b37a2fba3.jpg"},{"id":108401616,"identity":"c1206c3f-db18-45ce-b5fd-ec3e5b4f3cf7","added_by":"auto","created_at":"2026-05-04 09:06:22","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":831615,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eImportant features and prediction accuracy\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA, ’Feature importance’ refers to how much a feature contributed in building the model (measured by mean(|SHAP|) values). The features are ranked from highest to lowest (higher in the rank means more important). B, Receiver operating characteristic curve (ROC) obtained from XGBoost prediction of patient outcomes in the Test set. The AUC was 0.76. The inset under the ROC is the confusion matrix of the prediction.\u003c/p\u003e","description":"","filename":"Figure2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9396592/v1/5338f70cc6e39f0837617d11.jpg"},{"id":108401617,"identity":"ecf834b7-6256-4033-90a7-7f75c281d8ca","added_by":"auto","created_at":"2026-05-04 09:06:22","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1620523,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eThe effects of features on ICU mortality\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe effects of each feature on ICU mortality are shown, while keeping the other feature constant. Each dot represents one patient. The horizontal dashed line denotes log[OR] = 0 (equal probability of survival and death), and positive SHAP values correspond to higher likelihood of survival. Also shown in each panel are the same 5 patients between 69 and 71 years old randomly chosen (black dots with patient number) to illustrate the same features affect the same individual differently. Triangular black dots indicate the patients who survived and circular black dots indicate patients who died. Some patients did not appear in some panels due to missing data.\u003c/p\u003e","description":"","filename":"Figure3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9396592/v1/e6055b7802578f820450d832.jpg"},{"id":108493320,"identity":"b2635f33-f40b-4164-b57c-43c4ecaabeb5","added_by":"auto","created_at":"2026-05-05 09:59:55","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":1351270,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eXGBoost prediction of individual outcome\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWaterfall plots showing how the XGBoost algorithm predicted the outcomes for 6 randomly selected patients - 3 selected from patients who died and 3 from those survived. Based on the relationships between each of the feature and ICU survival odds, the algorithm considers each feature in turn and assigned the SHAP value for each available feature. Some features might favor survival in a patient (yellow arrows point to right), while some might favour death (dark arrows point left). SHAP values for each feature were added together to yield the total SHAP value for that patient (f(x)), which determines the prediction outcome. E[f(x)] denotes the mean SHAP value (log[OR]) for the Test population, hence f(x) lying to the left represents higher odds of death when compared to the average population while f(x) lying to the right represents higher than average OR for survival. Note that SHAP values for features with missing values (NA) were predicted for that patient based on the model.\u003c/p\u003e","description":"","filename":"Figure4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9396592/v1/f668a82f7e11bc103341c41a.jpg"},{"id":108492715,"identity":"c2f50d37-f25b-4000-a44c-775dec24133f","added_by":"auto","created_at":"2026-05-05 09:58:25","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":636160,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eHeatmap comparing predictions made by clinicians and XGBoost\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe ICU outcomes for a total of 189 patients randomly selected from the data were predicted by XGBoost and 5 clinicians (1 to 5) blinded to the actual outcomes. The actual outcomes are also shown on the right most column (Red = died, Blue = survived, and grey = not predicted by clinicians).\u003c/p\u003e","description":"","filename":"Figure5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9396592/v1/86239e7cf510b4d10447c062.jpg"},{"id":108495261,"identity":"2b2aaf18-d700-41fb-bef3-a8e8a14e0b0b","added_by":"auto","created_at":"2026-05-05 10:09:37","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4436792,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9396592/v1/40b3f0ad-1ccb-4a90-81cd-6c2d472b387c.pdf"},{"id":108493123,"identity":"8abafbfb-f5ce-43ac-b513-1a277ca908ee","added_by":"auto","created_at":"2026-05-05 09:59:26","extension":"jpg","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":163347,"visible":true,"origin":"","legend":"","description":"","filename":"FigureS1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9396592/v1/176d3ff141dcaf36eb528571.jpg"}],"financialInterests":"No competing interests reported.","formattedTitle":"Predicting ICU mortality in acute respiratory failure using echocardiographic parameters: a multicentre observational machine-learning study","fulltext":[{"header":"Introduction","content":"\u003cp\u003ePatients with severe SARS-CoV-2 viral infection (COVID-19) can develop acute respiratory failure (ARF) and progress to acute respiratory distress syndrome (ARDS) [1, 2]. \u0026nbsp;ARF patients have high mortality rates, making the prediction of ICU outcomes a critical task for intensivists [3]. Accurate prediction models can assist clinicians in making informed decisions, optimizing resource allocation, and improving patient outcomes. Traditional methods for predicting ICU mortality have relied on clinical surrogates and scoring systems such as the APACHE II (Acute Physiology and Chronic Health Evaluation II), SAPS score, and SOFA (Sequential Organ Failure Assessment) scores [4, 5]. While these methods provide valuable insights, they are often limited by their reliance on static snapshots of patient data and their application at a population level, leading to potential ecological fallacies when predicting individual outcomes. Further, these scores do not include echocardiographic data or their interactions with other clinical variables, which are \u0026nbsp; important because ICU patients with COVID-19 infection may develop ARDS [2] and septic shock, with variable effects on cardiac function [6, 7]. \u0026nbsp; Several echocardiographic features were associated with higher mortality in this population [8, 9]. \u0026nbsp;However, a majority of studies were relatively small and designed only to examine a single predictor for ICU mortality. \u0026nbsp;In this study, we sought to develop a machine learning model to predict individual ICU outcome based on early clinical information and echocardiographic surrogates for ventricular functions as predictors in the model.\u003c/p\u003e\n\u003cp\u003eExtreme gradient boosting (XGBoost), a machine learning algorithm, was used to predict ICU mortality in COVID-19 ARF patients. \u0026nbsp; XGBoost has shown promise in leveraging large datasets and considering multiple interacting features simultaneously [10]. \u0026nbsp;The study integrates early clinical and echocardiographic data from the ECHO-COVID study [11] and the Link\u0026ouml;ping ECHO dataset, aiming to provide a robust and useful predictive model to aid clinicians in discriminating patients at risk of death. The specific objectives were (1) to identify the most important features that can predict ICU mortality, (2) to examine the relationship between these features and actual ICU mortality, and (3) to use these features to predict individual ICU mortality.