A Unified Quantum-Resilient Security Framework for Intelligent Healthcare Ecosystems Integrating Zero-Trust Architecture with Explainable Artificial Intelligence

Abstract

The Internet of Medical Things (IoMT) adds the risk to patients of exposing life-critical telemetry to a simultaneous and prospective adversary. In particular, this paper demonstrates a variety of issues pervasive throughout the existing IoMT intrusion-detection literature, including optimising an isolated single classifier for accuracy where transport-layer cryptography will never survive the post-quantum transition, reporting opaque decisions and unfairly inflating headline scores by leaving identifier columns in the feature set. We propose a comprehensive and four-layered security framework that integrates

Its components include: (i) dual branch attention-fusion detector over the joint network-flow and biometric feature space, generating an exploration hybrid quantum-classical variational branch; (ii) explainable-AI layer which audits its attributions for faithfulness and stability via SHAP/LIME; (iii) Zero-Trust policy engine which converts both detector confidence & explanation signals into per-session trust scores + micro-segmentation decisions and; (iv) overhead analysis of post-quantum cryptography [PQC] - a must-have solution in our current constrained IoMT edge. The proposed detector achieved F1 = 0.681 ± 0.030, MCC = 0.637 ± 0.032 and PR-AUC = 0.832 ± 0.019 on the WUSTL-EHMS-2020 testbed dataset using a leakage-audited 5-fold ×3-repeat stratified evaluation protocol. The most accurate detectors are strongly-tuned gradient-boosted trees (LightGBM F1=0.788), and we transparently report the results of varying tuning settings;

However the model proposed is: (i) the strongest adversarially robust differentiable detector (0.935 stable accuracy under FGSM/PGD to ε = 0.2, collapse ≤0.14 by a 1-D CNN), (ii) edge-deployable (13.2k params, 0.24 ms CPU inference, Raspberry-Pi feasible) and iii) yields trustworthy explanations that are stable and interpretable overall(top-10 feature Jaccard of 0.93 across folds). SHAP assigns 57.1% of decision mass to network features and 42.9% to biometrics, estimating the benefit of multi-modal fusion. Zero-Trust micro-segmentation improves containment against lateral-movement attacks from 0.562 to 0.773 (+21.1 pp) in simulation, and ML-KEM/ML-DSA is shown to be practical at the IoMT edge with sub-millisecond handshake latency. The framework is released as an end-to-end, initialised-seed reproducible pipelined.