Abstract
<title>Abstract</title> <p>Introduction Early stroke detection and long-term risk assessment are critical components of preventative health and acute triage. Existing artificial intelligence solutions typically focus on either predictive modeling using tabular data or real-time clinical symptom tracking, but rarely both. We present PredictStroke, a modular framework designed to bridge the gap between preventative analytics and acute triage support by combining ensemble-based long-term risk assessment with real-time, vision- and speech-based acute symptom classification. The predictive engine utilizes soft voting probability averaging across an ensemble of three machine learning models: K-nearest neighbors (KNN), Random Forest, and Support Vector Machine (SVM). The real-time triage arm evaluates independent feature-engineered pipelines: an SVM classifier analyzing facial landmark asymmetry (eye, eyebrow, mouth corner height ratios) and a second SVM classifier analyzing speech fluency metrics derived from Mel-frequency cepstral coefficients (MFCCs), pitch variance, and paused intervals. Results The ensemble risk prediction engine achieved high robust performance with a precision of 0.90, recall of 0.80, and an F1-score of 0.85. Due to deployment constraints, the real-time detection arms were validated independently; the facial drooping detection module achieved an F1-score of 0.65, and the slurred speech classification module achieved an F1-score of 0.62. Conclusions Combining long-term mathematical risk modeling with independent acute symptom detection algorithms is a conceptually valuable and technically feasible framework for remote triage. The modular framework allows data modeling pipelines to scale independently, mitigating severe data availability and system integration bottlenecks common in consumer-facing digital health deployments.</p>