Cross-Domain Joint Dictionary Learning for ECG Inference From PPG

Xin Tian, Qiang Zhu, Yuenan Li, Min Wu

Abstract

The inverse problem of inferring clinical gold-standard electrocardiogram (ECG) from photoplethysmogram (PPG) that can be measured by affordable wearable Internet of Healthcare Things (IoHT) devices is a research direction receiving growing attention. It combines the easy measurability of PPG and the rich clinical knowledge of ECG for long-term continuous cardiac monitoring. The prior art for reconstruction using a universal basis, such as discrete cosine transform (DCT), has limited fidelity for uncommon ECG shapes due to the lack of representative power. To better utilize the data and improve data representation, we design two dictionary learning frameworks, the cross-domain joint dictionary learning (XDJDL), and the label-consistent XDJDL (LC-XDJDL), to further improve the ECG inference quality and enrich the PPG-based diagnosis knowledge. Building on the K-SVD technique, the proposed joint dictionary learning frameworks extend the expressive power by optimizing simultaneously a pair of signal dictionaries for PPG and ECG with the transforms to relate their sparse codes and disease information. The proposed models are evaluated with a variety of PPG and ECG morphologies from two benchmark datasets that cover various age groups and disease types. The results show the proposed frameworks achieve better inference performance than previous methods with average Pearson coefficients being 0.88 using XDJDL and 0.92 using LC-XDJDL, suggesting an encouraging potential for ECG screening using PPG based on the proactively learned PPG-ECG relationship.

Introduction

While ECG is the gold standard for diagnosing heart conditions, it usually requires sticky electrodes and bulky equipment, making long-term continuous monitoring difficult. Photoplethysmogram (PPG) sensors, conversely, are ubiquitous in modern wearables but lack the rich diagnostic detail of an ECG, such as the specific shapes of P-waves or the QRS complex. Reconstructing high-fidelity ECG signals from PPG data—an "inverse problem" in digital health—could bridge this gap, combining the convenience of wearables with clinical-grade diagnostic power.

The challenge lies in the non-linear and individually variable relationship between the heart's electrical activity (ECG) and resulting blood volume changes (PPG). Previous attempts using fixed mathematical bases like the Discrete Cosine Transform (DCT) were often too rigid to capture the diverse waveform morphologies found in clinical settings, especially for patients with cardiovascular diseases.

Methodology: Joint Dictionary Learning

To address these limitations, this research proposes a Cross-domain Joint Dictionary Learning (XDJDL) framework. Unlike fixed-basis methods, XDJDL learns custom "dictionaries"—one for PPG ($D_p$) and one for ECG ($D_e$)—simultaneously with a linear transform ($W$) that maps the sparse representation of a PPG cycle to its corresponding ECG cycle.

The end-to-end pipeline of the proposed ECG inference system, including preprocessing (alignment, segmentation, scaling), joint dictionary learning, and signal reconstruction.

The framework is further enhanced in the LC-XDJDL (Label-Consistent XDJDL) version, which incorporates disease labels (e.g., Myocardial Infarction) into the learning process. This forces the model to learn sparse codes that are not only accurate for reconstruction but also consistent with specific medical diagnoses. The optimization relies on the K-SVD algorithm, adapted here to handle two signal domains concurrently.

Experimental Setup

The study utilized the MIMIC-III (Medical Information Mart for Intensive Care) database, focusing on 33 patients with four types of cardiovascular diseases: Congestive Heart Failure (CHF), Myocardial Infarction (MI), Hypotension (HYPO), and Coronary Artery Disease (CAD). The dataset comprised over 34,000 individual ECG/PPG cycle pairs. Signals underwent rigorous preprocessing, including detrending and temporal scaling to a uniform length, ensuring the dictionaries captured morphological features rather than just heart rate.

Key Results and Discussion

The proposed XDJDL framework significantly outperformed previous state-of-the-art methods across multiple metrics:

  • Correlation: XDJDL achieved a Pearson correlation ($\rho$) of 0.88, compared to 0.71 for the DCT method.
  • Error Reduction: The relative Root Mean Squared Error (rRMSE) dropped from 0.67 (DCT) to 0.39 (XDJDL).
  • Label Boost: With ground-truth disease labels (LC2-XDJDL), the correlation climbed to 0.92.
  • Clinical Fidelity: The model was particularly successful at reconstructing the QRS complex, achieving a correlation of 0.94 in optimal scenarios.
Qualitative comparison of ECG reconstruction across different patient pathologies (CHF, MI, CAD) using various baseline methods versus the proposed XDJDL framework.

As shown in the qualitative results, XDJDL (bottom row) tracks the reference ECG (dashed red line) far more accurately than DCT or other dictionary methods, particularly in capturing the sharp peaks of the QRS complex and the subtle curves of the T-wave across different disease types.

Conclusion and Future Work

By enabling dynamic monitoring and analysis of an individual's health status, these frameworks contribute to the emerging digital twins paradigm for personalized healthcare. However, challenges remain regarding motion artifacts in real-world ambulatory settings and subject independence; testing on completely new patients showed a performance dip ($\rho = 0.74$). Future work will explore deep neural networks to further improve expressive power and incorporate robust denoising steps for practical wearable deployment.

Summary updated

Abstract

The inverse problem of inferring clinical gold-standard electrocardiogram (ECG) from photoplethysmogram (PPG) that can be measured by affordable wearable Internet of Healthcare Things (IoHT) devices is a research direction receiving growing attention. It combines the easy measurability of PPG and the rich clinical knowledge of ECG for long-term continuous cardiac monitoring. The prior art for reconstruction using a universal basis, such as discrete cosine transform (DCT), has limited fidelity for uncommon ECG shapes due to the lack of representative power. To better utilize the data and improve data representation, we design two dictionary learning frameworks, the cross-domain joint dictionary learning (XDJDL), and the label-consistent XDJDL (LC-XDJDL), to further improve the ECG inference quality and enrich the PPG-based diagnosis knowledge. Building on the K-SVD technique, the proposed joint dictionary learning frameworks extend the expressive power by optimizing simultaneously a pair of signal dictionaries for PPG and ECG with the transforms to relate their sparse codes and disease information. The proposed models are evaluated with a variety of PPG and ECG morphologies from two benchmark datasets that cover various age groups and disease types. The results show the proposed frameworks achieve better inference performance than previous methods with average Pearson coefficients being 0.88 using XDJDL and 0.92 using LC-XDJDL, suggesting an encouraging potential for ECG screening using PPG based on the proactively learned PPG-ECG relationship. By enabling the dynamic monitoring and analysis of the health status of an individual, the proposed frameworks contribute to the emerging digital twins paradigm for personalized healthcare.
BibTeX
@article{tian2022cross,
  author = {Xin Tian and Qiang Zhu and Yuenan Li and Min Wu},
  title = {{Cross-Domain Joint Dictionary Learning for ECG Inference From PPG}},
  journal = {IEEE Internet of Things Journal},
  year = {2022},
  doi = {10.1109/jiot.2022.3231862},
}
Figure