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cstrong\u003e\u003cem\u003eData source\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAcute respiratory failure patients were retrieved from data collected in the ECHO-COVID study (N = 689), a retrospective observational study involving 14 intensive care units of tertiary teaching hospitals in 8 countries [16], and from a prospective, observational COVID study from Link\u0026ouml;ping, Sweden (N = 157). \u0026nbsp;Exclusion criteria was PaO2/FiO2 ratio \u0026gt;300 mmHg. \u0026nbsp; Consecutive patients with confirmed SARS-CoV-2 infection between 1 February 2020 and 26 December 2022, admitted to an ICU and had at least one critical care echocardiography (CCE) in ICU, either a transthoracic (TTE) or transesophageal echocardiography (TEE), were included. \u0026nbsp;TTE or TEE were provided as part of routine care. \u0026nbsp;A total of 29 patients were excluded due to lack of data and a further 17 patients were excluded as their PF ratio \u0026gt;300 mmHg. \u0026nbsp; A total of 741 patients were included in this study (Supplementary Figure S1). \u0026nbsp;The original study was registered in ClinicalTrials.gov (no. NCT 04414410) and were approved by the respective ethics committees. Main investigators were part of the European Diploma in Advanced Critical Care Echocardiography (EDEC) expert group of the European Society of Intensive Care Medicine (ESICM).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eEchocardiography\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eOnly the first CCE (TTE or TEE) examination after admission was used for the present study. \u0026nbsp;All studies were recorded and interpreted off-line by international experts in the field, all of whom have EDEC or equivalent qualifications and experience. \u0026nbsp;For the purpose of this study, the following echocardiographic parameters (features) were used as surrogates for left ventricular (LV) function, right ventricular (RV) function, and hemodynamic status: LV ejection fraction (LVEF), RV tricuspid annular plane systolic excursion (TAPSE) and RV-to-LV end-diastolic area ratio (RVEDA/LVEDA) and LV end-diastolic volume (LVEDV).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eData analysis\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eFeatures inclusion\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe primary outcome (target variable) was ICU mortality. \u0026nbsp;The following features (variables) were selected in modelling: age, sex, body mass index (BMI), pre-existing diseases (cardiac failure, hypertension, lung disease, chronic renal disease (CRD), diabetes), mean arterial pressure (MAP), heart rate (HR)), LVEF, LVEDV, TAPSE, RVEDA/LVEDA and the ratio between partial arterial pressure of oxygen to the fraction of inspired oxygen [PaO2/FiO2 (PF ratio)]. \u0026nbsp;All physiological data were collected during the first CCE study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eData handling and machine learning with XGBoost\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe training model was built using extra-gradient boosting (XGBoost), a tree-based ensemble supervised learning algorithm [17]. \u0026nbsp;The dataset consisting of 741 patients and was randomly split into a Train set (n = 555) for model training and a Test set (n = 186) for model validation (in a 3:1 ratio). \u0026nbsp;The objective function was binary logistic and the loss function used was classification error rate. \u0026nbsp;Hyperparameters tuning was performed using grid search with 10-fold cross validation. \u0026nbsp; To assist interpretability of XGBoost results, we used SHAP (SHapley Additive exPlanation) values to summarize the results. \u0026nbsp;SHAP values are calculated by comparing a model\u0026rsquo;s prediction with and without a particular feature present [18,19]. \u0026nbsp;As the target variable (ICU mortality) was coded as binary in this study, SHAP values can be interpreted as the -log[OR] of ICU survival (positive value favors survival and negative favors non-survival). \u0026nbsp;The relative importance of the features in the prediction model were ranked using the mean(|SHAP|) value for each feature. \u0026nbsp;Large mean(|SHAP|) corresponds to greater importance in [20] outcome prediction. \u0026nbsp;The final model was validated using the Test set. \u0026nbsp;Individual\u0026rsquo;s outcome prediction was based on the total SHAP score which was calculated by summing all the SHAP values for the features for that patient. \u0026nbsp; Prediction accuracy was examined using a confusion matrix and receiver operating characteristic curve (ROC).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eComparison to clinician predictions\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo compare XGBoost prediction to that of clinicians, 5 clinicians with critical care echocardiographic background were asked to predict ICU survival for patients randomly sampled from the same cohort (n = 189). \u0026nbsp;These clinicians were not shown the results (and conclusion) of this study and were blinded from the actual outcomes. \u0026nbsp;The average sensitivity, specificity and accuracy were compared to XGBoost predictions.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eStatistics\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eUnless otherwise stated, data were reported as median [q1, q3] for continuous data and number (%) for categorical data. \u0026nbsp;Correlations between the features were performed using Pearson\u0026rsquo;s, Cramer\u0026rsquo;s V and ANOVA R-squared. \u0026nbsp;Since machine learning is an algorithm, it does not perform statistical hypothesis testing. \u0026nbsp;The associations between the SHAP values and each of the features was fitted by generalized additive model. \u0026nbsp;Statistical tests were 2-sided, and p-values were reported as appropriate. \u0026nbsp;All analysis were performed using R (version 4.3.1), and the main packages used were tidymodels, caret, and shapviz.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cstrong\u003e\u003cem\u003ePatient characteristics\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA total of 741 patients were included from the combined datasets. \u0026nbsp;The demography of the combined study populations is presented in Table 1. \u0026nbsp; There was no difference between the Train and Test sets.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eFeatures included and their importance\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe set of features included in machine learning modeling is shown in Table 1, and their correlations are shown in Figure 1. \u0026nbsp;Most of the features were only weakly correlated with each other (median r = 0.045 [0.005, 0.104]). \u0026nbsp;Maximal positive and negative correlations were observed between TAPSE and LVEF (r = 0.292), and between HR and PF ratio (r = -0.283), respectively.\u003c/p\u003e\n\u003cp\u003eFigure 2A ranks the importances of the features that contributed to the prediction in the Train dataset. \u0026nbsp;When all the features were considered together, cardiac rhythm, history of chronic renal disease (CRD) and lung disease, were deemed unimportant by the model. \u0026nbsp; The two features that contributed most to the model were age and PF ratio, followed by MAP. \u0026nbsp; Echocardiographic parameters LVEF, RVEDA/LVEDA, and LVEDV contributed significantly and similarly, sex contributed the least.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eModel performance\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe performance of the model obtained using the Train set was used to predict the outcome in the Test set. \u0026nbsp;The prediction accuracy (for non-survival) was 0.71, with sensitivity = 0.21 and specificity = 0.93 (Figure 2B). \u0026nbsp;The receiver-operating characteristics curve indicates that the model performance is acceptable with AUC of 0.76 (Figure 2B).