Log-Spectral Matching GAN: PPG-Based Atrial Fibrillation Detection can be Enhanced by GAN-Based Data Augmentation With Integration of Spectral Loss

Cheng Ding, Ran Xiao, Duc H. Do, David Scott Lee, Randall J. Lee, Shadi Kalantarian, Xiao HuIEEE JBHI 2023 · OpenReview

TL;DR

This paper introduces LSM-GAN, a novel Generative Adversarial Network (GAN) for augmenting Photoplethysmography (PPG) data to improve Atrial Fibrillation (AF) detection. By incorporating spectral information into the GAN's loss function, LSM-GAN generates more realistic synthetic PPG signals, significantly boosting classifier performance, especially in imbalanced datasets.

What Problem Does This Paper Address?

Detecting Atrial Fibrillation (AF) using Photoplethysmography (PPG) signals holds great promise for accessible, continuous cardiovascular monitoring. However, a major hurdle in training accurate AF detection models is the inherent class imbalance in PPG datasets. AF is relatively rare in the general population, and its paroxysmal nature means AF episodes are often interspersed with normal heart rhythms. This imbalance can lead to models that are biased towards the majority class (non-AF) and perform poorly in identifying actual AF cases. Traditional data augmentation techniques can help, but they often struggle to generate realistic PPG signals that capture the subtle spectral characteristics indicative of AF.

Methods

The core innovation is the Log-Spectral Matching GAN (LSM-GAN). Unlike standard GANs that primarily rely on adversarial loss (matching real vs. generated data distributions), LSM-GAN introduces a Log-Spectral Matching (LSM) loss. This loss function explicitly penalizes differences in the frequency domain between real and synthetic PPG signal segments.

Key methodological aspects include:

  • Novel Generator Architecture: The generator is designed to produce synthetic signals of the same length as the input noise, avoiding an up-sampling step common in other GANs. This is hypothesized to lead to a more interpretable filtering process.
  • LSM Loss Components:
    • Matching Distance ($d_m$): Measures the spectral difference between corresponding blocks of real and synthetic PPG signals.
  • Self-Consistency Distance ($d_s$): Measures the spectral variation within a single synthetic PPG signal, ensuring internal consistency.
  • Aggregation Function (F): Either Mean or Max, used to aggregate $d_m$ and $d_s$ across blocks, allowing for task-specific tuning (e.g., emphasizing irregular AF patterns vs. consistent sinus rhythms).
  • Hyperparameter Optimization: A guided grid search approach is used to automatically select optimal weighting parameters ($\lambda_1, \lambda_2$) for the LSM loss components and the aggregation function (F), aiming to minimize the Maximum Mean Discrepancy (MMD) of signal autocorrelations.
  • Baseline Comparisons: The proposed LSM-GAN is compared against several baseline methods, including simple data augmentation (Data-Copying, Permutation) and other GAN architectures (DCGAN, W-DCGAN) with both conventional and the new generator architectures.

Datasets

  • Training Data: Continuous fingertip PPG recordings from 126 in-hospital patients (aged 18-95) collected at UCLA Medical Center. AF episodes were marked by a cardiologist based on co-registered ECG. Data was segmented into 30-second records labeled "AF" or "non-AF".
  • Testing Data: Continuous PPG data from wearable devices (Empatica E4) worn by 13 acute stroke patients in the UCSF Medical Center NICU (aged 19-91). This dataset was also segmented into 30-second records.
  • External Validation Datasets:
    • DeepBeat Dataset: A public dataset with PPG and AF annotations, collected from wrist-type watches. Differences in signal length, sampling rate, and collection site (wrist vs. fingertip) required harmonization.
  • MIMIC-III Waveform Database: Synchronized PPG and ECG data from 60 patients, annotated by cardiologists. PPG signals were 10-second strips, which were repeated to create 30-second segments for compatibility.

