Introduction
Cardiovascular diseases (CVDs) remain a leading cause of global mortality. While phonocardiogram (PCG) analysis—the clinical practice of listening to heart sounds—is a powerful diagnostic tool, modern machine learning approaches often rely on supervised learning. Such methods require massive datasets of both normal and abnormal heart sounds, which are difficult to collect, expensive to label professionally, and frequently biased toward specific pathologies.
This paper proposes an unsupervised system that learns the fundamental characteristics of healthy heart sounds. By training exclusively on "normal" samples, the model identifies any significant deviation as a potential abnormality. This approach eliminates the need for labeled pathological data during training and enhances robustness against rare or previously unseen heart conditions.
Methodology
The proposed architecture utilizes a $eta$-Variational Auto-Encoder ($eta$-VAE). A standard VAE compresses input data into a lower-dimensional latent space and subsequently reconstructs the original input. The $eta$ parameter is critical, as it controls the trade-off between reconstruction accuracy and the degree to which the latent space adheres to a Gaussian distribution.
The processing pipeline involves:
- Pre-processing: PCG signals are transformed into Mel Spectrograms, normalized, and segmented into 3.07-second "super-frames" to capture multiple heartbeats.
- Architecture: A symmetric encoder-decoder setup featuring four fully-connected hidden layers (32, 32, 16, and 16 neurons) with a 16-dimensional latent space.
- Anomaly Scoring: The model is trained solely on normal samples. During inference, the Reconstruction Loss (Mean Squared Error) serves as the anomaly score. High reconstruction error indicates a sound that the model cannot accurately reproduce, flagging it as a likely anomaly.
Experimental Setup
The study utilizes the PhysioNet / CinC Heart Sound Dataset, comprising six subsets ('a' through 'f') from various sources and recording devices. This diversity facilitates testing across different recording environments, simulating real-world multi-domain challenges.
Results and Discussion
The experimental results highlight a phenomenon termed the "β Paradox." In typical generative modeling, β > 1 is preferred to encourage disentangled features. However, for PCG anomaly detection, the authors found that lower values (β < 1, specifically 0.01 or 0) yielded superior performance.
Key insights include:
- Reconstruction Priority: Low β values prioritize reconstruction accuracy over latent space organization. A "lightweight" KL divergence regularization helps the model generalize to new normal data without sacrificing the sensitivity required to detect anomalies.
- Superiority of Reconstruction Scores: For PCG analysis, anomaly scores derived from reconstruction loss outperformed those based on latent vectors. Increasing β increases the correlation between reconstruction loss and latent distribution, which paradoxically degrades detection performance in this context.
- Performance Metrics: The optimized model achieved a peak AUC of 0.91, making it competitive with state-of-the-art supervised systems. The β=0.01 configuration consistently outperformed standard VAEs (β=1) and basic Auto-Encoders across most subsets.
Conclusion
This research demonstrates the efficacy of unsupervised β-VAEs for heart abnormality detection. While the model showed strong cross-domain robustness, performance varied across subsets (e.g., subset 'b'), suggesting that specific noise profiles may require further architectural refinement. Future work will explore more complex architectures, such as Convolutional or Recurrent Neural Networks, to better capture the temporal dynamics of heart sounds within the unsupervised β-VAE framework.
Summary updated