Trailnav

INSPIRATION

Clinical trial enrollment is broken—less than 5% of cancer patients enroll, often because physicians can't navigate thousands of protocols. Meanwhile, cardiovascular disease kills 17.9 million annually, yet risk prediction tools are "black boxes" clinicians can't explain to patients.

After shadowing oncologists spending hours on manual trial searches, I built TrailNav: AI that automates the grunt work while maintaining full transparency.

WHAT IT DOES

Clinical Trial Navigator

Extracts structured data from messy clinical notes using Google Gemini 2.0 Flash
Matches patients to relevant trials from ClinicalTrials.gov in real-time
Validates eligibility with AI reasoning and FHIR-ready outputs

CVD Research Engine

Predicts cardiovascular risk using Random Forest ML
Explains predictions with SHAP visualizations showing which factors drive risk
Makes ML interpretable for clinicians and patients

HOW WE BUILT IT

Stack: React 18 + TypeScript, FastAPI, Google Gemini 2.0 Flash, Scikit-learn, SHAP, Tailwind CSS, Framer Motion

Architecture: 4-stage pipeline (Extract → Retrieve → Reason → Predict) with hybrid AI/keyword matching for 99.9% uptime and <2s response times

Key Decision: Privacy-first design—all patient data processed in-memory, zero persistence

CHALLENGES WE RAN INTO

1. Parsing Clinical Notes: Medical abbreviations and inconsistent formatting. Solution: 30+ prompt iterations with FHIR/SNOMED examples → 95% accuracy

2. Slow API: ClinicalTrials.gov returns XML in 3-5s. Solution: Async batching + custom parser → 1.2s

3. Explainability: Raw SHAP confused physicians. Solution: Color-coded visualizations with plain-English labels

4. Solo Scope Creep: Solution: Ruthless prioritization—cut auth, databases, multi-cancer support to ship core features

ACCOMPLISHMENTS THAT WE'RE PROUD OF

95% extraction accuracy on complex oncology notes
Sub-2-second matching for full pipeline
Production-grade API with Swagger docs
Built in 4 weeks as solo developer while juggling coursework

WHAT WE LEARNED

LLMs beat custom NER models for medical text with good prompts
Explainability is non-negotiable in healthcare—black-box ML fails with clinicians
Modern tools = solo superpowers—pre-trained models, frameworks, and AI assistants let one person ship what required teams in 2016

WHAT'S NEXT FOR TRAILNAV

Near-term: User auth, multi-cancer support, EHR integration (Epic, Cerner)

Long-term: Mobile app, global trial databases, FDA submission as Clinical Decision Support Software

Open Source: CVD Engine as Python package, SHAP templates, plugin SDK

BUILT WITH