Models /
Enhancing privacy-preserving deployable large language models for perioperative complication detection: a targeted strategy with LoRA fine-tuning
Shaowei Gao, Xu Zhao, Lihui Chen, Xia Feng et al. · npj Digital Medicine · 2025-12-13 · doi · code
What it does (our summary of the abstract)
The authors test large language models at finding and grading 22 kinds of postoperative complications in Chinese surgical records from two hospitals. They split the job into one focused question per complication and then LoRA fine-tune small open Qwen3 models (4B to 32B) on records from the first hospital, so that a model small enough to run on site comes close to, or beats, attending physicians on the second hospital's records.
Model at a glance (from the paper's methods)
| Input | Postoperative clinical notes, nursing notes, abnormal lab results and imaging reports (Chinese text, Markdown) |
|---|---|
| Output | Structured JSON list of complications with severity grade (EPCO definitions) |
| Data | 248 surgical cases from two Chinese hospitals; 100 used for fine-tuning |
| Weights | LoRA fine-tuned Qwen3 models released on Hugging Face |
Validated on SNUH data
clinicalqa@0 preview · ClinicalQA: Korean Medical Knowledge QA Benchmark, 1045 test instances
accuracy 0.722 (0.695–0.749) · #17 of 18 · agent (run by the medbench.kr model evaluator with the authors' released weights) · leaderboard · result
How the agent ran it: ✓ README ✓ weights released 0.66 GPU-h
Official entries will come from the model's authors, after their authorship is verified. This is planned; see the tags.