Doc Compass

Routes a plain-language health concern to a kind of doctor to book. It suggests a door to knock on; it does not diagnose and is not medical advice.

This repo holds the app code (doccompass/) and its routers, so the demo notebook can run without access to the project's GitHub repo.

What is here

File What it is
checkpoints/distilroberta_stage1/ distilbert/distilroberta-base fine-tuned (all weights) to route concerns to 12 labels
checkpoints/tfidf_stage1.joblib TF-IDF + logistic regression router, trained from scratch on the same data
doccompass/ emergency rules, routers, explanation and the Gradio app

Labels: Dermatology, Orthopedics, ENT, Gastroenterology, Neurology, Urology, Ob-Gyn, Mental health, Cardiology, Eye care, Dentistry, Start with a GP.

Training data

Stage 1 only: 6,252 unique comments from Patient Comments and Specialist Types (Mendeley Data, DOI 10.17632/2twgjzpn82.2, CC BY 4.0), with its 68 symptom categories remapped to the labels above. These are short, one-sentence comments, so treat results on this set as a sanity check only.

Results

On 938 held-out comments from that same public set, the fine-tuned router scores 94.3% top-1 (95% CI 92.7%–95.7%), 99.0% top-3 and 0.908 macro-F1. The comments are short and similar to each other, so this is an easy score: it says how well the router fits the public data, not how it does on real posts.

Limits

  • English only, and trained on short comments. It can be confidently wrong on wording the public set doesn't cover: "my gums bleed when I brush" is routed to Ob-Gyn.
  • The emergency check is a short list of written rules and will miss emergencies phrased in ways the rules don't anticipate. In an emergency call 911.
  • Not clinically validated.
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for ssg1/doc-compass

Finetuned
(786)
this model