Development of an efficient deep learning model for arrhythmia detection from consumer smartwatches
George DC, Hulubeac MA, Grămadă ŞL (equal contribution), Bálint Z, Dioşan L (Babeş-Bolyai University, Cluj-Napoca, Romania)
Summary
Compares a feature-based Random Forest with two end-to-end models (CNN-GRU, CNN-BiLSTM) for 3-class rhythm (NSR, AF, PAC/PVC) on 30 s windows of real-life smartwatch data from the Pulsewatch trial. Input is a 2-channel series: wrist PPG at 50 Hz plus a scalar accelerometer magnitude (the release ships ACC as one value per sample). Devices per the paper: Samsung Gear S3 or Galaxy Watch 3 with a chest ECG patch (250 Hz) as reference.
Key facts
- Form Factor
- wrist
- License
- Not stated
Reported benchmarks
AFib / arrhythmia
Pulsewatch · AUROC ↑
Same protocol; macro (one-vs-rest) ROC-AUC, Random Forest; not AF-vs-rest binary
AFib / arrhythmia
Pulsewatch · Macro-F1 ↑
3-class NSR/AF/PAC-PVC, 30 s, 2-fold subject-independent (72 subj); macro AUROC 0.970, acc 0.912; CNN-BiLSTM macro recall 0.874; only rows on this axis
Updated: Oct 13, 2026
Related Research
- AFib / arrhythmia leaderboard
Benchmark results for afib / arrhythmia
- LIMA-Mamba: A lightweight and distribution-robust Mamba network for PPG-based biometrics
Research paper
- A Comparative Study of Signal Representations Methods and Deep Learning Architectures for PPG-Based Cuffless Blood Pressure Estimation
Seven 1D→2D transforms × several CNN variants on the UCI BP database. Best: CWT, errors reported as
- A Decodability-Grounded Audit of Clinical Prediction Claims in PPG Foundation Models
Audits frozen representations of four encoders on 37,784 MC-MED encounters via concept decodability,