Introduction
Automated heart sound classification using Phonocardiogram (PCG) signals is a vital tool for early detection of cardiovascular diseases. However, the field faces two major hurdles: the high cost of expert labeling and significant class imbalance. In real-world clinical settings, "abnormal" heart sounds are far rarer than "normal" ones, making it difficult for standard supervised models to learn pathological characteristics effectively. This paper explores methods to reduce reliance on massive labeled datasets by utilizing unlabeled data and framing classification as an anomaly detection problem.
Data and the Challenge of Class Imbalance
The study utilized the PhysioNet/CinC 2016 Challenge dataset, a compilation of six sub-datasets (A-F) collected from various body locations and environments. A primary challenge in this dataset is the severe class imbalance.
As illustrated in Figure 1, the number of "normal" samples vastly outweighs "abnormal" ones in most subsets, particularly in dataset E. This imbalance necessitates specialized training strategies like synthetic over-sampling or alternative learning paradigms.
Methodological Framework
The researchers compared three distinct paradigms to address data scarcity:
1. Supervised Learning
The authors tested deep learning models (ResNet, DenseNet, LSTMs, 1D CNNs) and traditional machine learning (SVMs, Random Forests, Gradient Boosting). To mitigate class imbalance, they employed the Synthetic Minority Over-sampling Technique (SMOTE). Interestingly, Gradient Boosting on extracted audio features (MFCCs, Spectral Contrast) outperformed complex deep learning architectures, achieving a Mean Accuracy (MAcc) of 0.885.
2. Semi-supervised Learning (GAN-based)
A core contribution of this work is a Generative Adversarial Network (GAN) where the Discriminator performs a dual role: distinguishing real from fake samples and classifying heart sounds. By learning the distribution of heart sounds from a large pool of unlabeled data, the model regularizes the classification task, consistently outperforming supervised baselines when labeled samples are extremely scarce (e.g., under 20% of the dataset).
3. Unsupervised Learning (Anomaly Detection)
By treating "abnormal" heart sounds as outliers, the authors developed an unsupervised pipeline using a 1D CNN Autoencoder.
As shown in Figure 6, the autoencoder learns a compressed (latent) representation of "normal" heart sounds. Abnormalities are then detected by applying One-Class SVM or Isolation Forest algorithms to either the reconstruction loss or the latent embeddings.
Key Insights and Results
- The Power of Unlabeled Data: The GAN-based approach is superior in low-data regimes, as unlabeled samples help the model map the underlying "manifold" of heart sounds, making the decision boundary more robust.
- Anomaly Detection Effectiveness: The unsupervised One-Class SVM on autoencoder embeddings reached an AUROC of 0.842. This demonstrates that a useful diagnostic tool can be built without ever seeing a labeled abnormal sample during training.
- Latent Embeddings vs. Reconstruction Loss: Experiments revealed that using latent representations (embeddings) from the autoencoder yielded better detection performance than relying solely on reconstruction error.
Conclusion
This research demonstrates that semi-supervised and unsupervised methods are highly effective for heart sound classification, particularly when labeled data is limited. By leveraging the distribution of unlabeled "normal" sounds, these models can achieve performance comparable to or better than traditional supervised methods, offering a scalable path for clinical diagnostic tools in data-restricted environments.
Summary updated