Family-Connected Remote Monitoring for Stroke Survivors: A Multi-User Machine Learning Platform with Automated Caregiver Alert System

Background: Stroke survivors in low-to-middle-income countries frequently rely on family caregivers who may reside in different cities or countries. No accessible system simultaneously provides machine learning-based recovery monitoring for patients and automated alerts for geographically distributed family members. Methods: A dual-role Streamlit web application was developed in Python 3.10 with two authenticated portals: patient (daily ML-classified check-in with 18 features) and family (read- only monitoring dashboard). Three classifiers were trained on 2,000 synthetic records. An email alert system was implemented using Gmail SMTP (port 587, STARTTLS) triggered by three clinical conditions. SHA-256 password hashing secured all credentials. Data persistence used /tmp/ JSON storage for Streamlit Cloud compatibility. Results: LightGBM achieved accuracy 92.4% and ROC-AUC 0.991 on the held-out test set. The email alert system achieved 100% delivery across 50 triggered test events (20 BP exceedance, 20 adverse classification, 10 check-in completion). Mean alert latency was 8.2 seconds (SD = 1.4s). Role separation was verified across 12 access control tests. Conclusion: A multi-user stroke monitoring platform with automated family alerts is technically feasible and achieves robust ML performance. The platform addresses a specific gap in caregiver-connected rehabilitation technology for dispersed families.