MedAlign: Clinician-Centered Federated Meta-Learning for Medical IoT with Privacy and Interpretability Guarantees

Rong Fu, Muge Qi, Anqi Zhao, weizhi Tang, Simon James Fong

Keywords: Federated Learning, Representation Learning, Medical IoT, Edge Intelligence, Resource-Aware Optimization, Adaptive Gating, Differential Privacy, Multimodal Fusion

TL;DR: MedAlign is a privacy-preserving federated meta-learning framework that improves interpretability and efficiency for clinician-centered AI in medical IoT.

Abstract: We introduce MedAlign, a resource-aware federated meta-learning framework designed for medical Internet-of-Things deployments that face strong data heterogeneity, strict privacy constraints, and tight device resource budgets. MedAlign supports collaborative optimization across distributed clinical sites while enabling per-site personalization. The system couples ontology-driven feature selection with multimodal fusion and prototype-consistent representation learning to preserve stable diagnostic boundaries across non-identical client distributions. A lightweight adaptive gating controller (RL-gating) dynamically modulates module execution according to instantaneous compute, energy, and latency conditions on commodity edge hardware, allowing efficient on-device inference and iterative updates. Privacy is enforced through a formally calibrated aggregation protocol that composes sensitivity-aware noise with a multi-round Rényi-style accountant, yielding quantifiable confidentiality guarantees with minimal impact on clinical utility. We validate MedAlign on intensive-care and wearable-health benchmarks and on commodity edge platforms; the experimental suite includes ablation studies, privacy-accounting traces, and robustness tests against reconstruction and poisoning attacks. Results show that MedAlign consistently improves diagnostic performance and training efficiency while substantially lowering communication and energy costs compared to representative baselines.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 4632