Clinically oriented dual-tier screening for post-stroke epilepsy with interpretable machine learning in a severely...

Area of Science:

• Neurology
• Machine Learning
• Biostatistics

Background:

• Post-stroke epilepsy (PSE) is a significant complication following stroke.
• Early identification of PSE is hindered by its low incidence and severe class imbalance in prediction models.
• Conventional models struggle with the skewed data, limiting their clinical utility.

Purpose of the Study:

• To develop an interpretable machine learning framework for predicting post-stroke epilepsy.
• To address the challenge of severe class imbalance in prediction modeling.
• To create clinically distinct models for both balanced risk stratification and high-sensitivity alerting.

Main Methods:

• Retrospective cohort study of 21,459 stroke patients, with 936 developing PSE.
• Staged feature reduction applied to candidate predictors.
• Development of an imbalance-aware machine learning framework with a dual-tier screening approach.
• Utilized SHAP for model interpretation.

Main Results:

• The cohort exhibited significant class imbalance (approx. 21.9:1).
• The primary model achieved balanced performance (AUC 0.996, Precision-Recall AUC 0.970, F1 0.931).
• The secondary alert model provided higher sensitivity (0.971) for screening.
• Key predictors included neurological severity, hypertension, lactate, D-dimer, and AST.

Conclusions:

• A clinically oriented dual-tier framework effectively addresses class imbalance for PSE prediction.
• The framework provides both balanced risk stratification and high-sensitivity alerting.
• This approach can aid decision-making in follow-up care, pending external validation.