Development of transient ischemic attack risk prediction model suitable for initializing a learning health system...

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Public Health

Background:

• Transient ischemic attack (TIA) significantly increases stroke risk, yet screening rates are low, particularly in developing nations.
• Current TIA detection methods lack inclusivity and practicality for diverse healthcare settings.

Purpose of the Study:

• To develop an inclusive and practical machine learning (ML) model for transient ischemic attack (TIA) risk prediction.
• To establish the first ML-enabled learning health system (LHS) unit for routine and equitable TIA screening.

Main Methods:

• Standardized electronic medical record (EMR) data to build inclusive ML models using a data-centric approach.
• Applied quantitative distribution of TIA risk factors for feature engineering to create a practical, reduced-variable ML model.
• Initiated a TIA ML-LHS unit capable of continuous EMR data updates and validated externally.

Main Results:

• Inclusive ML models (150 variables) achieved 0.868 recall and 0.886 accuracy.
• A practical XGBoost model (20 variables) demonstrated 0.855 recall and 0.796 accuracy.
• The TIA ML-LHS unit achieved 0.830 recall, 0.726 precision, 0.816 ROC-AUC, and 0.812 accuracy, with strong external validation.

Conclusions:

• Developed the first inclusive, practical TIA XGBoost model and initiated the first TIA risk prediction ML-LHS unit.
• The 20-variable model enables equitable TIA screening in hospitals and resource-limited clinics.
• Significant implications for expanding TIA screening, improving early detection, and promoting health equity, with a novel protocol applicable to other diseases.