Improving sepsis mortality prediction with machine learning: A comparative study of advanced classifiers and performance metrics.

Abstract

Background: High sepsis mortality rates pose a serious global health problem. Machine learning is a promising technique with the potential to improve mortality prediction for this disease in an accurate and timely manner.

Objectives: This study aimed to develop a model capable of rapidly and accurately predicting sepsis mortality using data that can be quickly obtained in an ambulance, with a focus on practical application during ambulance transport.

Material and methods: Data from the Medical Information Mart for Intensive Care-IV (MIMIC-IV) dataset were used to compare the performance of 11 machine learning algorithms against the widely utilized quick Sequential Organ Failure Assessment (qSOFA) score. A dynamic updating model was implemented. Performance was evaluated using area under the curve (AUC) and precision-recall area under the curve (PRAUC) scores, and feature importance was assessed with SHapley Additive exPlanations (SHAP) values.

Results: The light gradient boosting machine (LightGBM) model achieved the highest AUC (0.79) and PRAUC (0.44) scores, outperforming the qSOFA score (AUC = 0.76, PRAUC = 0.40). The LightGBM also achieved the highest PRAUC (0.44), followed by Optuna_LightGBM (0.43) and random forest (0.42). The dynamically updated and tuned model further improved performance metrics (AUC = 0.79, PRAUC = 0.44) compared to the base model (AUC = 0.76, PRAUC = 0.39). Feature importance analysis offers clinicians insights for prioritizing patient assessments and interventions.

Conclusions: The LightGBM-based model demonstrated superior performance in predicting sepsis-related mortality in an ambulance setting. This study underscores the practical applicability of machine learning models, addressing the limitations of previous research, and highlights the importance of real-time updates and hyperparameter tuning in optimizing model performance.