Post-Stroke Epilepsy Prediction Models in a Cohort of 21,459 Patients

21,459 stroke patient records, including 936 who developed post-stroke epilepsy, were used to develop interpretable machine learning models. The dataset, created by Lijun Wu and last updated in May 2026, describes a dual-tier screening framework for a severely imbalanced clinical cohort. The primary model achieved a macro-AUC of 0.996, while a secondary alert model prioritized sensitivity.

Use Cases

• Developing risk stratification models based on clinical features like neurological severity and lab values
• Training high-sensitivity alerting systems for rare medical events based on the described dual-tier framework
• Benchmarking machine learning techniques for handling severe class imbalance in clinical data
• Interpreting model predictions for post-stroke complications using SHAP analysis on key clinical contributors

Strengths

• Dataset is based on a large cohort of 21,459 stroke patients
• Models achieved high reported performance metrics, including a macro-AUC of 0.996
• Framework is specifically designed for a severe class imbalance of approximately 21.9:1
• Model interpretation was performed using SHAP, identifying key clinical contributors

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count for the underlying data is unknown, which may limit suitability assessment
• The dataset is very small (26.4 KB), suggesting it contains model descriptions or summary statistics rather than the full patient-level data

Provenance

Source

figshare

Collection Method

Retrospective cohort study

Freshness

Last updated 2026-05-21 04:19:24; freshness should be verified