TL;DR
This paper investigates how different data augmentation techniques impact the performance of Self-Supervised Learning (SSL) models for phonocardiogram (PCG) classification. They found that specific augmentations, particularly low-pass filters, significantly improve model robustness and generalization to unseen data, outperforming traditional fully-supervised methods.
What Problem Does This Paper Address?
Deep learning models show great promise for analyzing medical signals like phonocardiograms (PCGs) to detect heart conditions. However, their real-world adoption is hampered by the scarcity of high-quality, labeled datasets. This often leads to models that perform well on the data they were trained on but fail when encountering new, slightly different data (out-of-distribution or OOD data). Self-Supervised Learning (SSL), especially contrastive learning, offers a way to leverage abundant unlabeled data to build more robust models. However, the effectiveness of SSL heavily relies on choosing the right data augmentation strategies, which is a complex and time-consuming process. This paper aims to systematically evaluate various audio augmentations for PCG classification using SSL to identify the most effective ones and provide guidelines for practitioners.
Methods
The researchers employed a contrastive Self-Supervised Learning (SSL) framework, inspired by SimCLR, for PCG representation learning. The core idea is to train a model to recognize different augmented versions of the same PCG signal while distinguishing them from augmented versions of other signals. This pretraining phase uses unlabeled data. After pretraining, the learned feature extractor's weights are frozen, and a classification head is added and fine-tuned on labeled data for specific downstream tasks (e.g., normal vs. abnormal heart sound classification).
They systematically evaluated a range of audio augmentations, including:
- Noise: Adding zero-mean Gaussian noise.
- Cut-off Filters: Applying high-pass or low-pass filters.
- Scale: Randomly scaling the signal amplitude.
- Reverse: Reversing the signal sequence.
- Inversion: Multiplying the signal by -1.
- Random Flip: A combination of reverse and inversion with a certain probability.
They also explored combinations of these augmentations (1vs1, 1vs2, 2vs2). The effectiveness of these augmentations was assessed by measuring their impact on both in-distribution (data from the same distribution as training) and out-of-distribution (data from different distributions) performance.
Datasets
The study utilized five distinct public PCG datasets:
- Ephnogram: Unlabeled, multi-channel ECG and PCG recordings.
- FPCGDB (Fetal PCG Database): Unlabeled fetal PCG signals.
- Pascal: Labeled heart sound classification data.
- PhysioNet2016: Labeled data for binary classification of normal/abnormal heart sounds.
- PhysioNet2022: Labeled data for heart murmur detection.
The data was preprocessed by discarding the first and last 2 seconds of recordings, splitting into 5-second overlapping windows, and resampling to 2 kHz to homogenize them.
Key Insights and Ideas
- Augmentation is Crucial for SSL: The choice of augmentations significantly impacts the ability of SSL models to learn robust and generalizable representations. Poor choices can even degrade performance.
- Low-Pass Filters Shine: Low-pass filters consistently emerged as the most impactful augmentation for improving PCG classification robustness, especially in out-of-distribution scenarios. This is attributed to their ability to remove high-frequency noise while preserving essential low-frequency signal components.
- Random Flip is Powerful: Applying random flips (combining reversal and inversion) with a probability of 0.5 showed a strong positive effect on training.
- Noise Can Be Detrimental: Adding artificial noise, especially to already noisy datasets, can be counterproductive and lead to degraded performance.
- SSL Offers Superior OOD Robustness: SSL-trained models demonstrated significantly better generalization to unseen data compared to fully-supervised models, with a much smaller drop in performance.
Key Results
- Fully-supervised models experienced up to a 32% drop in effectiveness when evaluated on OOD data.
- SSL models, in contrast, showed a drop of only up to 10% or even improved in some cases on OOD data.
- The study identified specific augmentation combinations that led to the best performing models, with low-pass filters, random flip, and uniform noise being frequently used in top-performing models.
- Cohen's d analysis revealed that low-pass filters had the largest positive effect size (d > 1.0) on robust SSL PCG learning.
- T-SNE visualizations showed that SSL-trained models produced more structured and clustered feature representations compared to baseline fully-supervised models, indicating better generalization.
Limitations and Future Work
- Data Corpus Size: SSL pretraining requires a large unlabeled data corpus to maximize representation learning capabilities.
- Computational Constraints: The contrastive loss function often relies on large batch sizes, which can be computationally demanding.
The authors plan to:
- Refine augmentation techniques by exploring new functions and parameters.
- Extend the approach to transfer learning tasks.
- Investigate optimal augmentation policies for other 1D biosignals like ECG, EMG, and EEG.
Summary updated