MACHINE LEARNING-ENHANCED SECURE PLATFORM FOR ELECTRONIC HEALTH RECORD SHARING WITH PROACTIVE THREAT DETECTION

Abstract

The need to achieve high availability, quick clinical access, and robust protection of electronically protected health information (ePHI) is critical to EHR exchange. However, these defenses, authentication, role-based authorization, encryption, and audit logging, are often reactive, as they fail to protect when a user with legitimate access to the network attempts to misuse it, or when a user enters a disguised document into a network that prevents even the best static signatures. This paper introduces Secure Health ML, a web-based prototype of an EHR sharing system that features fine-grained access control, authenticated encryption, secure transport, audit logging, and machine learning-based threat detection. The platform's application layer is based on Next.js/React; it is persisted in PostgreSQL; it includes an additional component for malware scanning; and it has an isolated Python microservice for supervised document-risk inference. Structural and metadata-derived features are used to evaluate the Random Forest (RF) and XGBoost (XGB) classifiers on a held-out dataset of PDF malware. Random Forest was chosen for deployment due to its highest test accuracy (99.25%), precision (99.02%), recall (99.64%), F1 score (99.33%), and ROC-AUC (0.9997). Security validation showed it successfully implemented policies such as authentication, RBAC/ABAC, encrypted storage, alert generation, and quarantine-driven response to high-risk uploads. The results suggest that an integrated solution combining policy-based access-control governance and model-based detection for insider and credential compromise can minimize the window of exposure caused by these activities. The prototype is constrained by the limited scope of the test environment and the simulated data set. Still, it offers a technically viable basis for designing the adaptive, security-aware sharing system for EHRs.