MEDAL: Sequential adapter learning for privacy-preserving multicenter clinical language models.

BACKGROUND: Scalable reasoning over clinical notes is a major goal in medical artificial intelligence. Large Language Models (LLMs) are promising for this task, particularly when fine-tuned on multicenter note corpora, but progress has been limited by barriers to data sharing and technical gaps in Federated Learning (FL). METHODS: We developed Multi-institutional Efficient/Distributed Adapter Learning (MEDAL), a framework that sequentially fine-tunes lightweight LLM adapters across health systems without transferring patient-level data. We applied MEDAL to fine-tune LLaMA-3.1-8B for two discharge-summary tasks: identifying sepsis and in-hospital mortality. We trained this LLM using three siloed datasets: MIMIC-IV, a University of California, San Francisco (UCSF) adult ICU cohort, and a UCSF pediatric ICU cohort, with one GPU per site across the University of Texas at Austin and UCSF. We compared the multicenter MEDAL model with a centralized pooled-data model, single-center models, and GPT-5. To further evaluate MEDAL's stability at scale, we partitioned MIMIC-IV into 10 virtual centers and sequentially fine-tuned adapters across them. RESULTS: Using MEDAL to fine-tune LLaMA-3.1-8B yielded high F1 scores approaching the upper bound of centralized data pooling, outperforming frontier models such as GPT-5 by up to 16.6 percentage points, and consistently outperforming center-specific locally fine-tuned models across the target tasks. In the 10-center experiment, performance converged quickly and remained stable across training rounds. CONCLUSION: MEDAL is a scalable method for fine-tuning LLMs across many health systems without sharing patient-level data, enabling high-performance local models for reasoning over clinical notes. MEDAL may also be useful for training multimodal healthcare AI models.