Jul 2026· International Conference on Digital Health· pp. 23-32· 0 citations· 37 references
Abstract
Large Language Models (LLMs) offer significant potential for clinical decision support, but their deployment is constrained by hallucination, privacy risks, and limited interpretability. Retrieval-Augmented Generation (RAG) improves reliability by grounding outputs in patient-specific electronic health records (EHRs); however, it introduces new challenges related to sensitive data exposure. This paper proposes a governanceaware, privacy-preserving RAG framework for clinical triage decision support using the MIMIC-IV-Ext dataset. The framework integrates retrieval-based grounding, differential privacy, explainability, and an agentic orchestration layer within a unified architecture. An $\epsilon$-ledger is introduced to track cumulative privacy expenditure, and a Privacy-Utility-Explainability (PUE) framework is used to systematically evaluate trade-offs. Experimental results demonstrate that the proposed approach improves classification performance (e.g., achieving up to 82% accuracy and 88% AUROC) and reduces under-triage risk, while privacy mechanisms introduce only modest performance degradation. Retrieval-based attribution further enhances interpretability by linking predictions to clinically relevant cases. These results highlight the effectiveness of combining retrieval, privacy, and explainability for developing trustworthy and deployable healthcare AI systems.
This paper empirically evaluates ChatGPT 3.5 and 4.0 using over 23,000 real user-generated medical queries, assessing their susceptibility to privacy breaches through quasi-identifiers such as age, location, phone number and national registration number and proposes a scalable privacy evaluation model that combines k-anonymity, l-diversity, t-closeness, entropy, re-identification risk and delta-disclosure.
Foad Jalali, Mehran Alidoost Nia· Journal of Supercomputing· 0 citations
Significant challenges remain in AI-driven educational systems in balancing privacy preservation with accurate cognitive diagnosis. To overcome this, we propose a federated inference framework in which several commercial LLM APIs collaborate without requiring access to raw student data or proprietary model internals. Using multiple federated entities, such as LLaMA-3.3-70B, GPT-4o-mini, and Claude-3-Haiku, our framework builds upon a heterogeneous multi-LLM architecture. The predictions generated by these entities are combined with epsilon-local differential privacy by adding Laplace noise locally to each entity's prediction output before aggregation, while residual-based aggregation mitigates model heterogeneity. Our approach is predicated on an honest-but-curious trust paradigm in which API providers are presumed not to abuse submitted queries, and our differential privacy mechanism shields the published diagnostic results from external inference. We conduct rigorous privacy-utility analysis showing strong privacy guarantees with minimal accuracy loss, and extensive real-world evaluations across three educational benchmarks confirm the framework's practical usability and cross-domain generalizability.
Initial evaluations using various machine-learning algorithms on pre-and post-generalized datasets demonstrate the privacy framework’s effectiveness in mitigating privacy risks while preserving data usability.
Ze-Yang Zhu, Matthias N. Louws, Roland V. Bumbuc et al.· 0 citations
LLM-powered healthcare analytics provides clinicians and government agencies with critical insights for effective public health management through intuitive natural language queries, thereby eliminating the need for dedicated IT support. Given the sensitive nature of medical records, our proposed framework is motivated by the requirement to incorporate strong privacy-preserving mechanisms to maintain the confidentiality of patient information. Our model achieves increased level of privacy and utility of healthcare data through five complementary mechanisms: 1) federated framework simulated across multiple healthcare sites to enable secure and scalable analysis of patient records; 2) Complementary NER and Regex based redaction of raw sensitive clinical data; 3) ontology-guided semantic abstraction; 4) access-controlled mechanisms; and 5) a novel two-stage self-regulating differential privacy with heterogeneous privacy budgets. Experimental results indicate that Gemini 3.1 Flash-Lite is the most effective language-model to implement our proposed system, outperforming both Claude Haiku 4.5 and GPT-4o models across a carefully curated set of four application-specific metrics namely numerical accuracy, intent accuracy, hallucination rate and task success rate. During the ablation analysis performed on three distinct datasets including a real-time hospital dataset, the observed progressive increase in consistency score, as components are added, underscores the significance of each mechanism and demonstrates that their combined integration is critical for achieving the maximum consistency score of 96%. Furthermore, scalability experiments conducted over a range of 1 to 10 federated sites indicate that the proposed system achieves a favourable balance between aggregation accuracy and computational efficiency, making it well-suited for large-scale deployments.
K. Bhuvaneswari, M. Varalakshmi· IEEE Access· 0 citations
Chronic kidney disease (CKD) is a major public health concern that requires early and reliable diagnosis to reduce disease progression and associated complications. Existing machine learning approaches often rely on centralized training, limiting their applicability in healthcare environments where data privacy, heterogeneity, and resource constraints are critical concerns. To address these challenges, this study proposes a privacy-preserving federated meta-ensemble stacking framework for CKD prediction. The proposed approach integrates a hybrid preprocessing pipeline, federated learning with prediction-level aggregation, differential privacy, and SHAP-based explainability to enable secure, interpretable, and decentralized model learning. Experimental evaluation demonstrates a mean cross-validation accuracy of 98.85% with an F1-score of 98.58%. The framework maintains robust performance under Gaussian noise, achieving 97.50% accuracy, while cross-client evaluation reaches 99.25%, demonstrating strong generalization across distributed healthcare institutions. The model also exhibits excellent discrimination and calibration, achieving a ROC-AUC of 0.998, PR-AUC of 0.997, and a Brier score of 0.00038. Furthermore, communication overhead is reduced from 0.82 MB to 0.25 MB through compression and sparse updates. These findings demonstrate that the proposed framework provides an accurate, privacy-aware, and interpretable solution for decentralized CKD prediction.
Komal Kumar, D. Sathyanarayanan, Raguraman Purushothaman et al.· Discover Artificial Intellig...· 0 citations
The emergence of large language models offers unprecedented opportunities to transform clinical workflows. However, reliance on proprietary closed-source models poses significant risks to data privacy, institutional autonomy, and reasoning transparency. This study investigates the optimization of open-weight architectures, specifically the Llama-3.1-8B and Qwen2.5-14B-Instruct families, to function as specialized diagnostic tools rather than general-purpose conversational agents. Utilizing the MedQA USMLE-style benchmark, we evaluate the synergistic impact of context window scaling and multi-stage supervised fine-tuning within a localized Retrieval-Augmented Generation framework. Our findings identify a critical scaling threshold at 512 tokens, beyond which diagnostic accuracy stabilizes while preserving the computational efficiency required for on-site clinical deployment. Furthermore, we document a pronounced Brevity Shift, in which multi-stage supervised fine-tuning reduced model verbosity by over 99.7%. The models transitioned from long-form reasoning with an average of 465 tokens to single-token outputs for Qwen and near-single-token outputs for Llama. This shift effectively mitigated verbosity compensation, defined as the tendency of models to mask uncertainty through excessive text, while simultaneously improving diagnostic performance, which peaked at 70.0% accuracy for the Qwen architecture on the benchmark task. Additionally, our analysis reveals that systematic fine-tuning substantially reduces option-preference bias, as measured by Total Variation Distance, leading to more objective and consistent model outputs. Conducted under the INFOSTRATEG Strategic Program (PARROT AI) and funded by the National Centre for Research and Development (NCBR), Poland, this work establishes a robust technical framework for developing trustworthy, high-efficiency medical AI systems capable of operating entirely within hospital-controlled infrastructure.
L. Pawlik, Stanisław Deniziak· Scientific Reports· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
