FedKGC: Federated Knowledge-Grounded Calibration Framework for Hallucination Mitigation in Medical LLMs

The integration of Large Language Models (LLMs) into clinical workflows offers transformative potential for decision support and documentation, yet widespread adoption is currently stalled by the prohibitive risks of hallucination and strict privacy regulations preventing data centralization. To bridge this gap, we propose FedKGC (Federated Knowledge-Grounded Calibration), a novel framework that synergizes privacy-preserving Federated Learning (FL) with retrieval-augmented calibration. FedKGC introduces a dual-phase mitigation strategy: employing a Knowledge-Grounded Loss ($\mathcal {L}_{KG}$) at the local level to anchor Low-Rank Adaptation (LoRA) fine-tuning to private clinical guidelines, and utilizing differentially private metadata during global aggregation to construct a Global Conflict Map that down-weights contradictory concepts. Evaluation on a non-IID partition of the MIMIC-IV dataset and the MedHallu benchmark demonstrates that FedKGC achieves a 28% reduction in hallucination rates compared to standard federated baselines while strictly adhering to $(\epsilon, \delta)$-Differential Privacy guarantees. These results demonstrate that high-fidelity, hallucination-resistant medical AI can be trained collaboratively across siloed healthcare institutions without compromising patient privacy, paving the way for trustworthy distributed medical intelligence.