\u003c/p\u003e\n\u003cp\u003eThe effects of the main features with continuous data on ICU survival, i.e. SHAP values, after taking the effects of other features into consideration in the Test dataset are shown in Figure 3. \u0026nbsp;All features exhibited non-linear relationships with ICU outcome. \u0026nbsp; Echocardiographic features displayed complex relationships with ICU survival. \u0026nbsp;Of interest, most of these features had inflection points, around their corresponding “normal” ranges. \u0026nbsp;For example, better survival was observed when MAP was between 80 to 100 mmHg, when LVEDV was around 100 ml, or when RVEDA/LVEDA was around 0.6. \u0026nbsp;Higher LVEF (hyperdynamic LV) seemed to be associated with higher mortality and lower LVEF seemed to be associated with survival.\u003c/p\u003e\n\u003cp\u003eAlso displayed in Figure 3 are the data points for five randomly selected patients with age between 69 and 71. \u0026nbsp;While they were approximately the same age, the mortality risks (SHAP) were different. \u0026nbsp; Also, the individuals’ mortality risks for each of the features shown were also different: an individual could have higher mortality risk for one of the features, but lower risk for another.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eXGBoost predictions of individual outcomes: examples\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFigure 4 shows the waterfall plots explaining the prediction of individual outcome by XGBoost using the features included in the model. \u0026nbsp;The figure shows the predicted outcome for 6 randomly selected patients: 3 from those who died and 3 who survived. \u0026nbsp;Out of these 6 randomly chosen patients, XGBoost correctly predicted the outcome of 4 patients, and incorrectly predicted 2 patients who died as survived. The features contributing to the prediction differ from patient to patient, as do their relative importance.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eComparison of XGBoost vs clinicians’ predictions\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFigure 5 compares the prediction results of 5 clinicians and XGBoost using 189 patients randomly selected from the same cohort. \u0026nbsp; The prediction accuracy of XGBoost was 0.70 (sensitivity = 0.30 and specificity = 0.92). \u0026nbsp; On the other hand, the mean accuracy of clinicians’ predictions was 0.69 (range = 0.67 to 0.70), mean sensitivity was 0.30 (range = 0.20 to 0.38), and specificity was 0.90 (range = 0.88 to 0.93). \u0026nbsp; The agreement between the clinicians’ predictions was moderate (Fleiss kappa = 0.574 [0.468, 0.68]).\u003c/p\u003e\n\n\n\n\n\n\n\n\n\n\n\n\n\n"},{"header":"Discussion","content":"\u003cp\u003eIn COVID-19 patients with ARF, we showed that machine learning (XGBoost) was able to predict individual ICU outcome using a combination of early clinical and echocardiographic features with acceptable accuracy, but with low sensitivity. \u0026nbsp;The prediction performance was comparable to clinicians’ predictions. \u0026nbsp; The most useful (important) echocardiographic features were LVEF, RVEDA/LVEDA ratio and LVEDV which contributed similarly and significantly in the prediction model.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003e\u003cem\u003eRelationship between each feature and mortality: mean effects\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\u003cp\u003eUsing XGBoost algorithm, we managed to tease out the relationships between different feature individually and their contributions to ICU mortality in ARF patients. \u0026nbsp;To the best of our knowledge, this is the only study that described these relationships in one single analysis. \u0026nbsp; Our results showed that echocardiographic features displayed different (non-linear) relationships and had complex interactions with other features in predicting ICU mortality.\u003c/p\u003e\u003cp\u003eWhen considered in isolation, the average relationships between each feature on ICU mortality displayed complex non-linear relationships when other features were kept constant. \u0026nbsp;For example, LV volume has long been reported to be associated with mortality, especially in sepsis, albeit with conflicting results [12, 13]. \u0026nbsp;Our results showed that, on average, small LVEDV increased the risk of ICU mortality, consistent with the findings of Furian et al. [13]. \u0026nbsp; In patients with chronic obstructive pulmonary diseases, small LV size has been shown to be an independent predictor of all-cause mortality [14]. \u0026nbsp;This may due to increased pulmonary vascular resistance, hence reduced pulmonary blood flow, leading to underfilling of the LV [15]. \u0026nbsp;In ARDS patients, a post-hoc analysis of the HEMOPRED study demonstrated that hypovolemia was associated with ICU mortality [16]. \u0026nbsp;Since LVEDV is related to intravascular volume, hypovolemia may partly explain our observation. \u0026nbsp;This notion is supported from the relationships seen between other related features with ICU mortality: high heart rate, hyperdynamic LV and low MAP all suggested hypovolemia (Figure 3). \u0026nbsp; Although not directly shown in this study, the association between LV volume and mortality could also be due to severe RV dilatation or distributive shock [17].\u003c/p\u003e\u003cp\u003eLV systolic dysfunction is common in COVID-19 patients. \u0026nbsp;For example, our previous COVID-ECHO study showed that approximately 22% of the patients exhibited LV systolic dysfunction [11]. \u0026nbsp;Other studies report a reduced LVEF in 16% to 34.7% of patients [18, 19]. \u0026nbsp; While LV longitudinal strain has been reported to be associated with higher mortality [20], no association was found between ICU or in-hospital mortality and LVEF in other studies [11, 18, 21]. \u0026nbsp; Therefore, the relationship between reduced LV strain and mortality is inconsistent [22] despite it being more sensitive in detecting subclinical LV dysfunction [23]. \u0026nbsp; We did not have data on LV strain in this study and LVEF was instead used as a global indicator of LV systolic function.\u003c/p\u003e\u003cp\u003eAccording to our model, higher LVEFs had more negative SHAP values contributing to model‐predicted mortality (i.e. higher risk), while lower LVEFs had positive SHAP values (i.e. lower risk). Thus, in our model, ‘hyperdynamic’ LVEF was deleterious, while a low LVEF appeared to be protective. \u0026nbsp;This seems contrary to prior belief that low LVEF is associated with increased mortality. The results in this study also suggest that patients with high LVEF (hyperdynamic LV) were more likely to die consistent with previous findings in COVID-19 [24] and septic patients [25]. In septic shock patients, hyperdynamic LV seems frequent, particularly in the early phase, and may be associated with LV outflow obstruction and increased mortality [26] \u0026nbsp; although the latter was not assessed in the present study. A hyperkinetic LV has previously been demonstrated in COVID-19 patients, either alone or in association with RV dilation and/or ACP [11]. The pending question is whether hyperdynamic LV itself increases mortality or if it is the association with RV dilation. Although XG boost takes into account the codependencies between echocardiographic variables, these findings require confirmation. \u0026nbsp;LV hyperkinesia and outflow obstruction could be precipitated by hypovolemia. \u0026nbsp;In this study, hypovolemia surrogates, namely low MAP, small LVEDV, tachycardia and high LVEF, were all associated with higher mortality (see above). \u0026nbsp; \u0026nbsp;Meanwhile, having RV dilation and LV hyperkinesia increases the mortality in these patients should be considered. \u0026nbsp;One can hypothesize that the restriction of the LV by a dilated RV further