Post-Stroke Epilepsy Prediction with Interpretable ML in a 21,459-Patient Cohort

21,459 stroke patient records, including 936 who developed post-stroke epilepsy, form a retrospective cohort for a machine learning study. The research developed a dual-tier screening framework to address severe class imbalance, achieving a model sensitivity of 0.907 and specificity of 0.998. Author Lijun Wu published the study on figshare in May 2026.

Use Cases

• Developing imbalance-aware prediction models for post-stroke epilepsy based on the described dual-tier screening framework.
• Interpreting clinical risk factors for epilepsy using SHAP analysis based on key contributors like neurological severity and D-dimer.
• Comparing model performance metrics (e.g., F1-score, AUPRC) for sensitivity-prioritized versus balanced risk stratification tasks.
• Studying feature selection and reduction techniques in a clinical cohort based on the described staged feature reduction process.

Strengths

• Dataset is derived from a substantial cohort of 21,459 stroke patients.
• The study explicitly addresses and models severe class imbalance (approximately 21.9:1).
• Model performance is reported with specific metrics, including a macro-AUC of 0.996 and an AUPRC of 0.970.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the underlying data tables is unknown, which may limit suitability assessment.
• The primary file is a 22.4 KB DOCX document, suggesting the dataset itself may not be directly included or is very small in scale.

Provenance

Source

figshare

Collection Method

Retrospective cohort study

Freshness

Last updated 2026-05-21 04:19:23; freshness should be verified.