Early Stroke Risk Detection via AI and Self-Reported Symptoms

A new study presents a passive surveillance system based on artificial intelligence, designed for the early detection of stroke risk in high-risk individuals, such as diabetic patients. The system analyzes symptoms reported by the patients themselves, filling a critical gap in the timeliness of care.

System Architecture

The system is based on a symptom taxonomy derived from patient language and a machine learning pipeline that combines a heterogeneous graph and EN/LASSO models. This approach allows the identification of symptom patterns associated with subsequent stroke. The results were translated into a hybrid screening system that integrates symptom relevance and temporal proximity.

Performance and Evaluation

The system was evaluated through simulations based on electronic health records (EHR) over time windows of 3-90 days. By adopting conservative thresholds to minimize false alerts, the system achieved high specificity (1.00) and a prevalence-adjusted positive predictive value of 1.00, with good sensitivity (0.72), an intentional trade-off to prioritize precision. Sensitivity was highest in the 90-day window.

Implications

The ability to detect stroke risk with high precision and low burden, based solely on patient language, offers a valuable time window for clinical evaluation and intervention in high-risk individuals.