enhance the deleterious consequence of uncoupling between LV and aorta in the setting of profound vasoplegia. Overall result is further reduction of stroke volume and potential tissue hypoperfusion. \u0026nbsp; Another question remaining is: was low LVEF associated with higher mortality? \u0026nbsp;From our results, we are unable to confirm or rule out this trend due to a lack of sample size in this range (LVEF \u0026lt; 30%). \u0026nbsp;Similarly, our findings should be interpreted cautiously due to the low numbers of patients at the extremes of LVEF, and that isolated LVEF is not meaningful in the context of the current prediction algorithm that takes into account interactions with other features. As illustrated by examples in Figure 4, features contributing to outcome prediction differ from patient to patient, as do their relative importance. Therefore, ML models such as the present may provide a paradigm shift for prediction in critically ill patients with ARDS. Interestingly, clinicians’ predictions performed similarly to the ML model, raising the question of whether it is able to break down clinical gestalt and intuition into explainable features\u003c/p\u003e\u003cp\u003eCOVID-19 as well as ARDS patients with RV dysfunction or dilatation displayed higher mortality [27, 28]. \u0026nbsp; In a previous analysis, we demonstrated that age and ACP were related to mortality in the COVID-ECHO cohort [7]. In another analysis of the same dataset, we found mortality seemed different for different RV phenotypes [29]. Along with other studies, this study confirmed that the presence of RV dysfunction, dilation or ACP were associated with higher mortality among patients with COVID-19 infection [29-31].\u003c/p\u003e\u003cp\u003e\u003cstrong\u003e\u003cem\u003ePrediction of individual outcome\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\u003cp\u003eAssessing features or clusters of features in isolation does not provide individual prediction, as their relationships change and differ from patient to patient. \u0026nbsp;To illustrate this, note that there was a range (vertical dispersion) of SHAP values for each value of a particular feature. \u0026nbsp;For example, for age around 70, the SHAP value ranged from -0.332 to 0.496 (i.e. mortality differed for similar age). \u0026nbsp; Since age did not completely account for all the variability (i.e. R2 = 0.95) observed, other features might have contributed to the variability. \u0026nbsp; In other words, for a patient of the same age, the SHAP value (survival) also depends on the values of other features. \u0026nbsp;This is illustrated in the randomly selected 5 patients in Figure 3. \u0026nbsp; The results are consistent with heterogeneous findings at population level – if each feature is considered individually, LV size and function as well as RV size and function are associated with ICU mortality to different degrees [24, 29].\u003c/p\u003e\u003cp\u003eTraditional statistical prediction model, while useful in predicting group outcomes, it is limited incapabilty of predicting individual outcome. \u0026nbsp;On the other hand, machine learning has the capacity to predict individual patient’s outcome based on a collection of features (predictors). \u0026nbsp;In this study, we perform prediction on individual outcome using one of the machine learning algorithm - XGBoost. \u0026nbsp; The prediction process takes all the features into account, and the magnitude and direction of any feature may not be the same depending on the values and combination of other features- a process which resembles to real life clinical practice. \u0026nbsp;Despite its superior prediction performance in other areas [32], its prediction performance was only acceptable in this study. \u0026nbsp; Although the specificity was high (0.93), the sensitivity was low (0.21). \u0026nbsp;Interesting, and perhaps surprisingly, this prediction performance was comparable to the average predictions performance made by clinicians, in terms of accuracy, sensitivity and specificity. \u0026nbsp; This is to be expected given the fact that the data used for prediction in this study were collected in early ICU admission stage. \u0026nbsp;Disease progress, treatment options, complications and other subsequent events might change patients’ outcomes, and unless the patient was in the extreme (and clear) case, making early prediction about a particular patient would be a very challenging task. \u0026nbsp;Further, the lack of other relevant data, either due to missingness or not collected, also rendered the prediction less accurate both for the clinicians and the model.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003e\u003cem\u003eStrengths and limitations\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\u003cp\u003eApart from the large sample size, one other strength of this study is the use of multiple features (predictors) and their interactions in predicting ICU outcome in ARF patients. \u0026nbsp; In many association studies to date, outcome predictions are often made with one predictor, which is seldom practiced in real life. \u0026nbsp;For example, one would not ignore the age of the patients, comorbidities, and the severity of diseases when making outcome predictions using LVEF. \u0026nbsp;The strength of machine learning is that it takes all the features that are supplied to the model, ranked these features according to their importances then makes predictions, a process akin to real clinical practice. \u0026nbsp;Another strength is the use of XGBoost as the prediction algorithm, a tree-based model which is closer to our thinking process. \u0026nbsp;In many statistical analyses, e.g. t-test, ANOVA and regression method, normal distribution is assumed and in many cases this requirement may not be satisfied. \u0026nbsp;XGBoost on the other hand, makes no assumption about distribution and can handle missing data and outliers, making it one of the most robust algorithms. \u0026nbsp; However, as in many machine learning algorithms, interpretation of model and results can be challenging due to its complexity. \u0026nbsp;Combining two large datasets, COVID-ECHO and Linköping ECHO, offered another benefit: better generalization of the results. \u0026nbsp;That said, our cohort only comprised of COVID-19 ARF patients, while we believe the results can be generalized to non-COVID ARF patients, this may not be true and needs to be confirmed.\u003c/p\u003e\u003cp\u003eThe biggest limitation of this study is that we only used the first echocardiography study data early clinical data for outcome prediction. \u0026nbsp;Some other relevant outcome-related data were not taken into account, e.g. medications and subsequent complications, and that complete longitudinal data were not available to monitor disease progress – a limitation of retrospective study. \u0026nbsp;This might have affected the accuracy and sensitivity of the predictions. \u0026nbsp; Future research should incorporate prospective validation, include strain imaging, use external multicenter cohorts, and explore whether real-time incorporation of evolving variables (e.g. daily echo or hemodynamic changes) improves mortality prediction.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eWe demonstrated that a machine learning model using early echocardiographic and clinical markers can yield individualized risk predictions for ICU mortality in COVID-19 ARF patients, with performance comparable to clinician judgment. Echocardiographic features including LVEDV, LVEF, and RVEDA/LVEDA interact in complex, patient-specific ways to influence risk. Our findings suggest that bedside echocardiography, when combined with interpretable ML models, may enhance early risk stratification for induvial patients.