Project 3: Symptom-to-Specialist Router¶
Mentors: To be announced
Problem: Patients describing symptoms in free text don't know which kind of specialist to book, and a generic prompted LLM hasn't been shown to route that decision reliably.
Context: Fine-tunes a distilbert-base-uncased text classifier on Symptom2Disease (public, on Hugging Face), the first project in the track that trains a model instead of only prompting one.
Goals: Given a free-text symptom description, classify which of Symptom2Disease's 22 diagnosis categories it falls under, then map that category to the type of specialist the patient likely needs to see.
Deliverables: A short fine-tuning script (Hugging Face Trainer, a handful of epochs on a small DistilBERT) against the Symptom2Disease training split, a small hand-written lookup table mapping each of its 22 diagnosis categories to a specialist type, the checkpoint served behind a small API, and a free-text input + result card (predicted specialist + confidence) frontend, intentionally simple, so the focus stays on watching a model go from generic to task-specific after training, compared against Projects 1-2's pure-prompting approach.
Showcase: TBD
References:
- Optimizing Classification of Diseases Through Language Model Analysis of Symptoms (Scientific Reports, 2024): fine-tunes BERT-family classifiers on a same-shape symptom-to-disease dataset (24 categories, versus this project's 22), reaching over 99% accuracy with MCN-BERT.
- Evaluating the Diagnostic Performance of Symptom Checkers: Clinical Vignette Study (JMIR AI, 2024): a peer-reviewed, vendor-neutral benchmark of symptom-checker accuracy against physicians, useful context for how well this project's classify-then-route approach could be expected to perform.