rPPG Vitals
Camera-based vital signs. Patented.
- 8×
- Inference speedup
- 0.8+
- AUC health models
- 50GB
- Video processed
The problem
Vital-sign monitoring usually means hardware: cuffs, wearables, finger clips. For a digital-health product, every extra device is friction between the user and their own health data. The question behind this work: how much can a phone camera alone tell you?
Remote photoplethysmography (rPPG) recovers the blood-volume pulse from subtle color changes in facial video. This system which was granted as US Patent 12,525,360 extracts rPPG signals from facial video and uses them to model metabolic health risks including hypertension, cholesterol, and diabetes. This was my work at Hibiscus Health.
The pipeline
The system processes facial video through signal extraction, quality filtering, and downstream risk models:
- Signal extraction: rPPG waveforms recovered from facial regions across 50GB of facial video.
- Multi-threaded inference: parallelizing the pipeline delivered an 8× inference speedup, which is the difference between a research artifact and a feature that ships inside a mobile app.
- Risk modeling: trained and evaluated models predicting hypertension, cholesterol, and diabetes risk, reaching AUCs above 0.8.
Everything runs behind a HIPAA-compliant ETL pipeline and deploys via AWS Lambda, feeding a live mobile app serving thousands of users.
Results
- US Patent 12,525,360 — granted January 2026, sole technical inventor
- 8× faster inference via multi-threading
- 0.8+ AUC across health-risk models
- Deployed to a live mobile app via AWS Lambda
Read more
The full system is described in the granted patent: Systems and methods for assessing health measurements from facial video data.