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAbbreviation\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDefinition\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eACP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eacute cor pulmonale\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eARDS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eacute respiratory distress syndrome\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eARF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eacute respiratory failure\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eBMI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003ebody mass index\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eCCE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003ecritical care echocardiography\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eCRD\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003echronic renal disease\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eHR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eheart rate\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eICU\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eintensive care unit\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eLV\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eleft ventricle/left ventricular\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eLVEDV\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eleft ventricular end-diastolic volume\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eLVEF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eleft ventricular ejection fraction\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eMAP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003emean arterial pressure\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003ePF ratio\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003ePaO2/FiO2 ratio\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003ePSM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eparadoxical septal motion\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eROC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003ereceiver operating characteristic\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eRV\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eright ventricle/right ventricular\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eRVEDA/LVEDA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eright-to-left ventricular end-diastolic area ratio\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eSHAP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eSHapley Additive exPlanation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eTAPSE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003etricuspid annular plane systolic excursion\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eTEE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003etransesophageal echocardiography\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003eTTE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 314px;\"\u003e\n \u003cp\u003etransthoracic echocardiography\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003e\u003cem\u003eEthics approval and consent to participate\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe original studies were approved by the respective ethics committees of the participating centers. The ECHO-COVID study was registered in ClinicalTrials.gov (NCT04414410). Requirement for informed consent was handled according to local regulations and ethics approvals.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eConsent for publication\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eAvailability of data and materials\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets analysed during the current study are available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eCompeting interests\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eFunding\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNo funding\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eAuthors\u0026apos; contributions\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eStudy design: MS Chew, P Vignon, M Slama\u003c/p\u003e\n\u003cp\u003ePatient inclusion: S Tran, G Prat, M Balik, F Sanfilippo, G Banauch, F Clau-Terre, A Morelli, D De Backer, B Cholley, C Charron, M Goudelin, F Bagate, P Bailly, PB Blixt, P Masi, B Evrard, S Orde, P Mayo, A Vieillard-Baron, M Slama\u003c/p\u003e\n\u003cp\u003eStatistical analysis: S Huang\u003c/p\u003e\n\u003cp\u003eManuscript drafting: M Slama, MS Chew, S Huang\u003c/p\u003e\n\u003cp\u003eCritical revision and approval of the final manuscript: All authors\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eAcknowledgements\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eCollaborators are listed below:\u003cbr\u003e\u0026nbsp;Anne-Marie Welsh (Nepean Hospital, Sydney, Australia), H Didriksson (Link\u0026ouml;ping University hospital, Sweden), Yoann Zerbib (University hospital of Amiens, Amiens, France), Cl\u0026eacute;ment Brault (University hospital of Amiens, Amiens, France), Laetitia Bod\u0026eacute;nes (CHU La Cavale Blanche, Brest, France), Nicolas Ferri\u0026egrave;re (CHU La Cavale Blanche, Brest, France), Gabor Zilahi (St George\u0026rsquo;s University hospital, London, UK), Sue Wright (St George\u0026rsquo;s University hospital, London, UK), S Clavier (H\u0026ocirc;pital Europ\u0026eacute;en Georges Pompidou, AP-HP and Universit\u0026eacute; de Paris, Paris, France), I Ma (H\u0026ocirc;pital Europ\u0026eacute;en Georges Pompidou, AP-HP and Universit\u0026eacute; de Paris, Paris, France), JB Rius (Vall d\u0026rsquo;Hebron University hospital, Barcelona, Spain), JR Palomares (Vall d\u0026rsquo;Hebron University hospital, Barcelona, Spain), Fernando Piscioneri (Department of Clinical Internal, Anesthesiological and Cardiovascular Sciences, University of Rome, \u0026quot;La Sapienza\u0026quot;, Policlinico Umberto Primo), S Giglioli (CHIREC Hospitals, Universit\u0026eacute; Libre de Bruxelles, Brussels, Belgium), Marine Goudelin (University hospital of Limoges, France), Bruno Evrard (University hospital of Limoges, France) \u0026nbsp;\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eBerlin DA, Gulick RM, Martinez FJ. Severe Covid-19. N Engl J Med. 2020;383:2451-2460.\u003c/li\u003e\n\u003cli\u003eYang X, Yu Y, Xu J, Shu H, Xia J, Liu H, et al. Clinical course and outcomes of critically ill patients with SARS-CoV-2 pneumonia in Wuhan, China: a single-centered, retrospective, observational study. Lancet Respir Med. 2020;8:475-481.\u003c/li\u003e\n\u003cli\u003eGBD Causes of Death Collaborators. Global burden of 288 causes of death and life expectancy decomposition in 204 countries and territories and 811 subnational locations, 1990-2021: a systematic analysis for the Global Burden of Disease Study 2021. Lancet. 2024;403:2100-2132.\u003c/li\u003e\n\u003cli\u003eSalluh JI, Soares M. ICU severity of illness scores: APACHE, SAPS and MPM. Curr Opin Crit Care. 2014;20:557-565.\u003c/li\u003e\n\u003cli\u003eVincent JL, de Mendonca A, Cantraine F, Moreno R, Takala J, Suter PM, et al. Use of the SOFA score to assess the incidence of organ dysfunction/failure in intensive care units: results of a multicenter, prospective study. Crit Care Med. 1998;26:1793-1800.\u003c/li\u003e\n\u003cli\u003eCastiello T, Georgiopoulos G, Finocchiaro G, Claudia M, Gianatti A, Delialis D, et al. COVID-19 and myocarditis: a systematic review and overview of current challenges. Heart Fail Rev. 2022;27:251-261.\u003c/li\u003e\n\u003cli\u003eGul M, Ozyilmaz S, Bastug Gul Z, Kacmaz C, Satilmisoglu MH. Evaluation of cardiac injury with biomarkers and echocardiography after COVID-19 infection. J Physiol Pharmacol. 2022;73.