Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Multiclass Arrhythmia Classification using Smartwatch Photoplethysmography Signals Collected in Real-life Settings

Dong Han, Jihye Moon, Luís Roberto Mercado Díaz, Darren Chen, Devan Williams, Eric Y. Ding, Khanh-Van Tran, David D. McManus, Ki H. Chon

TL;DR

The paper tackles multiclass arrhythmia classification from wearable smartwatch PPG in real-life settings, addressing motion artifacts and PAC/PVC discrimination alongside AF detection. It introduces a lightweight multimodal approach using PPG, heart rate, and accelerometer data fed into a 1D-Bi-GRU, optimized for real-time wearable deployment. On real-world Pulsewatch data, it achieves a PAC/PVC sensitivity around 83% and AF detection accuracy around 97%, outperforming prior methods while using significantly fewer parameters and lower computational cost. The work demonstrates the practical potential of real-time, multimodal wearables for accurate, accessible arrhythmia monitoring outside clinical environments.

Abstract

Most deep learning models of multiclass arrhythmia classification are tested on fingertip photoplethysmographic (PPG) data, which has higher signal-to-noise ratios compared to smartwatch-derived PPG, and the best reported sensitivity value for premature atrial/ventricular contraction (PAC/PVC) detection is only 75%. To improve upon PAC/PVC detection sensitivity while maintaining high AF detection, we use multi-modal data which incorporates 1D PPG, accelerometers, and heart rate data as the inputs to a computationally efficient 1D bi-directional Gated Recurrent Unit (1D-Bi-GRU) model to detect three arrhythmia classes. We used motion-artifact prone smartwatch PPG data from the NIH-funded Pulsewatch clinical trial. Our multimodal model tested on 72 subjects achieved an unprecedented 83% sensitivity for PAC/PVC detection while maintaining a high accuracy of 97.31% for AF detection. These results outperformed the best state-of-the-art model by 20.81% for PAC/PVC and 2.55% for AF detection even while our model was computationally more efficient (14 times lighter and 2.7 faster).

Multiclass Arrhythmia Classification using Smartwatch Photoplethysmography Signals Collected in Real-life Settings

TL;DR

The paper tackles multiclass arrhythmia classification from wearable smartwatch PPG in real-life settings, addressing motion artifacts and PAC/PVC discrimination alongside AF detection. It introduces a lightweight multimodal approach using PPG, heart rate, and accelerometer data fed into a 1D-Bi-GRU, optimized for real-time wearable deployment. On real-world Pulsewatch data, it achieves a PAC/PVC sensitivity around 83% and AF detection accuracy around 97%, outperforming prior methods while using significantly fewer parameters and lower computational cost. The work demonstrates the practical potential of real-time, multimodal wearables for accurate, accessible arrhythmia monitoring outside clinical environments.

Abstract

Most deep learning models of multiclass arrhythmia classification are tested on fingertip photoplethysmographic (PPG) data, which has higher signal-to-noise ratios compared to smartwatch-derived PPG, and the best reported sensitivity value for premature atrial/ventricular contraction (PAC/PVC) detection is only 75%. To improve upon PAC/PVC detection sensitivity while maintaining high AF detection, we use multi-modal data which incorporates 1D PPG, accelerometers, and heart rate data as the inputs to a computationally efficient 1D bi-directional Gated Recurrent Unit (1D-Bi-GRU) model to detect three arrhythmia classes. We used motion-artifact prone smartwatch PPG data from the NIH-funded Pulsewatch clinical trial. Our multimodal model tested on 72 subjects achieved an unprecedented 83% sensitivity for PAC/PVC detection while maintaining a high accuracy of 97.31% for AF detection. These results outperformed the best state-of-the-art model by 20.81% for PAC/PVC and 2.55% for AF detection even while our model was computationally more efficient (14 times lighter and 2.7 faster).
Paper Structure (15 sections, 2 figures, 3 tables)

This paper contains 15 sections, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Dataset Description
  3. Pulsewatch dataset
  4. Training and testing datasets
  5. Methods
  6. Signal Processing of the Time-Series Data
  7. 1D Time Series Data Preparation
  8. Extraction of PPG heart rates (HR)
  9. Machine Learning Model Design: 1D-Bi-GRU Model
  10. Machine Learning Model Training Process
  11. Evaluation Metrics
  12. Results and Discussion
  13. Effectiveness of using HR as an Input for Multiclass Arrhythmia Classification
  14. Efficiency in Computational Cost using HR as an Input
  15. CONCLUSION

Figures (2)

  • Figure 1: Example 30-sec segments from (a) NSR, (b) AF, and (c) PAC/PVC with less than 5 sec of motion noise artifact recorded from the Pulsewatch system. Rows: (1) reference ECG, (2) filtered PPG, (3) heart rates (HR) calculated from the peaks of ECG and PPG, and the interpolated PPG HR that was used for our proposed model, (4) normalized PPG HR with minimum and maximum values from 30 to 220 BPM range, (5) “zoom-in” PPG HR for regular heart rate to better accentuate dynamic ranges, and (6) accelerometer (ACC) signals. Signals from rows (2), (4), (5), and (6) were used in our best proposed model.
  • Figure 2: The architecture of our proposed 1D bi-directional Gated Recurrent Unit (1D-Bi-GRU) model.