Arrhythmia Detection from Smartwatch PPG
Detecting and counting ectopic beats from consumer smartwatch photoplethysmography at scale.
Photoplethysmography from a consumer smartwatch is noisy, motion-corrupted, and indirect compared to an ECG — but it is worn continuously by millions of people, which makes it uniquely suited to catching intermittent cardiac events that a clinical recording would miss.
This work, done with the Digital Health team at Samsung Research America, asked whether ectopic beats can be reliably detected and counted from wrist PPG alone.
Contributions
- Designed and implemented a scalable end-to-end data pipeline to ingest, preprocess, and manage over 1 million real-world smartwatch PPG segments, enabling large-scale training and evaluation.
- Built and optimized deep learning models — ResNet, VGG, and attention-based architectures — reaching >85% F1 for arrhythmia detection while reducing per-segment count estimation error to <0.05 MAE.
- Integrated arrhythmia detection as a task in a multimodal healthcare foundation model, contributing reusable components to an internal ML platform used by several downstream teams.
Published at ICASSP 2026 (Lu et al., 2026).
Work conducted during an industry internship; described here at the level of publicly shareable methods and outcomes.