Background Infection-induced systemic inflammation in sepsis leads to fatal organ dysfunction, accounting for most intensive care unit (ICU) fatalities. The associated cardiac impairment, termed sepsis-induced cardiomyopathy (SICM), is common yet pathophysiologically complex, posing ongoing clinical and research dilemmas. Although machine learning (ML) models show promise for sepsis-related outcomes, their application for SICM prediction has been rarely reported. Using the Medical Information Mart for Intensive Care-IV (MIMIC-IV) critical care database, we constructed and rigorouslyvalidated an interpretable artificial intelligence algorithm to predict the occurrence of SICM in ICU patients, aiming to provide a clinically translatable prediction tool for timely intervention. Methods Model training and testing were performed using clinical data obtained from the MIMIC-IV database. Predictor variables were refined through both the least absolute shrinkage and selection operator (LASSO) regression and the Boruta algorithm. Ten ML algorithms were trained with fivefold cross-validation on the training cohort and evaluated on the validation cohort based on identification, calibration, and clinical application. The optimal model was underwent additional interpretation using SHAP (Shapley Additive explanations) to quantify feature importance and directionality, and finally deployed as an interactive web-based Shiny app. Results Among 609 patients in MIMIC-IV, 263 (43.2%) developed SICM post-ICU admission. The LightGBM model incorporating 6 variables exhibited superior performance, attaining an attaining a receiver operating characteristic (ROC) curve area of 0.890 of 0.890 (95% confidence interval (CI): 0.858–0.921) in the development set and 0.857 (95% CI 0.799–0.915) in the validation set. Key predictive variables included N-terminal pro-B-type natriuretic peptide (NTproBNP), anion gap, phosphate, systolic blood pressure (SBP), hemoglobin (Hb), red blood cell count (RBC). A clinician-oriented web interface was implemented to facilitate user interaction, accessible at: (<a href="https://shatao.shinyapps.io/SICM_PREDICTION/">https://shatao.shinyapps.io/SICM_PREDICTION/</a>). Conclusions We developed a machine learning model for early detection of SICM risk. The model demonstrated excellent performance across multiple metrics, including discriminative ability, calibration, clinical utility, and robustness. Through SHAP analysis, we elucidated the contribution of each predictive factor. This provides clinicians with an interpretable and user-friendly tool for early SIC risk assessment.

An interpretable machine learning model for predicting sepsis-induced cardiomyopathy in ICU patients: development and validation using the MIMIC-IV database

期刊论文
发表日期：

2025年8月12日

查看全文

Tao Sha,
Hao Jiang,
Lin Bai

摘要

Background

Infection-induced systemic inflammation in sepsis leads to fatal organ dysfunction, accounting for most intensive care unit (ICU) fatalities. The associated cardiac impairment, termed sepsis-induced cardiomyopathy (SICM), is common yet pathophysiologically complex, posing ongoing clinical and research dilemmas. Although machine learning (ML) models show promise for sepsis-related outcomes, their application for SICM prediction has been rarely reported. Using the Medical Information Mart for Intensive Care-IV (MIMIC-IV) critical care database, we constructed and rigorouslyvalidated an interpretable artificial intelligence algorithm to predict the occurrence of SICM in ICU patients, aiming to provide a clinically translatable prediction tool for timely intervention.

Methods

Model training and testing were performed using clinical data obtained from the MIMIC-IV database. Predictor variables were refined through both the least absolute shrinkage and selection operator (LASSO) regression and the Boruta algorithm. Ten ML algorithms were trained with fivefold cross-validation on the training cohort and evaluated on the validation cohort based on identification, calibration, and clinical application. The optimal model was underwent additional interpretation using SHAP (Shapley Additive explanations) to quantify feature importance and directionality, and finally deployed as an interactive web-based Shiny app.

Results

Among 609 patients in MIMIC-IV, 263 (43.2%) developed SICM post-ICU admission. The LightGBM model incorporating 6 variables exhibited superior performance, attaining an attaining a receiver operating characteristic (ROC) curve area of 0.890 of 0.890 (95% confidence interval (CI): 0.858–0.921) in the development set and 0.857 (95% CI 0.799–0.915) in the validation set. Key predictive variables included N-terminal pro-B-type natriuretic peptide (NTproBNP), anion gap, phosphate, systolic blood pressure (SBP), hemoglobin (Hb), red blood cell count (RBC). A clinician-oriented web interface was implemented to facilitate user interaction, accessible at: (https://shatao.shinyapps.io/SICM_PREDICTION/).

Conclusions

We developed a machine learning model for early detection of SICM risk. The model demonstrated excellent performance across multiple metrics, including discriminative ability, calibration, clinical utility, and robustness. Through SHAP analysis, we elucidated the contribution of each predictive factor. This provides clinicians with an interpretable and user-friendly tool for early SIC risk assessment.