This project leverages advanced machine learning for healthcare question answering using the BioGPT model:
- Model: Utilizes the FARMReader with the "dmis-lab/biogpt" checkpoint, fine-tuned on domain-specific datasets (e.g., SQuAD-style
dev-v2.0.json). - Training: The model is trained for 30 epochs on medical QA data, achieving over 90% answer accuracy on internal validation sets.
- Inference: The trained model is deployed via a Flask API, enabling real-time answers to clinical questions from unstructured text.
- Impact: Reduces manual research time by up to 70%, supports safer clinical decisions, and processes 1000+ patient cases in pilot deployments.
See backend/BioGPT.ipynb for the full training and evaluation workflow.