Key Insights and Ideas

  • Spectral Information is Crucial for Realistic PPG Augmentation: Incorporating spectral matching into the GAN's loss function significantly improves the realism of generated PPG signals, leading to better downstream classification performance. This suggests that capturing the frequency-domain characteristics is vital for mimicking physiological signals.
  • LSM-GAN Generates More Dispersed and Realistic Distributions: Visualization using PaCMAP shows that LSM-GAN generates synthetic AF and non-AF signals whose distributions are more similar to real signals compared to other GAN methods. This indicates a better capture of the underlying data manifold.
  • Generator Architecture Matters: Using a generator that avoids up-sampling and maintains input noise dimension leads to improved performance over conventional GAN architectures, even before adding the LSM loss.
  • Hyperparameter Tuning is Key: The proposed hyperparameter selection method, guided by MMD of autocorrelations, is shown to be effective in finding optimal settings for LSM-GAN, leading to further performance gains.
  • Robustness to Artifacts: While all methods struggle with highly artifact-ridden PPG signals, LSM-GAN demonstrates strong resilience and maintains a leading performance across various signal quality levels.

Key Results

  • Experiment 1 (Performance Comparison): LSM-GAN achieved the highest accuracy (0.9612) and sensitivity (0.9284) among all augmentation methods, outperforming even the improved DCGAN-1200 and W-DCGAN-1200 by approximately 3% in accuracy and 5% in sensitivity.
  • Experiment 2 (Resilience to Artifacts): LSM-GAN achieved an F1 score of 99% on perfect quality signals and maintained strong performance (96%, 90%, 72%) across decreasing signal quality groups, significantly outperforming the unaugmented model (91%, 76%, 73%, 38%).
  • Experiment 3 (Sample Size): LSM-GAN consistently showed an increasing trend in F1 score with larger training sample sizes, even in low-quality signal groups where other methods plateaued.
  • External Validation: LSM-GAN demonstrated superior performance on both the DeepBeat and MIMIC datasets compared to data-copying and other GANs, showing its generalizability.

Limitations and Future Work

  • Optimal Data Amount: The paper does not investigate the ideal amount of data required to train an effective LSM-GAN.
  • Broader Applications: The study focused solely on AF detection. Future work could explore LSM-GAN's effectiveness for detecting other cardiac arrhythmias using PPG.
  • Exploring Different Classifiers: While ResNet-50 was used for fair comparison, exploring other deep learning architectures could reveal further improvements.
  • Performance Plateau: Even with augmentation, performance plateaus after approximately one million training samples, suggesting that acquiring more real-world data might be necessary for further gains.
  • Interpretability of Filters: The authors note that understanding the learned filters in their novel generator architecture is an area for future research.

Summary updated

Abstract

Photoplethysmography (PPG) is a ubiquitous physiological measurement that detects beat-to-beat pulsatile blood volume changes and hence has a potential for monitoring cardiovascular conditions, particularly in ambulatory settings. A PPG dataset that is created for a particular use case is often imbalanced, due to a low prevalence of the pathological condition it targets to predict and the paroxysmal nature of the condition as well. To tackle this problem, we propose log-spectral matching GAN (LSM-GAN), a generative model that can be used as a data augmentation technique to alleviate the class imbalance in a PPG dataset to train a classifier. LSM-GAN utilizes a novel generator that generates a synthetic signal without a up-sampling process of input white noises, as well as adds the mismatch between real and synthetic signals in frequency domain to the conventional adversarial loss. In this study, experiments are designed focusing on examining how the influence of LSM-GAN as a data augmentation technique on one specific classification task - atrial fibrillation (AF) detection using PPG. We show that by taking spectral information into consideration, LSM-GAN as a data augmentation solution can generate more realistic PPG signals.
BibTeX
@article{ding2023log,
  author = {Cheng Ding and Ran Xiao and Duc H. Do and David Scott Lee and Randall J. Lee and Shadi Kalantarian and Xiao Hu},
  title = {{Log-Spectral Matching GAN: PPG-Based Atrial Fibrillation Detection can be Enhanced by GAN-Based Data Augmentation With Integration of Spectral Loss}},
  journal = {IEEE J. Biomed. Health Informatics},
  year = {2023},
  url = {https://openreview.net/forum?id=tml0t3P1za},
}
Figure