AI-assisted clinical evidence scoping tool for physiotherapy.
Turn a patient profile into a curated, quality-rated evidence brief through a 6-agent, PRISMA-informed workflow across 7+ free scientific databases. Designed to reduce the time required for initial evidence scoping.
⚠️ PhysioFlow is a decision-support tool, not a substitute for a formal systematic review. All outputs are labeled and time-stamped.
Evidence-based physiotherapy is stuck in a bottleneck: a clinician has ~5 minutes between patients, but a proper literature search takes hours. Existing tools (PubMed, Google Scholar) return raw lists with no screening, no quality rating, and no synthesis. PhysioFlow closes that gap by assisting with structured evidence scoping through a traceable, PRISMA-informed workflow.
Input a patient profile (age, condition, comorbidities, goals). The pipeline:
- Translates the profile into a structured PICOS query
- Searches 7+ databases in parallel (PubMed, OpenAlex, Europe PMC, SciELO, ClinicalTrials.gov, medRxiv, Semantic Scholar)
- Screens results through a 3-level, PRISMA-informed screening workflow with two model-assisted review passes
- Rates quality using Cochrane RoB 2, PEDro, and Oxford CEBM levels
- Extracts quantitative outcomes from included studies
- Meta-analyzes using random-effects DerSimonian-Laird with Hedges' g, then outputs SVG forest plots
- Rates confidence with GRADE and delivers a plain-language brief with citations
Every step is logged and traceable back to the source paper.
The screening, RoB 2, PEDro, Oxford CEBM, extraction, meta-analysis, and GRADE outputs are model-assisted preliminary results that require expert verification. PhysioFlow supports evidence scoping; it does not replace a formal, human-conducted systematic review.
Six independent agents coordinated by a pipeline orchestrator:
| Agent | Responsibility |
|---|---|
picos_translator |
Patient profile to structured PICOS |
search |
Parallel query across 7+ databases |
screener |
3-level, PRISMA-informed model-assisted screening (title/abstract/full-text) |
quality |
RoB 2 / PEDro / Oxford CEBM rating |
extractor |
Quantitative outcome extraction |
synthesizer |
Random-effects meta-analysis + GRADE + brief |
Each agent returns strict JSON and is independently testable.
- Framework: Next.js 16 (App Router, Turbopack)
- Language: TypeScript
- UI: Tailwind CSS + shadcn/ui
- LLM: DeepSeek API (OpenAI-compatible) for agent reasoning
- Statistics: Custom random-effects meta-analysis implementation (DerSimonian-Laird, Hedges' g)
- Visualization: SVG forest plots (no charting library dependency)
- Deployment: Vercel
# 1. Clone
git clone https://github.com/victorgomezadapty/physioflow.git
cd physioflow
# 2. Install
npm install
# 3. Configure
cp .env.example .env.local
# Fill in DEEPSEEK_API_KEY and optional PUBMED_API_KEY / SEMANTIC_SCHOLAR_API_KEY
# 4. Run
npm run dev
# Open http://localhost:3000The app works without a PubMed API key. The key only raises rate limits.
MVP functional. The end-to-end pipeline runs with real searches and real papers, under active development toward:
- Public demo deployment
- Editable inclusion/exclusion criteria per query
- Session persistence and shareable briefs
- Additional condition-specific PICOS templates
- User-facing quality rating explanations
See EXECUTION_PLAN.md for the detailed build plan.
- Transparency over convenience. Every claim in the output is linked to its source.
- Honest labeling. This is a scoping tool, not a systematic review, and it says so.
- Bilingual by default. Spanish patient-facing labels, English technical terms.
- Zero patient data storage. Queries are stateless and nothing is retained server-side.
Built by Víctor Andrés Gómez López. Based in Riyadh, Víctor works as a Personal Trainer and supports the implementation, integration, and consistent delivery of physiotherapy services across three fitness clubs. He is also a doctoral researcher and founder of ADAPTY, building human-supervised health-data and decision-support prototypes.
Contact: LinkedIn
Source available for review and evaluation. Contact the author for commercial or clinical deployment inquiries.