A Web-Based Interface and Machine Learning for Intelligent Stroke Risk Assessment

Abstract— Early identification of people with common, easily available health markers that indicate a pattern of higher risk is helpful to prevent stroke. We have developed a ML screening method for brain-stroke vulnerability using structured demographical, clinical and lifestyle data which is available via the web. The proposed method is based on the EBM along with the Mondrian CP and SA. LR, DT, KNN and Gaussian NB are used as baseline classifiers. The EBM is trained to learn additive feature functions that are interpretable and the conformal layer generates prediction sets output for each class with a desired coverage target. If the circumstances are ambiguous, no attempt to label them as risky is made, but instead the interface provides a Review Required state. Experimental results of the EBM point classifier produce results of 94.7% accuracy, 92.8% precision, 95.9% recall, 94.3% F1-score and 97.1% ROC-AUC. The proposed uncertainty layer has an empirical coverage of 90.8% at nominal conformal coverage level of 90% and an abstention rate of 8.6%. The final architecture represents an interpretable and uncertainty-aware web screening approach that relies on intrinsic additive modelling and uncertainty in the prediction set. “Index Terms—stroke prediction, machine learning, Explainable Boosting Machine, conformal prediction, selective prediction, web application, uncertainty-aware screening.”