ClearanceAI

Inspiration

It started with a phone call.

A close friend of mine had just undergone abdominal surgery and was discharged after what seemed like a smooth recovery. Three days later, he was back in the Emergency Department — dehydrated, febrile, and in considerably more pain than when he first went in. The surgical team had caught a post-op complication that, in hindsight, had been quietly waving red flags in his vitals and labs before he was ever sent home. He recovered fully, but the experience shook me.

When I later learned that 1 in 5 Medicare patients is readmitted within 30 days of discharge, and that many of these are considered preventable, I couldn't stop thinking about him. The clinical data was there. The signs were there. What was missing was a way to synthesize them clearly — before the patient walked out the door.

What it does

ClearanceAI helps hospitals prevent premature patient discharge and unnecessary 30-day readmissions by giving clinicians an AI-powered second opinion at the moment of discharge.

For each admitted patient, it:

Computes a 0–100 readmission risk score using a hybrid engine — 13 validated clinical signals (vitals, labs, comorbidities, length of stay) combined with Claude AI's analysis of free-text clinical notes Generates a structured discharge readiness checklist across 7 domains: vitals stability, mobility, medications, social support, follow-up, wound care, and patient education Displays a live ward dashboard showing occupancy, high-risk patient alerts, discharge queue, and 30-day readmission rate

How we built it

Challenges we ran into

• Calibrating the hybrid score — Finding the right 45/55 weighting between rule-based and AI components required iterative testing against synthetic patient profiles to ensure the composite score was clinically plausible across low, moderate, high, and critical risk cases
• Graceful AI degradation — Building seamless fallback to rule-based-only scoring when the AI model is unavailable, with no hard failures and still a valid clinical output
• Async SQLite in FastAPI — SQLAlchemy async sessions with SQLite require careful connection management to avoid threading race conditions under concurrent API requests
• Making the demo feel real — Seeding believable, varied patient profiles across all four risk tiers so the dashboard tells a meaningful clinical story, not just dummy test data

Accomplishments that we're proud of

• A fully working end-to-end clinical decision support tool — hybrid risk scoring, AI discharge checklists, live ward dashboard, and full patient records — built and polished within the hackathon window
• A hybrid AI architecture that is both interpretable (structured rule outputs with feature importance) and powerful (LLM clinical reasoning), rather than a black-box model
• Zero-infrastructure deployment — the entire stack runs locally with a single Python command and npm run dev, making it immediately usable in any hospital IT environment
• A graceful fallback system ensuring the tool remains clinically useful even without an active AI API key

What we learned

How readmission risk is quantified in clinical practice (LACE index, HOSPITAL score) — and how to computationally approximate it by fusing structured vitals/labs with unstructured clinical notes That the most dangerous moment in a patient's hospital journey may be discharge itself — a 5-minute AI review at that inflection point could prevent days of suffering and tens of thousands of dollars in avoidable readmission costs How much critical signal lives in unstructured free-text notes that structured fields completely miss — the LLM reads what the rule engine cannot Practical lessons building production-quality async FastAPI services and a real-time React dashboard under tight time pressure

What's next for ClearanceAI

EHR integration — Connect to real hospital systems via HL7 FHIR APIs so patient data flows in automatically, no manual entry On-premise AI — Fine-tune a BioGPT model on historical readmission datasets for fully local, privacy-preserving inference — no data ever leaves the hospital network Clinician feedback loop — Allow physicians to confirm or override AI recommendations, feeding corrections back into model calibration over time Predictive alerting — Push notifications when a patient's overnight vitals or labs push their risk score past a critical threshold Outcome tracking — Close the loop by tracking 30-day post-discharge outcomes and measuring whether ClearanceAI-flagged patients had lower readmission rates

Built With

• anthropicapi
• fastapi
• node.js
• npm
• pip
• pydantic
• python
• query
• react
• recharts
• sqlalchemy
• sqlite
• tanstack
• typescript
• uvicorn
• vite