Ultra lightweight residual network for multiclass arrhythmia classification using raw smartwatch PPG signals

Tiny-ResNet (stem + 2 residual blocks, 23,727 parameters) on raw single-channel smartwatch PPG, no HR/ACC inputs, for AF vs PAC/PVC vs normal on the Pulsewatch dataset (106 subjects, 14 days real-life). Test results: overall accuracy 98.00%; PAC/PVC sensitivity 93.96%, specificity 98.80%, F1 92%; AF sensitivity 99.20%, specificity 99.00%. Compared against 2D DenseNet (221,303 params) and retrained VGG-16 (1.6 M). Split protocol (subject-wise or not) not stated in the abstract.

• Form factor: wrist

Objective: Smartwatches with photoplethysmographic (PPG) sensors are ideal for continuous heart monitoring, but accurately distinguishing between Atrial Fibrillation (AF) and Premature Atrial/Ventricular Contractions (PAC/PVC) remains a challenge because these rhythms often have similar PPG waveforms. Many current state-of-the-art models rely on multimodal data, including heart rate (HR), magnified HR, and accelerometer (ACC) signals, to achieve high sensitivity, but this increases computational complexity. We propose an ultralightweight 1D residual network that uses only raw single-channel PPG signals for multiclass arrhythmia classification. Methods: We used the NIH-funded Pulsewatch dataset, which contains PPG data from 106 subjects recorded in real-life settings for 14 days. We developed a “Tiny-ResNet” architecture and tested different configurations with one, two, and three residual blocks. Our final model consists of a stem layer for initial feature extraction, followed by two residual blocks to capture longer temporal dependencies and morphological features. Results: Our model is extremely efficient with only 23,727 trainable parameters, making it significantly lighter than the 2D DenseNet with 221,303 parameters and the retrained VGG-16 model with 1.6 million parameters. In the test dataset, the model achieved an overall accuracy of 9 8. 0 0 %. For the difficult PAC/PVC class, we achieved a sensitivity of 93.96 %, specificity of 98.80 % and F1-score of 92 %. For AF detection, our model maintained a sensitivity of 99.20 % and specificity of 99.00 %. Conclusion: This study demonstrates that multiclass arrhythmia classification is achievable using only raw 1D PPG waveforms without the need for additional heart rate or accelerometer data. Due to its very small size and high accuracy, our model is well-suited for real-time, on-device deployment on smartwatches and other low-power wearables for continuous cardiac screening.

Verbatim from Semantic Scholar (opens in a new tab) · Abstract via Semantic Scholar; © the authors/publisher

No results from this paper on ChatPPG leaderboards yet. Browse the leaderboards.

• Pulsewatchambulatory smartwatch PPG with AF adjudication (Synapse gated)TrainingEvaluation106 subjects · PPG · ECG · IMU