Secondary Prevention of AFAIS: Deploying Traditional Regression, Machine Learning, and Deep Learning Models to...

Area of Science:

• Neurology
• Cardiology
• Data Science

Background:

• Atrial fibrillation (AF) significantly increases ischemic stroke risk, with higher recurrence rates in secondary prevention patients.
• Current guidelines recommend the CHA2DS2-VASc score for oral anticoagulation but its predictive value for recurrence is not well-established.
• Poor reporting quality and challenges with machine learning (ML) highlight the need for transparent, validation-oriented clinical prediction rule (CPR) protocols.

Purpose of the Study:

• To validate and update the CHA2DS2-VASc score for predicting 90-day recurrence risk in AF-related acute ischemic stroke (AFAIS) patients.
• To develop improved clinical prediction rules (CPRs) for secondary prevention in AFAIS.
• To enhance the systematic and transparent reporting of CPR development and validation.

Main Methods:

• Utilized the Virtual International Stroke Trials Archive (VISTA) data from 2763 AFAIS patients and 7809 non-AF AIS patients.
• Validated CHA2DS2-VASc for 90-day recurrence and secondary outcomes using discrimination, calibration, and clinical utility metrics.
• Trained logistic regression, XGBoost, and multilayer perceptron (MLP) models with nested cross-validation, employing transfer learning for the MLP model.

Main Results:

• Models will be assessed on a hold-out test set using AUC, calibration curves, and F1 score.
• SHAP values will interpret models and inform the construction of updated CPRs.
• Performance of new CPRs will be compared against CHA2DS2-VASc and default strategies.

Conclusions:

• The study aims to deliver updated CPRs that better capture 90-day recurrence risk for AFAIS patients.
• Evaluating CPRs based on discrimination, calibration, and clinical utility will guide selection.
• Trade-off analysis will balance the ease-of-use and clinical utility of the developed CPRs.