An interpretable machine learning model for predicting sepsis-induced cardiomyopathy in ICU patients: development and...

Area of Science:

• Critical Care Medicine
• Cardiology
• Artificial Intelligence in Medicine

Background:

• Sepsis causes fatal organ dysfunction, with sepsis-induced cardiomyopathy (SICM) being a complex and common impairment in intensive care units (ICUs).
• Machine learning (ML) models show potential for predicting sepsis outcomes, but their application to SICM prediction is limited.
• Developing accurate prediction tools for SICM is crucial for timely clinical intervention and improved patient outcomes.

Purpose of the Study:

• To construct and validate an interpretable artificial intelligence algorithm for predicting SICM in ICU patients.
• To provide a clinically translatable prediction tool to facilitate early detection and intervention for SICM.
• To leverage the MIMIC-IV database for robust model development and validation.

Main Methods:

• Utilized the Medical Information Mart for Intensive Care-IV (MIMIC-IV) database for model training and testing.
• Employed LASSO regression and the Boruta algorithm to refine predictor variables.
• Trained and evaluated ten ML algorithms using fivefold cross-validation, with the optimal model interpreted using SHAP (Shapley Additive explanations) and deployed as a web app.

Main Results:

• Identified 263 out of 609 (43.2%) patients developing SICM post-ICU admission.
• The LightGBM model, using 6 variables, achieved an ROC AUC of 0.890 in the development set and 0.857 in the validation set.
• Key predictors included NTproBNP, anion gap, phosphate, SBP, Hb, and RBC; an interactive web tool was created.

Conclusions:

• Developed a high-performing ML model for early SICM risk detection.
• The model demonstrated excellent discriminative ability, calibration, clinical utility, and robustness.
• Provided clinicians with an interpretable, user-friendly tool for early SIC risk assessment via SHAP analysis and a web interface.