Which Augmentation Should I Use? An Empirical Investigation of Augmentations for Self-Supervised Phonocardiogram Representation Learning

Aristotelis Ballas, Vasileios Papapanagiotou, Christos Diou

Abstract

Despite recent advancements in deep learning, its application in real-world medical settings, such as phonocardiogram (PCG) classification, remains limited. A significant barrier is the lack of high-quality annotated datasets, which hampers the development of robust, generalizable models that can perform well on newly collected, out-of-distribution (OOD) data. Self-Supervised Learning (SSL) contrastive learning has shown promise in mitigating the issue of data scarcity by using unlabeled data to enhance model robustness. This research explores and evaluates a wide range of audio-based augmentations to uncover combinations that enhance SSL model performance in PCG classification. Our findings reveal that augmentation choice significantly influences model robustness: while fully-supervised models experienced up to a 32% drop in effectiveness when evaluated on unseen data, SSL models demonstrated greater resilience, losing only 10% or even improving. This study highlights the most promising augmentations for PCG signal processing, providing valuable guidelines for developing reliable medical AI.

Introduction: The "Data Wall" in Heart Sound Analysis

Automating heart sound analysis (phonocardiogram or PCG classification) could revolutionize early cardiac screening, but it faces a significant "data wall." High-quality, annotated medical data is expensive and scarce. Furthermore, models trained on one dataset often fail when deployed in the real world—a phenomenon known as poor Out-of-Distribution (OOD) generalization. Differences in microphone placement, patient body type, or ambient room noise can render a model trained in a controlled environment ineffective in a clinical setting.

While Self-Supervised Learning (SSL)—specifically contrastive learning—has addressed similar problems in computer vision, its application to 1D medical audio is under-researched. The "secret sauce" of contrastive learning is data augmentation (creating two different "views" of the same signal), yet the optimal augmentation policies for heart sounds remained largely unknown until now.

Methodology and Experimental Framework

The researchers adapted the SimCLR framework for 1D PCG signals. The core objective is to train a "backbone" encoder to recognize that two differently augmented versions of the same heart sound are related, while versions from different recordings are distinct.

The six-step experimental pipeline for SSL pre-training and OOD evaluation.

The study utilized five distinct public PCG datasets to ensure diversity: Ephnogram, FPCGDB (Fetal PCG), Pascal (2011 Heart Sound Challenge), PhysioNet2016, and PhysioNet2022. By pre-training on a mix of these and testing on others, the team could strictly measure how well the models handled OOD data. Six primary augmentation types were evaluated: Noise, Cut-off Filters, Scale, Reverse, Inversion, and Random Flip. Over 150 combinations were explored to identify the optimal policy for heart sound representation.

Key Insights: Which Augmentations Matter?

  • Low-Pass Filters are King: The most significant finding was that low-pass filters (cutting off frequencies above 250–750 Hz) had the highest positive effect size (Cohen’s d > 1.0). Most diagnostic heart sound information resides in lower frequencies, while high-frequency components are often environmental noise or sensor artifacts.
  • SSL as a Robustness Tool: The primary value of SSL in this context is "learning to ignore the noise." SSL models learned features that were more structured and less chaotic than supervised models when visualized via t-SNE.
  • The "Random Flip" Effectiveness: Applying a 50% chance of both reversing and inverting the signal proved highly effective. This forces the network to learn the underlying rhythmic morphology of the heart cycle rather than relying on the specific direction of the wave.
  • Inhibitory Augmentations: Adding Gaussian noise or high-pass filters actually inhibited training, likely because they destroyed the subtle physiological markers the model needed to learn.

Quantitative Results and OOD Resilience

  • Resilience to Distribution Shift: When evaluated on OOD data (unseen datasets), fully-supervised models saw their performance drop by as much as 32%. In contrast, SSL models typically lost only 10% and, in some instances, actually outperformed the supervised baseline on new data.
  • Pascal Dataset Breakthrough: On the Pascal dataset—notoriously difficult due to its small size—the SSL approach outperformed the supervised baseline by 7.6% in binary classification.
  • Effect Size Ranking: The study provides a ranked "cheat sheet" of augmentations by Cohen’s d, prioritizing Random Flip and Low-Pass filters while advising against High-Pass filters and Gaussian Noise.

Conclusion and Future Directions

This research provides a rigorous empirical roadmap for using SSL to classify heart sounds. While SSL is computationally expensive, requiring large batch sizes and significant unlabeled data, its ability to build robust models that generalize across clinical environments is a major step forward for medical AI.

Future work will focus on refining audio-specific augmentations like pitch shifting, investigating if these policies hold for other 1D biosignals (such as ECG or EMG), and testing the transferability of heart-sound pre-trained backbones to entirely different medical tasks.

Summary updated

Abstract

Despite recent advancements in deep learning, its application in real-world medical settings, such as phonocardiogram (PCG) classification, remains limited. A significant barrier is the lack of high-quality annotated datasets, which hampers the development of robust, generalizable models that can perform well on newly collected, out-of-distribution (OOD) data. Self-Supervised Learning (SSL) contrastive learning, has shown promise in mitigating the issue of data scarcity by using unlabeled data to enhance model robustness. Even though SSL methods have been proposed and researched in other domains, works focusing on the impact of data augmentations on model robustness for PCG classification are limited. In particular, while augmentations are a key component in SSL, selecting the most suitable policy during training is highly challenging. Improper augmentations can lead to substantial performance degradation and even hinder a network's ability to learn meaningful representations. Addressing this gap, our research aims to explore and evaluate a wide range of audio-based augmentations and uncover combinations that enhance SSL model performance in PCG classification. We conduct a comprehensive comparative analysis across multiple datasets, assessing the impact of various augmentations on model performance. Our findings reveal that depending on the training distribution, augmentation choice significantly influences model robustness, with fully-supervised models experiencing up to a 32\% drop in effectiveness when evaluated on unseen data, while SSL models demonstrate greater resilience, losing only 10\% or even improving in some cases. This study also highlights the most promising and appropriate augmentations for PCG signal processing, by calculating their effect size on training. These insights equip researchers with valuable guidelines for developing reliable models in PCG signal processing.
BibTeX
@inproceedings{ballas2024which,
  author = {Aristotelis Ballas and Vasileios Papapanagiotou and Christos Diou},
  title = {{Which Augmentation Should I Use? An Empirical Investigation of Augmentations for Self-Supervised Phonocardiogram Representation Learning}},
  booktitle = {IEEE Access},
  year = {2024},
  url = {https://openreview.net/forum?id=1UUZ1cIrJ2},
}
Figure