JMIR Publications

ABSTRACT

Background:

Patients with septic shock complicated by heart failure exhibit a significantly high mortality rate in the Intensive Care Unit (ICU). Timely intervention and early identification of high-risk patients are essential.

Objective:

In recent years, machine learning has garnered increasing attention for its applications in medical prediction; however, its use in the context of patients with septic shock and heart failure remains relatively underexplored.

Methods:

This study employed the Medical Information Mart for Intensive Care IV (MIMIC-IV)and eICU Collaborative Research Database (eICU-CRD) to identify patients with concurrent diagnoses of septic shock and heart failure. The MIMIC-IV database was used for model training, while the eICU-CRD database served for model validation.Using least absolute shrinkage and selection operator (LASSO) regression, we identified 18 variables significantly associated with prognosis and developed predictive models using eight machine learning algorithms: decision trees, random forests, extreme gradient boosting, light gradient boosting machine, support vector machines, multilayer perceptrons, elastic net, and logistic regression.The optimal model was determined through performance comparison of all models, followed by validation.

Results:

The study included 1,263 patients, comprising 860 from the MIMIC-IV database and 403 from the eICU-CRD database.The ICU mortality rates were 25.5% in the MIMIC-IV database and 19.6% in the eICU-CRD database.The random forest model demonstrated superior performance over other models in all evaluation metrics, achieving an area under curve (AUC) of 0.9349, accuracy of 87.8%, sensitivity of 0.829, F1 score of 0.763, Positive Predictive Value (PPV) of 0.707, and Negative Predictive Value (NPV) of 0.944 in the training set.In the test set, the random forest (RF) model maintained a leading position with an AUC of 0.7289 and an accuracy of 69.5%.Analysis of model interpretability identified acute physiology score III(APSIII), length of ICU stay, lactate levels, sequential organ failure assessment (SOFA) score, and prothrombin time as significant predictors of ICU mortality risk.

Conclusions:

The random forest model demonstrated excellent performance in predicting ICU mortality risk among patients with septic shock and heart failure, exhibiting high accuracy and reliability.This model is a reliable predictive tool that assists clinicians in identifying high-risk patients early, facilitating timely intervention.