Analyzing and predicting the risk of death in stroke patients using machine learning

Area of Science:

• Neurology
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Stroke is a leading cause of death and disability worldwide, necessitating improved prediction and treatment strategies.
• Accurate prediction of stroke-related mortality and understanding treatment heterogeneity are crucial for effective patient management.

Purpose of the Study:

• To develop a machine learning model for predicting stroke patient mortality using comorbidities, lab tests, and demographics.
• To infer heterogeneous treatment effects of warfarin and human albumin in stroke patients for personalized treatment planning.

Main Methods:

• Utilized the Medical Information Mart for Intensive Care (MIMIC)-IV database, including 7,483 stroke patients.
• Developed and interpreted machine learning models, including Shapley Additive Explanation (SHAP), to predict mortality.
• Employed meta-learners to analyze heterogeneity in warfarin and human albumin treatment effects.

Main Results:

• The machine learning model achieved state-of-the-art accuracy in predicting stroke patient mortality.
• Significant differences in demographic factors (age, marital status, insurance, BMI) were observed between survival and death groups.
• Patients aligned with model predictions showed significantly better survival outcomes compared to the overall treatment group.

Conclusions:

• Highly interpretable machine learning models accurately predict stroke prognosis and identify heterogeneous treatment effects.
• Findings align with clinical knowledge and suggest potential for improved stroke patient management and personalized therapies.