\u003c/li\u003e\n\u003cli\u003eMosallami Aghili SM, Khoshfetrat M, Asgari A, Arefizadeh R, Mohsenizadeh A, Mousavi SH. Association of echocardiographic findings with in-hospital mortality of COVID-19 patients and their changes in one-month follow-up: a cohort study. Arch Acad Emerg Med. 2022;10:e85.\u003c/li\u003e\n\u003cli\u003eBeyls C, Martin N, Booz T, Viart C, Boisgard S, Daumin C, et al. Prognostic value of acute cor pulmonale in COVID-19-related pneumonia: a prospective study. Front Med (Lausanne). 2022;9:824994.\u003c/li\u003e\n\u003cli\u003eChen T, Guestrin C. XGBoost: a scalable tree boosting system. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 2016. p. 785-794.\u003c/li\u003e\n\u003cli\u003eHuang S, Vignon P, Mekontso-Dessap A, Tran S, Prat G, Chew M, et al. Echocardiography findings in COVID-19 patients admitted to intensive care units: a multinational observational study (the ECHO-COVID study). Intensive Care Med. 2022;48:667-678.\u003c/li\u003e\n\u003cli\u003eZanotti Cavazzoni SL, Guglielmi M, Parrillo JE, Walker T, Dellinger RP, Hollenberg SM. Ventricular dilation is associated with improved cardiovascular performance and survival in sepsis. Chest. 2010;138:848-855.\u003c/li\u003e\n\u003cli\u003eFurian T, Aguiar C, Prado K, Ribeiro RV, Becker L, Martinelli N, et al. Ventricular dysfunction and dilation in severe sepsis and septic shock: relation to endothelial function and mortality. J Crit Care. 2012;27:319.e9-319.e15.\u003c/li\u003e\n\u003cli\u003eAbdo M, Watz H, Alter P, Kahnert K, Trudzinski F, Groth EE, et al. Characterization and mortality risk of impaired left ventricular filling in chronic obstructive pulmonary disease. Am J Respir Crit Care Med. 2025;211:477-485.\u003c/li\u003e\n\u003cli\u003eAlter P, Watz H, Kahnert K, Pfeifer M, Randerath WJ, Andreas S, et al. Airway obstruction and lung hyperinflation in COPD are linked to an impaired left ventricular diastolic filling. Respir Med. 2018;137:14-22.\u003c/li\u003e\n\u003cli\u003eJoseph A, Evrard B, Petit M, Goudelin M, Prat G, Slama M, et al. Fluid responsiveness in acute respiratory distress syndrome patients: a post hoc analysis of the HEMOPRED study. Intensive Care Med. 2024;50:1850-1860.\u003c/li\u003e\n\u003cli\u003eMercado P, Maizel J, Kontar L, Nalos M, Huang S, Orde S, et al. Moderate and severe acute respiratory distress syndrome: hemodynamic and cardiac effects of an open lung strategy with recruitment maneuver analyzed using echocardiography. Crit Care Med. 2018;46:1608-1616.\u003c/li\u003e\n\u003cli\u003eKrishna H, Ryu AJ, Scott CG, Mandale DR, Naqvi TZ, Pellikka PA. Cardiac abnormalities in COVID-19 and relationship to outcome. Mayo Clin Proc. 2021;96:932-942.\u003c/li\u003e\n\u003cli\u003eJain SS, Liu Q, Raikhelkar J, Fried J, Elias P, Poterucha TJ, et al. Indications for and findings on transthoracic echocardiography in COVID-19. J Am Soc Echocardiogr. 2020;33:1278-1284.\u003c/li\u003e\n\u003cli\u003eWibowo A, Pranata R, Astuti A, Tiksnadi BB, Martanto E, Martha JW, et al. Left and right ventricular longitudinal strains are associated with poor outcome in COVID-19: a systematic review and meta-analysis. J Intensive Care. 2021;9:9.\u003c/li\u003e\n\u003cli\u003ePark J, Kim Y, Pereira J, Hennessey KC, Faridi KF, McNamara RL, et al. Understanding the role of left and right ventricular strain assessment in patients hospitalized with COVID-19. Am Heart J Plus. 2021;6:100018.\u003c/li\u003e\n\u003cli\u003eCroft LB, Krishnamoorthy P, Ro R, Anastasius M, Zhao W, Buckley S, et al. Abnormal left ventricular global longitudinal strain by speckle tracking echocardiography in COVID-19 patients. Future Cardiol. 2021;17:655-661.\u003c/li\u003e\n\u003cli\u003eShmueli H, Shah M, Ebinger JE, Nguyen LC, Chernomordik F, Flint N, et al. Left ventricular global longitudinal strain in identifying subclinical myocardial dysfunction among patients hospitalized with COVID-19. Int J Cardiol Heart Vasc. 2021;32:100719.\u003c/li\u003e\n\u003cli\u003eJansson S, Blixt PJ, Didriksson H, Jonsson C, Andersson H, Hedstrom C, et al. Incidence of acute myocardial injury and its association with left and right ventricular systolic dysfunction in critically ill COVID-19 patients. Ann Intensive Care. 2022;12:56.\u003c/li\u003e\n\u003cli\u003eSato R, Sanfilippo F, Hasegawa D, Prasitlumkum N, Duggal A, Dugar S. Prevalence and prognosis of hyperdynamic left ventricular systolic function in septic patients: a systematic review and meta-analysis. Ann Intensive Care. 2024;14:22.\u003c/li\u003e\n\u003cli\u003eChauvet JL, El-Dash S, Delastre O, Bouffandeau B, Jusserand D, Michot JB, et al. Early dynamic left intraventricular obstruction is associated with hypovolemia and high mortality in septic shock patients. Crit Care. 2015;19:262.\u003c/li\u003e\n\u003cli\u003ePaternoster G, Bertini P, Innelli P, Trambaiolo P, Landoni G, Franchi F, et al. Right ventricular dysfunction in patients with COVID-19: a systematic review and meta-analysis. J Cardiothorac Vasc Anesth. 2021;35:3319-3324.\u003c/li\u003e\n\u003cli\u003eDong D, Zong Y, Li Z, Wang Y, Jing C. Mortality of right ventricular dysfunction in patients with acute respiratory distress syndrome subjected to lung protective ventilation: a systematic review and meta-analysis. Heart Lung. 2021;50:730-735.\u003c/li\u003e\n\u003cli\u003eHuang S, Vieillard-Baron A, Evrard B, Prat G, Chew MS, Balik M, et al. Echocardiography phenotypes of right ventricular involvement in COVID-19 ARDS patients and ICU mortality: post hoc exploratory analysis of repeated data from the ECHO-COVID study. Intensive Care Med. 2023;49:946-956.\u003c/li\u003e\n\u003cli\u003eLi YL, Zheng JB, Jin Y, Tang R, Li M, Xiu CH, et al. Acute right ventricular dysfunction in severe COVID-19 pneumonia. Rev Cardiovasc Med. 2020;21:635-641.\u003c/li\u003e\n\u003cli\u003eSanchez PA, O\u0026apos;Donnell CT, Francisco N, Santana EJ, Moore AR, Pacheco-Navarro A, et al. Right ventricular dysfunction patterns among patients with COVID-19 in the intensive care unit: a retrospective cohort analysis. Ann Am Thorac Soc. 2023;20:1465-1474.\u003c/li\u003e\n\u003cli\u003eMoore A, Bell M. XGBoost, a novel explainable AI technique, in the prediction of myocardial infarction: a UK Biobank cohort study. Clin Med Insights Cardiol. 2022;16:1-16.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Table","content":"\u003cp\u003e\u003cstrong\u003eTable 1. Patient characteristics\u003c/strong\u003e\u003c/p\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eCharacteristic\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eOverall N = 7411\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eTrain N = 5551\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eTest N = 1861\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003ep-value2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eq-value3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003ePatient Characteristics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003ePatient Characteristics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003ePatient Characteristics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003ePatient Characteristics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003ePatient Characteristics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003ePatient Characteristics\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eSex\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e0.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026gt;0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e518 (70%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e382 (69%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e136 (73%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e223 (30%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e173 (31%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e50 (27%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eAge\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e65 (56, 73)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e66 (56, 73)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e64 (56, 71)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e0.