README.md

metadata

license: apache-2.0
base_model:
  - google/medgemma-1.5-4b-it
  - AlphaRaven/medgemma-4b-antihallu
tags:
  - medical
  - dermatology
  - adverse-event-detection
  - medgemma

MedGemma AE Detection

MedGemma 1.5-4B fine-tuned for visual adverse event detection from clinical photographs using paired image comparison.

• Base model: AlphaRaven/medgemma-4b-antihallu
• Method: LoRA fine-tuning (r=8, alpha=16, target: q_proj/v_proj)
• Training: 100 epochs
• Task:
  • 1. Paired comparison of baseline vs. current patient photos.
    2. Classify skin AEs from patient photos into 21 categories (normal + 7 AE types × 3 CTCAE grades; alopecia G1-G2 only)
• AE types: maculopapular rash, acneiform rash, periorbital edema, SJS, stomatitis, pruritus, alopecia

Performance (test set, 42 paired images):

- Accuracy: 95.2%
- Weighted F1: 0.951
- 

Dataset**: Synthetic clinical photographs (Gemini-generated)

- 147 train / 21 val / 42 test paired images

Usage

from transformers import AutoModelForImageTextToText, AutoProcessor

model = AutoModelForImageTextToText.from_pretrained("AlphaRaven/medgemma-ae-detection")
processor = AutoProcessor.from_pretrained("google/medgemma-1.5-4b-it")

Input: Two images (baseline photo + current photo) with a clinical reasoning prompt using CTCAE grading criteria. Output: JSON array of detected AEs with ae_term, grade, and reasoning.

Part of the Clinical Trial Simulation Engine pipeline.