Building On-Device Health Models with Small, Medium, and Massive Datasets

Most models never make it from the notebook to the device - they're too big, too slow, too power-hungry, or trained on data that doesn't look like what the sensor actually sees. Drawing from my experience shipping women's health algorithms to millions of Oura users, building sensor models at Verily and Google, and advising early-stage health wearable startups, this talk will cover lessons learned from deploying on-device models when your dataset is small, medium, or large. We will delve into how label collection, validation, and deployment strategy change at each scale, using examples from fetal monitoring, wearable research cohorts, and clinical trials, and what breaks when you apply the wrong approach for the data you have. This session is for engineers taking ML models from the notebook to the device. 

Key Takeaways: 

1. What changes at each data scale: from expert-adjudicated recordings to weak labels from millions of users. 

2. Sensor, phone, or cloud: how to split inference with real latency and battery numbers. 3. The places your exported model silently diverges from the one you trained, and how to test for each.