Hybrid Framework for Sustainable Wearable Stroke Monitoring

The objective is to create a real-time, leakage-resistant stroke-risk pipeline that incorporates privacy-preserving learning methodologies and is deployable on smart watches. This paper employed 38 clinically curated variables to implement plausibility filtering, stratified splitting, correlation trimming, imputation, winsorization, robust scaling, rare-level grouping, and one-hot encoding on a harmonized dataset of 10,000 individuals. The RFECV wrapper was employed for feature selection, while Cat Swarm Optimization was utilized to optimize these features while searching for a regularization hyper-parameter and a binary feature mask. The creation of two-layered hybrid ensembles involved SCT optimization, which included threshold tuning, isotonic probability calibration on the validation set, and train-only SMOTE. Hybrid Model 1 attained the highest accuracy of 96.4% on the reserved test set at the optimal operating threshold (threshold = 0.50). The performance consistently demonstrates accuracy in threshold-sweep studies around the ideal point, surpassing both its default settings and the superior configurations of Hybrid Model 2. This research encompasses calibrated thresholding, a dual-stage selection procedure employing both wrapper and swarm methodologies, edge-ready artefact packaging comprising the pre-processor, features, model, and threshold, as well as comprehensive leakage control throughout the entire process. It provides enhancements for privacy preservation through automated reporting and notifications, along with a federated learning update mechanism designed for nonIID contexts. It is recommended to conduct multi-centre external validation, incorporate wearable and longitudinal data streams, uphold calibration and thresholds to address drift, implement safety overrides and bias audits, and pursue future trials utilizing secure, versioned federated updates and energy-efficient, on-device inference for sustainable healthcare.