AI-Powered Smart Healthcare Monitoring System Using Federated Learning

This is the authors' abstract. We don't add key points for this paper.

Healthcare artificial intelligence is still commonly built by pooling patient data from every hospital on a single server and training a model there. This approach delivers accuracy but creates a serious privacy risk, and it conflicts with regu-lations such as HIPAA, GDPR and India’s Digital Personal Data Protection Act, which is a major reason why hospitals are reluctant to release raw records [1], [2]. This paper pro-poses a system that avoids centralization instead of work-ing within it. Wearable devices at each hospital node con-tinuously collect physiological signals (heart rate, ECG, SpO2 and blood pressure), and a Long Short-Term Memory (LSTM) model [3] is trained on these data locally and on-site. Only an encrypted set of model weights leaves the hospital; a central server combines the weights using Fed-erated Averaging (FedAvg) [4] and returns an improved global model. Over 20 communication rounds, the global model reaches 97.6% accuracy and a 97.2% F1-score for early disease prediction over five categories, which is with-in about half a percentage point of a centralized model trained with full access to the same pooled data. Clinicians interact with the system through a Streamlit dashboard that shows prediction confidence, patient trends and alerts without exposing another hospital’s raw records.