Patient Care Portal Triage

Patient portals have a quiet failure mode: a question lands in a generic inbox, waits, and gets forwarded twice before it reaches someone who can answer it. The patient learns nothing in the meantime.

This project attacks both halves of that. A classifier triages incoming queries to the correct team — insurance, scheduling, medical advice, and so on — and an information retrieval step surfaces relevant evidence-based research so the patient has something credible to read while they wait for a professional.

How it works

I built the classifier in PyTorch on top of sentence embeddings from Google’s EmbeddingGemma 300M model. Working on embeddings rather than raw text means the classifier stays small, trains fast, and generalises across the enormous variety of ways people describe the same problem.

What I’d want before it touched a real patient

The honest answer is: a lot. Confidence thresholds with a human fallback, explicit handling for anything that reads as an emergency, and retrieval that can only cite a vetted corpus. Healthcare is exactly the domain where the interesting question is not “how accurate is it” but “what does it do when it is wrong.”