Inferring ECG from PPG for Continuous Cardiac Monitoring Using Lightweight Neural Network

li yuenan, Xin Tian, Qiang Zhu, Min WuarXiv 2020 · DOI · arXiv

Abstract

This paper presents a computational solution that enables continuous cardiac monitoring through cross-modality inference of electrocardiogram (ECG). While some smartwatches now allow users to obtain a 30-second ECG test by tapping a built-in bio-sensor, these short-term ECG tests often miss intermittent and asymptomatic abnormalities of cardiac functions. It is also infeasible to expect persistently active user participation for long-term continuous cardiac monitoring in order to capture these and other types of cardiac abnormalities. To alleviate the need for continuous user attention and active participation, we design a lightweight neural network that infers ECG from the photoplethysmogram (PPG) signal sensed at the skin surface by a wearable optical sensor. We also develop a diagnosis-oriented training strategy to enable the neural network to capture the pathological features of ECG, aiming to increase the utility of reconstructed ECG signals for screening cardiovascular diseases (CVDs). We also leverage model interpretation to obtain insights from data-driven models, for example, to reveal some associations between CVDs and ECG/PPG and to demonstrate how the neural network copes with motion artifacts in the ambulatory application. The experimental results on three datasets demonstrate the feasibility of inferring ECG from PPG, achieving a high fidelity of ECG reconstruction with only about 40K parameters.

Introduction

Electrocardiogram (ECG) monitoring is the gold standard for diagnosing cardiovascular diseases (CVDs). However, continuous monitoring remains a significant hurdle. Traditional Holter monitors are cumbersome, and modern smartwatch ECG features require active user participation—stopping all activity to touch a sensor for 30 seconds. This requirement often results in missing intermittent or asymptomatic cardiac events that occur during daily life.

In contrast, Photoplethysmogram (PPG) sensors, commonly found in wearables as green optical lights, monitor blood volume changes continuously and passively. The challenge lies in "cross-modality inference": transforming the simpler, mechanical PPG signal into the complex, electrical detail of an ECG waveform.

Comparison of ECG and PPG waveforms during a single cardiac cycle.

The PPG2ECG Network

To bridge the gap between PPG and ECG, the researchers developed the PPG2ECG network, a lightweight architecture designed for resource-constrained wearable devices. The network employs a specialized encoder-decoder structure. To minimize computational overhead, it utilizes grouped convolutions and channel shuffling, allowing the model to capture multi-scale features—such as the brief P-wave versus the longer blood ejection phase—with only approximately 40,000 parameters.

The PPG2ECG network architecture and diagnosis-oriented training (DOT) strategy.

Key Components

  • Dual Attention Gates: The model uses both temporal and channel attention to highlight informative parts of the PPG cycle and re-weight features based on their importance for reconstruction.
  • Diagnosis-Oriented Training (DOT): Beyond standard visual similarity (Mean Squared Error), the authors introduced a "feature loss" and "contrastive loss" based on a pre-trained CVD classifier. This ensures the reconstructed ECG preserves pathological markers, such as ST-segment changes, critical for clinical diagnosis.
  • Knowledge Distillation: A teacher-student framework was used to compress the model further, reducing parameter count by over 60% while maintaining high fidelity.

Experimental Results

The model was validated across three datasets: MIMIC-III (clinical data), BIDMC (generalizability), and a Motion PPG-ECG dataset (ambulatory performance).

Performance and Efficiency

  • High Fidelity: The model achieved a Pearson correlation ($\rho$) of 0.94 on the MIMIC-III dataset, indicating exceptional similarity to ground-truth ECGs.
  • Clinical Utility: 93% of reconstructed ECGs were correctly classified by a CVD diagnostic tool, with $F_1$ scores exceeding 0.95 for conditions like Myocardial Infarction and Heart Failure.
  • Resource Efficiency: With only 40K parameters and 28 million FLOPs per cycle, the model is roughly 270 times smaller than previous state-of-the-art models like TANN, making it ideal for real-time on-device processing.

Handling Motion Artifacts

In ambulatory settings, motion can significantly degrade PPG signals. By incorporating 3-axis accelerometer data as an auxiliary input, the network learned to effectively "subtract" motion noise, with motion accounting for approximately 27% of the signal's influence during activities like running.

Conclusion and Future Directions

The PPG2ECG network demonstrates that the mechanical activity of the heart captured by PPG is intrinsically linked to its electrical activity. By prioritizing diagnostic utility over simple visual similarity, this approach enables continuous, passive cardiac monitoring. Future work will focus on leveraging unsupervised learning for unpaired data and integrating user metadata (age, weight, history) to further personalize and enhance reconstruction accuracy.

Reference: Li, Y., Tian, X., Zhu, Q., & Wu, M. (2024). Inferring Electrocardiography From Optical Sensing Using Lightweight Neural Network. IEEE Transactions on Artificial Intelligence.

Summary updated

Abstract

This paper presents a computational solution that enables continuous cardiac monitoring through cross-modality inference of electrocardiogram (ECG). While some smartwatches now allow users to obtain a 30-second ECG test by tapping a built-in bio-sensor, these short-term ECG tests often miss intermittent and asymptomatic abnormalities of cardiac functions. It is also infeasible to expect persistently active user participation for long-term continuous cardiac monitoring in order to capture these and other types of cardiac abnormalities. To alleviate the need for continuous user attention and active participation, we design a lightweight neural network that infers ECG from the photoplethysmogram (PPG) signal sensed at the skin surface by a wearable optical sensor. We also develop a diagnosis-oriented training strategy to enable the neural network to capture the pathological features of ECG, aiming to increase the utility of reconstructed ECG signals for screening cardiovascular diseases (CVDs). We also leverage model interpretation to obtain insights from data-driven models, for example, to reveal some associations between CVDs and ECG/PPG and to demonstrate how the neural network copes with motion artifacts in the ambulatory application. The experimental results on three datasets demonstrate the feasibility of inferring ECG from PPG, achieving a high fidelity of ECG reconstruction with only about 40K parameters.
BibTeX
@misc{yuenan2020inferring,
  author = {li yuenan and Xin Tian and Qiang Zhu and Min Wu},
  title = {{Inferring ECG from PPG for Continuous Cardiac Monitoring Using Lightweight Neural Network}},
  year = {2020},
  eprint = {2012.04949},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2012.04949},
}
Figure