What I do
I synthesize lesion detection results with historical similar cases to generate a comprehensive risk profile. I use WebLLM (SmolLM2) for offline inference to calculate risk scores with equalized odds correction.
When to use me
Use this when:
- •Similar case search is complete and you need risk scoring
- •You need a numerical risk assessment for clinical decision support
- •You're combining multiple signals into an overall risk profile
Key Concepts
- •WebLLM: Browser-based LLM for offline inference
- •SmolLM2: Efficient LLM model for risk synthesis
- •Equalized Odds Correction: Fairness-aware risk calibration
- •Risk Score: Numerical assessment (Low/Medium/High)
- •risk_assessed: State flag after assessment complete
Source Files
- •
services/vision.ts: Risk assessment implementation - •
types.ts: AnalysisResult interface
Code Patterns
- •Synthesize lesion data with historical patterns
- •Apply equalized odds correction for demographic fairness
- •Return risk level (Low/Medium/High) with supporting evidence
Operational Constraints
- •Must provide equalized odds across demographics
- •Heavy model - must expose unload() method
- •Confidence scores required for all risk assessments