Answers are grounded in this paper and cite their sources.
Intelligent edge computing frameworks are crucial for real-time healthcare monitoring, where privacy, adaptability, security, explainability, and regulatory compliance are all critical. This paper introduces a novel system-level solution, called the Self-Learning Edge-IoT Ecosystem, which is an integration of trust-weighted federated learning, confidence-driven self-learning, homomorphic encryption, blockchain-based trust assurance, adaptive wireless communication and explainable artificial intelligence into a single healthcare monitoring platform. The proposed paradigm is different from previous ones that have emphasised individual components, as it introduces an integrated paradigm towards trustworthy and adaptive edge healthcare analytics. The architecture extends canonical Federated Averaging (FedAvg) to a dynamically weighted client-trust aggregation mechanism (FedAvg-TW), with Elastic Weight Consolidation (EWC)-based self-learning to mitigate catastrophic forgetting which occurs during continuous adaptation. Preserving privacy is done by encrypting model modifications with additive homomorphic encryption; being able to track clinical events and model transactions immutably is done with a lightweight Proof-of-Authority blockchain. SHAP and Grad-CAM are added for increased interpretability and transparency for clinicians. Experimental evaluation was performed using 1,000 samples derived from the publicly available MIT-BIH Arrhythmia Database and MIMIC-III Waveform Database via PhysioNet. Feature engineering was applied to raw ECG, HRV, and SpO₂ waveforms to produce a five-feature tabular dataset (ECG Mean Amplitude, HRV, Signal Entropy, SpO₂, Heart Rate) suitable for resource-constrained edge deployment; this tabular representation necessitated the adoption of a lightweight Multi-Layer Perceptron (MLP) as the local edge classifier in place of a raw-waveform CNN-LSTM, and all experimental results pertain to this MLP-on-tabular-features configuration. The 1,000 samples were partitioned at the patient level using a stratified 70/15/15 split, with the 700 training windows distributed across 10 simulated federated clients using a non-IID Dirichlet partition; no patient overlap was permitted between clients or data splits. Results show that the average classification accuracy is 61.3% ± 1.7%, the average AUC is 0.572, and the average inference latency is 202 ms on the NVIDIA Jetson Nano hardware. With respect to the cloud baseline (identical MLP served via REST API), the proposed framework reduces communication overhead by 38% and inference latency by 52%. The addition of the self-learning module leads to a 4.1%-point improvement in classification accuracy, as confirmed by ablation analysis, and statistically significant improvements over recent federated healthcare frameworks of the past ( p < 0.05) as shown by the statistical tests. The proposed ecosystem is architecturally designed to align with key HIPAA, GDPR, and EU AI Act principles (data minimisation, on-device processing, and decision explainability), and is capable of delivering privacy-preserving, explainable and adaptive healthcare intelligence at the network edge; a formal regulatory compliance assessment has not been conducted and would be required before any compliance claim.
No takes yet. Share an insight, caveat, or question.
P. Sundaravadivel (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: