Development and Validation of an Interpretable Machine Learning-Based Clinical Prediction Model for Short-Term...

Area of Science:

• Critical Care Medicine
• Neurology
• Data Science

Background:

• Intracerebral hemorrhage (ICH) with thrombocytopenia is linked to poor patient outcomes.
• Limited early risk prediction tools exist for this specific patient subgroup.
• Developing accurate prediction models is crucial for timely intervention.

Purpose of the Study:

• To develop and externally validate an interpretable machine learning model.
• The model aims to predict 28-day all-cause mortality in ICU patients with ICH and thrombocytopenia.
• To provide an accessible tool for early risk stratification.

Main Methods:

• Utilized large internal datasets (MIMIC-III/IV, eICU, NWICU) and an external validation cohort from China.
• Trained and evaluated five machine learning models, selecting LightGBM for its performance.
• Employed SHAP analysis for model interpretability and developed a web-based tool.

Main Results:

• The LightGBM model demonstrated strong performance with AUROCs of 0.840 (internal) and 0.764 (external).
• Key predictors identified include Glasgow Coma Scale (GCS), diastolic blood pressure, glucose, and platelet count.
• The model achieved good internal and acceptable external discrimination.

Conclusions:

• The developed LightGBM model shows potential as an early risk stratification tool for ICH patients with thrombocytopenia.
• The model offers interpretable feature contributions, enhancing clinical utility.
• Further prospective multicenter validation is recommended before widespread clinical implementation.