Toward Trustworthy Healthcare AI: A Federated Explainable Deep Learning Framework for Secure and Privacy-Preserving Clinical Decision Support
Applied IT & Engineering · 01.10.2023 · 1 (1) · 1-11
DOI: 10.25163/engineering.1110824 ↗Kurz gefasst
Fachartikel in der Zeitschrift Applied IT & Engineering, online erschienen am 01.10.2023.
Abstract (Originalsprache)
Background
Artificial intelligence (AI) holds extraordinary promise for transforming clinical decision-making in modern healthcare. Yet three persistent barriers — compromised patient data privacy, inadequate model transparency, and insufficient security infrastructure — continue to constrain its adoption at scale. These are not small technical inconveniences; they represent fundamental gaps between the potential of AI and its safe, ethical deployment in clinical environments.
Methods
We propose the Federated Explainable AI (FEXAI) Framework — a unified architecture that integrates TensorFlow Federated (TFF)-based distributed model training with SHapley Additive exPlanations (SHAP) for post-hoc interpretability, and a blockchain-anchored secure aggregation protocol for model update integrity. Using a composite distributed dataset from three simulated healthcare institutions (n = 42,830 de-identified patient records drawn from MIMIC-III and UCI repository benchmarks), local deep neural network models were trained independently at each node. Federated Averaging (FedAvg) was applied across 50 communication rounds to converge a global model, with no raw patient data ever leaving the local institution.
Results
The FEXAI Framework achieved a diagnostic accuracy of 96.8% (F1-score: 95.5%), surpassing centralized deep neural network (92.7%) and traditional machine learning baselines including Random Forest (89.4%) and Support Vector Machine (86.9%). Privacy preservation and security metrics — assessed against defined scoring rubrics — reached 98.3% and 97.6%, respectively. SHAP-based explainability produced clinician-interpretable feature attribution scores with an explainability rating of 94.8%. Mean inference latency was 0.42 seconds per case.
Conclusion
The FEXAI Framework demonstrates that accuracy, privacy, security, and transparency are not mutually exclusive objectives in healthcare AI. The findings suggest a viable and reproducible path toward clinical AI systems that clinicians can trust, regulators can audit, and patients can rely on.
Zitieren
APA
(2023). Toward Trustworthy Healthcare AI: A Federated Explainable Deep Learning Framework for Secure and Privacy-Preserving Clinical Decision Support. Applied IT & Engineering, 1(1), 1-11. https://doi.org/10.25163/engineering.1110824
Vancouver
Toward Trustworthy Healthcare AI: A Federated Explainable Deep Learning Framework for Secure and Privacy-Preserving Clinical Decision Support. Applied IT & Engineering. 2023;1(1):1-11. doi:10.25163/engineering.1110824
BibTeX
@article{euvimed2023,
title = {Toward Trustworthy Healthcare AI: A Federated Explainable Deep Learning Framework for Secure and Privacy-Preserving Clinical Decision Support},
journal = {Applied IT & Engineering},
year = {2023},
volume = {1},
number = {1},
pages = {1--11},
doi = {10.25163/engineering.1110824}
}
EUVIMED