TracHeal — AI Patient Care Continuity Agent

Catching patients before they fall through the cracks between hospital discharge and follow-up care.

Team Cache Hit Dubai · HSIL Hackathon 2026 · Challenge #7: Fragmented Care Pathways

---

The Problem

Every year, millions of patients are discharged from hospital with complex follow-up plans — referrals, appointments, medication changes — that never get executed. Thirty days later, they're back in the ER.

• 22.8% of patients miss their first follow-up appointment (Ambade et al., BMJ Open 2025)
• $21 billion annual cost of preventable readmissions in the US (Vizient 2025 / CMS HRRP)
• 27% of hospital readmissions are entirely preventable (van Walraven et al., CMAJ 2011)

Discharge isn't the finish line — it's where fragmented health systems lose track of patients.

What TracHeal Does

TracHeal is an AI-powered care continuity agent that ingests unstructured discharge notes, identifies patients at risk of falling through the cracks, and surfaces actionable interventions for care coordinators — with full source attribution and clinician-in-the-loop review.

Three core pillars:

1. EHR Parsing — Ingests unstructured clinical text (via paste, PDF, or FHIR APIs in future)
2. AI Risk Profiling — Identifies hidden barriers: medication complexity, missed referrals, jargon, social determinants
3. Proactive Care Dashboard — Unified timeline with prioritised interventions and real-time alerts

Live Demo

trachealhackathon.vercel.app/demo

Paste a discharge summary and see:

• Risk score + risk level (high / medium / low)
• Identified barriers by category (clinical, medication, social, literacy, transport)
• Prioritised interventions with rationale
• RAG source attribution with relevance scores
• Patient-facing insight tags (jargon, complexity, cost, psychological barriers)

Input + RAG Retrieval	Risk Assessment Output

Barriers + Interventions	Source Attribution

Architecture

Clinical Notes (text or PDF)
        │
        ▼
┌───────────────────┐
│  Entity Extraction │   ← NLP parsing
└─────────┬─────────┘
          ▼
┌───────────────────┐
│   RAG Retrieval    │   ← TF-IDF cosine similarity over patient records
└─────────┬─────────┘
          ▼
┌───────────────────┐
│   LLM Inference    │   ← Gemini → DeepSeek → mock fallback cascade
└─────────┬─────────┘
          ▼
┌───────────────────┐
│  Structured Output │   ← risk_score, barriers, interventions, sources
└───────────────────┘
          │
          ▼
   Clinician Review  ← Human-in-the-loop, non-negotiable

Stack:

• Frontend: Static HTML/CSS/JS deployed on Vercel CDN
• Backend: Python Flask as Vercel serverless function
• LLM: Google Gemini (primary) with DeepSeek fallback and deterministic mock fallback
• RAG: Pure-Python TF-IDF cosine similarity (no NumPy — keeps Vercel bundle under 50MB)
• PDF Ingestion: PyPDF2
• Deployment: Vercel (static pages + Python serverless API)

Project Structure

.
├── index.html              # Landing page
├── demo.html               # Interactive demo with live API
├── pilot.html              # DHA pilot deployment plan
├── business.html           # Business model and ROI
├── stakeholders.html       # Stakeholder map
├── styles.css              # Shared styling
├── api/
│   └── index.py            # Unified Flask backend (RAG + LLM + PDF)
├── requirements.txt        # Python dependencies
├── vercel.json             # Vercel routing + Python function config
├── .vercelignore           # Excludes reference dirs from deployment
└── citations/              # Source PDFs for problem statistics

Running Locally

# Clone the repo
git clone https://github.com/riyashet-hds/trachealhackathon.git
cd trachealhackathon

# Install Python dependencies
pip install -r requirements.txt

# Set API keys (optional — app falls back to mock mode without them)
export GEMINI_API_KEY="your_key_here"
export DEEPSEEK_API_KEY="your_key_here"

# Run the backend
python api/index.py

# Serve static pages (any static server works)
python -m http.server 8000

Then open http://localhost:8000/demo.html.

API

POST /api/generate

Accepts either:

• application/json with {"note": "discharge text..."}
• multipart/form-data with a PDF file

Returns:

{
  "success": true,
  "source": "gemini|deepseek|mock",
  "summary": "...",
  "risk_score": 78,
  "risk_level": "high",
  "risk_factors": [...],
  "barriers": [...],
  "interventions": [...],
  "risk_drivers": [...],
  "rag_sources": [...],
  "audit": { "documents_scanned": 3, "chunks_used": 3, ... }
}

Deployment Status

This is a hackathon MVP built in 48 hours — functional, deployed, and ready for pilot validation, but not yet production-grade. Specifically:

What works today:

• ✅ Deployed live on Vercel with working AI inference
• ✅ End-to-end pipeline: text/PDF input → risk analysis → source-attributed output
• ✅ Graceful fallback cascade (Gemini → DeepSeek → deterministic mock)
• ✅ Clinician-in-the-loop design — outputs are decision-support, never decision-making

What's not yet built (post-hackathon roadmap):

• ⏳ Real FHIR integration (currently uses paste/PDF input for demo)
• ⏳ NABIDH integration (Phase 1 of DHA pilot plan)
• ⏳ Authentication and role-based access control
• ⏳ Full HIPAA / UAE Federal Law No. 2 of 2019 compliance audit
• ⏳ Automated test suite
• ⏳ Embedding-based retrieval (currently TF-IDF for deployment size constraints)
• ⏳ Multi-language support (Arabic UI)

Roadmap

Phase 1 — NABIDH Integration (Q3 2026) Read-only connection to Dubai's Health Information Exchange to pull patient records securely.

Phase 2 — DHA Hospital Pilot (Q4 2026 – Q1 2027) Live pilot in a DHA endocrinology unit, tracking high-risk diabetic patients post-discharge. Primary endpoint: 30-day readmission rate.

Phase 3 — Scale (2027+) Expand to private UAE facilities, then international health systems.

Team — Cache Hit Dubai

• Riya — Health Data Science (EHR modelling, health economics)
• Om — AI Engineering (RAG architectures, cloud ML)
• Wenhui Yang — Computer Science (AI/ML)
• Xu Liu — Data pipelines, analytical rigour

Built at the HSIL Dubai hub, April 10–11, 2026.

Ethics & Safeguards

• Clinician-in-the-loop — every AI recommendation is reviewed by a human before action
• Source-attributed outputs — every insight links back to the retrieved record it came from
• No data egress — patient data stays within the authorised institutional boundary
• Bias auditing — planned for every model version against demographic subgroups
• UAE regulatory alignment — designed to operate within UAE Federal Law No. 2 of 2019 and DHA data governance

References

1. Ambade, P., Hoffman, Z., Mehra, K., Gunja, M., Yi, M., MacKinnon, B.H. and MacKinnon, N.J. (2025) 'Hospital discharge communication problems in 10 high-income nations: a secondary analysis of an international health policy survey', BMJ Open, 15, e094724. doi: 10.1136/bmjopen-2024-094724.
2. van Walraven, C., Bennett, C., Jennings, A., Austin, P.C. and Forster, A.J. (2011) 'Proportion of hospital readmissions deemed avoidable: a systematic review', CMAJ, 183(7), pp. E391–E402. doi: 10.1503/cmaj.101860.
3. Laurent, A. (2025) 'Hospital readmission rates by state: US data & analysis', IntuitionLabs, 11 November. Revised 13 March 2026.

License

This project was built for the HSIL Hackathon 2026. Licensing TBD pending Bootcamp progression.

---

"Ensuring no patient is left behind after leaving the hospital doors."