← All work Research · MIT · First-author publication

Circadian Phenotyping

Digital phenotyping pipelines that infer circadian rhythm and behavioral states from smartphone and wearable sensors, deployed on-device, HIPAA-compliant, and validated in two MIT clinical studies.

Bayesian networks Deep learning TensorFlow Lite On-device inference Clinical validation

Summary

At MIT, across CSAIL, the Media Lab, and the Mobile Technology Lab, I built digital phenotyping pipelines that turn passive smartphone and wearable signals into clinically meaningful markers of sleep, stress, and behavior. The headline result: inferring circadian rhythm from phone sensors at accuracy on par with commercial sleep wearables like Oura, with everything packaged to run privately on the device.

The method

The pipeline combines Bayesian networks and deep learning to infer circadian rhythm and fine-grained smartphone-interaction states from raw sensor streams such as GPS, accelerometer, and battery. Models are packaged with TensorFlow Lite for HIPAA-compliant, on-device inference, so sensitive behavioral data never has to leave the phone. The work spans the full stack: signal processing, feature engineering, model training, and deployment.

Validation

2
MIT clinical studies led
≈ Oura
circadian accuracy parity
On-device
HIPAA-compliant inference
1st author
peer-reviewed publication

I led two MIT clinical studies that ended in a first-author, peer-reviewed publication. The circadian inference reached accuracy on par with dedicated commercial sleep wearables, using only the sensors already in a person's pocket.

Why it matters

Sleep, stress, and circadian disruption are upstream of a wide range of health outcomes, yet measuring them at scale usually requires dedicated hardware. Phenotyping from the phone makes continuous, long-term monitoring available to anyone, and running the inference on-device keeps it private. This research connects my work in wearables, mobile health, and the health mission behind Avenir.