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026gt;0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eBMI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e29.0 (25.6, 33.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e28.7 (25.6, 33.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e29.4 (25.5, 33.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e0.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026gt;0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eComorbidities\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eComorbidities\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eComorbidities\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eComorbidities\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eComorbidities\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eComorbidities\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eCardiac failure\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e100 (15%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e75 (15%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e25 (15%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026gt;0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eHypertension\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e411 (56%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e303 (55%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e108 (59%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e0.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026gt;0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eLung disease\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e145 (20%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e106 (19%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e39 (21%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e0.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026gt;0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eDiabetes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e218 (30%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e160 (29%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e58 (31%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e0.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026gt;0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eChronic renal disease\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e60 (8.2%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e48 (8.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e12 (6.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e0.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026gt;0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eClinical information\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eClinical information\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eClinical information\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eClinical information\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eClinical information\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eClinical information\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eHR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e85 (71, 100)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e85 (70, 100)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e88 (72, 100)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e0.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026gt;0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eCardiac rhythm\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e0.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026gt;0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eSinus\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e660 (92%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e492 (91%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e168 (92%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eAtrial fibrillation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e61 (8.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e47 (8.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e14 (7.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eMAP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e82 (72, 94)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e83 (72, 95)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e80 (72, 92)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e0.12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026gt;0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003ePF ratio\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e122 (89, 170)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e121 (86, 167)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e130 (96, 175)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e0.11\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026gt;0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eEchocardiographic information\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eEchocardiographic information\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eEchocardiographic information\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eEchocardiographic information\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eEchocardiographic information\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eEchocardiographic information\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eLVEF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e58 (50, 65)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e58 (49, 65)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e59 (50, 65)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026gt;0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eLVEDV\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e91 (74, 118)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e91 (75, 119)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e92 (70, 115)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e0.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026gt;0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eRVEDA/LVEDA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e0.60 (0.50, 0.78)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e0.60 (0.50, 0.78)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e0.60 (0.50, 0.75)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e0.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026gt;0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eTAPSE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e20.0 (17.0, 24.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e20.0 (17.0, 24.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e20.4 (18.0, 24.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e0.