PhysioGPT.
AI assistance for neurological physiotherapy. Practice pilot.
Therapeut
Patient nach Schlaganfall, 6 Wochen post-akut. Gangschulung?
PhysioGPT
Bobath → PNF-Diagonalmuster → Laufband mit Entlastung.
Guidelines are correct — and hard to find.
Christine works in a neurology practice. Her patients come in after strokes, with multiple sclerosis or Parkinson’s. Therapy guidelines are documented — across multiple PDFs, in textbooks, in conference protocols. In day-to-day practice she has 15 minutes between two patients to clarify a question. PDF search isn’t enough.
General-purpose LLMs aren’t a solution either. They hallucinate on medical questions, don’t cite sources, conflate Bobath with PNF with Vojta. Without a controlled knowledge base, the risk is too high.
RAG on a curated base.
PhysioGPT takes a specific patient question (anonymised), searches the curated knowledge base (guidelines, Bobath texts, PNF material), and returns an answer with page-number citations. If the base has no answer, it says so — no hallucinations.
Currently piloted in Christine’s practice. Next phase: RLS for multi-user (multiple practices, separate logs), persistent chat history, mobile app. Privacy is non-negotiable — patient data never leaves the system.
/ tech-decisions
Three decisions that mattered.
RAG
Curated knowledge base, not open search
Bobath concept, PNF, stroke guidelines — vectorised in pgvector. Answers come from this base, not from generic model memory. Source citations include page numbers.
Edge Function
AI inference on Supabase Edge
No dedicated backend server. Edge function (Deno) calls Claude, writes logs to Supabase, replies via streaming. Scales automatically, no idle cost.
Privacy
No patient data sent to the model
Use case is anonymised case descriptions. Therapist enters symptoms and context, not names. Logs are traceable for 30 days, then auto-deleted.
/ other apps