Fine-Tuned MedGemma 27B Extracts Clinical Data from Cancer Discharge Letters

A fine-tuned MedGemma 27B model extracted tumor, lab, and medication information from synthetic cancer discharge letters with F1-scores up to 0.99, but procedure extraction was lower (F1 0.63). On 30 real letters, case-level correctness was 61-93%.

• Proof-of-concept study using 75,000 synthetic discharge letters from 41,175 cancer patients' structured FHIR data to fine-tune MedGemma 27B; tested on 7,500 synthetic and 30 real letters.
• Fine-tuned model outperformed general-purpose LLMs (Qwen3, LLaMA, GPT-OSS) on nearly all extraction categories in a one-shot comparison.
• Limitation: small real-world test set (n=30) from likely a single center; procedure extraction F1 only 0.63 (synthetic) / 61.3% correctness (real); not prospectively validated.

RadPigeon summaries are original and for information only. They are not clinical advice.