Predicting Mortality in Intensive Care Unit Patients With Heart Failure Using an Interpretable Machine Learning...

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Cardiology

Background:

• Heart failure (HF) poses a significant public health challenge, necessitating accurate mortality prediction for personalized treatment.
• Current HF mortality prediction models often lack the interpretability required for clinical implementation.

Purpose of the Study:

• To develop an interpretable machine learning model for predicting mortality risk in intensive care unit (ICU) patients with heart failure.
• To utilize the SHapley Additive exPlanation (SHAP) method for explaining an extreme gradient boosting (XGBoost) model and identifying key prognostic factors in HF.

Main Methods:

• A retrospective cohort study using the eICU Collaborative Research Database (eICU-CRD).
• Data from the first 24 hours of ICU admission were used for training (70%) and validation (30%) of machine learning models.
• The XGBoost model's performance was compared against three other models using the area under the curve (AUC), and SHAP was employed for model interpretability.

Main Results:

• The XGBoost model demonstrated superior predictive performance (AUC=0.824) compared to other models, including support vector machine (AUC=0.701).
• The SHAP analysis identified the top 20 predictors of HF mortality, with average blood urea nitrogen levels being the most significant.
• The XGBoost model showed superior net benefit over other models within specific probability thresholds.

Conclusions:

• An interpretable predictive model can enhance physicians' ability to accurately assess mortality risk in ICU patients with HF.
• This interpretability facilitates better treatment planning, optimal resource allocation, and increased trust in predictive model reliability.