ChartLens

FastAPI backend handles extraction via Google's langextract library with few-shot examples. Notes are optionally de-identified using Microsoft Presidio before being sent to cloud LLM providers (Gemini, OpenAI). Extracted entities are re-identified and char intervals remapped to original text coordinates. Next.js frontend provides tabbed views for browsing structured data with source highlighting.

Clinical notes are parsed into 10 extraction classes: diagnosis, medication, lab_result, vital_sign, imaging, procedure, assessment, event, allergy, and social_history. Each extraction includes the source text, character interval for highlighting, and structured attributes (e.g., dose, route, frequency for medications). A summary builder groups extractions into clinical categories with timeline sorting.

For cloud providers, notes pass through a Presidio-based de-identification layer that detects PERSON, DATE_TIME, US_SSN, PHONE_NUMBER, EMAIL_ADDRESS, LOCATION, MEDICAL_RECORD_NUMBER, and AGE_OVER_89 entities. Custom pattern recognizers handle MRN formats and HIPAA age thresholds. Offset shift tracking enables accurate re-identification and char interval remapping after extraction.

When using Ollama (local models), notes never leave the machine — de-identification is skipped entirely. The app auto-discovers available Ollama models via the local API, making it easy to run fully private chart review with no cloud dependencies.