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026gt;0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003ePSM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e229 (32%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e171 (32%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e58 (32%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026gt;0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026gt;0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eICU outcome\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eICU outcome\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eICU outcome\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eICU outcome\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eICU outcome\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eICU outcome\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eICU length of stay\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e14 (7, 26)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e13 (7, 25)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e16 (7, 29)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e0.084\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026gt;0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eICU outome\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026gt;0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026gt;0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eDied\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e225 (30%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e168 (30%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e57 (31%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003eSurvived\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e516 (70%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e387 (70%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e129 (69%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e1n (%); Median (Q1, Q3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e1n (%); Median (Q1, Q3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e1n (%); Median (Q1, Q3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e1n (%); Median (Q1, Q3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e1n (%); Median (Q1, Q3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e1n (%); Median (Q1, Q3)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e2Pearson\u0026apos;s Chi-squared test; Wilcoxon rank sum test\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e2Pearson\u0026apos;s Chi-squared test; Wilcoxon rank sum test\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e2Pearson\u0026apos;s Chi-squared test; Wilcoxon rank sum test\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e2Pearson\u0026apos;s Chi-squared test; Wilcoxon rank sum test\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e2Pearson\u0026apos;s Chi-squared test; Wilcoxon rank sum test\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e2Pearson\u0026apos;s Chi-squared test; Wilcoxon rank sum test\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e3Bonferroni correction for multiple testing\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e3Bonferroni correction for multiple testing\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e3Bonferroni correction for multiple testing\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e3Bonferroni correction for multiple testing\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e3Bonferroni correction for multiple testing\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e3Bonferroni correction for multiple testing\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eData are presented as n (%) or median (Q1, Q3). P values were obtained using Pearson\u0026apos;s chi-squared test or the Wilcoxon rank-sum test. q values correspond to Bonferroni correction for multiple testing. Abbreviations: BMI, body mass index; HR, heart rate; MAP, mean arterial pressure; PF ratio, PaO2/FiO2 ratio; LVEF, left ventricular ejection fraction; LVEDV, left ventricular end-diastolic volume; RVEDA/LVEDA, right-to-left ventricular end-diastolic area ratio; TAPSE, tricuspid annular plane systolic excursion; PSM, paradoxical septal motion.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"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":"critical-care","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"cric","sideBox":"Learn more about [Critical Care](http://ccforum.biomedcentral.com/)","snPcode":"13054","submissionUrl":"https://submission.nature.com/new-submission/13054/3","title":"Critical Care","twitterHandle":"@Crit_Care","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"acute respiratory failure, COVID-19, echocardiography, intensive care, machine learning, mortality prediction, XGBoost","lastPublishedDoi":"10.21203/rs.3.rs-9396592/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9396592/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAcute respiratory failure (ARF) is associated with high ICU mortality. Early prediction of individual ICU outcome is helpful for personalizing management strategies to improve survival. However, individual prediction requires the consideration of a number of risk factors simultaneously, which may be challenging for the human brain. In this study, we explore if machine learning can predict individual ICU outcome based on early clinical and echocardiographic information.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eEarly clinical and the first echocardiographic data (features) from COVID-19 patients with ARF from two previous studies were combined. The importances of a collection of features (risk factors) were ranked and individual ICU outcome predictions based on these features were made using machine learning (XGBoost) algorithm. Machine learning prediction accuracy metrics were reported and were also compared to clinicians’ predictions on the same dataset.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWhile age, PaO2/FiO2 (PF) ratio and MAP ranked the top three in the feature importance list, echocardiographic features (LVEF, RVEDA/LVEDA ratio and LVEDV) also contributed significantly to ICU mortality prediction but to a lesser degree. The relationships between each feature with ICU mortality risk were consistent with early studies but not in simple linear relationships. The performance of machine learning prediction of individual ICU outcome was reasonable with accuracy = 0.71 (ROC AUC = 0.76). While the specificity was high (93%), the sensitivity was poor (21%). Interestingly, the prediction performance was similar to clinicians’ predictions (sensitivity = 30%, specificity = 90%, accuracy = 0.69).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn patients with ARF from COVID, machine learning (XGBoost) shows promise in predicting individual ICU outcomes using a combination of early clinical and echocardiographic information. The prediction was a complex process and needed to take all the risk factors into consideration – a process which is similar to actual clinical practice.\u003c/p\u003e","manuscriptTitle":"Predicting ICU mortality in acute respiratory failure using echocardiographic parameters: a multicentre observational machine-learning study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-04 09:06:17","doi":"10.21203/rs.3.rs-9396592/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2026-04-28T10:13:17+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"95202267519772255114088634703907229753","date":"2026-04-23T07:07:38+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-04-22T05:55:16+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-15T10:01:45+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-15T10:00:46+00:00","index":"","fulltext":""},{"type":"submitted","content":"Critical Care","date":"2026-04-12T19:45:37+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"critical-care","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"cric","sideBox":"Learn more about [Critical Care](http://ccforum.biomedcentral.com/)","snPcode":"13054","submissionUrl":"https://submission.nature.com/new-submission/13054/3","title":"Critical Care","twitterHandle":"@Crit_Care","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"c858d2a0-9f2f-4165-bd38-b464e940bceb","owner":[],"postedDate":"May 4th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-05-04T09:06:17+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-04 09:06:17","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9396592","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9396592","identity":"rs-9396592","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}