Abstract
High-dimensional time series data, such as multi-lead electrocardiograms (ECG), present significant computational challenges for processing and analysis. Traditional dimensionality reduction techniques, including Wavelet transforms and Functional PCA, often rely on fixed mathematical bases that fail to capture complex temporal dependencies. This paper evaluates the use of Convolutional Variational Autoencoders (CNN-VAEs) to learn data-driven, compressed representations of ECG signals. Our results demonstrate that CNN-VAEs significantly outperform traditional methods in reconstruction accuracy, particularly at high compression rates, while providing robustness to noise and enhanced utility for anomaly detection.
Introduction
The "curse of dimensionality" is a major hurdle in analyzing high-dimensional time series like ECGs, where a standard 12-lead recording can contain upwards of 12,000 data points. Efficient dimensionality reduction is crucial for downstream tasks such as clustering, classification, and real-time monitoring. Traditional methods like PCA often ignore temporal dependencies, while Wavelet transforms use fixed bases that are not optimized for specific signal morphologies. Functional PCA (FPCA) further struggles with the lack of synchronization inherent in heartbeat datasets. This study proposes VAEs as a superior alternative capable of learning tailored filters for cardiac signal components.
Methods
The researchers implemented a symmetrical 3-layer deep Convolutional Variational Autoencoder (CNN-VAE). The model maps input signals to a lower-dimensional latent space defined by probability distributions. To enhance generalization, data augmentation techniques including random cropping and artificial rhythm changes (up-sampling/down-sampling) were employed. The VAE's performance was benchmarked against two wavelet-based approaches: a standard "Global" wavelet decomposition and an "Oracle" approach, which represents a theoretical upper bound by selecting the most significant coefficients for each signal. Evaluations were conducted across three large-scale ECG datasets: PTB-XL, the Georgia Dataset, and CPSC2018.
Results
Reconstruction Accuracy and Compression
The CNN-VAE demonstrated a superior ability to reconstruct signals from highly compressed representations. By learning tailored convolutional filters, the VAE effectively identifies key cardiac features like P-waves, QRS complexes, and T-waves.
In multivariate scenarios involving 12-lead recordings, the VAE's advantage became even more pronounced. At a coefficient retention rate of just 0.5%, the VAE maintained a Mean Squared Error (MSE) of approximately 0.022, while the Global Wavelet method's error was more than double at ~0.050. Notably, at compression rates below 6%, the VAE even outperformed the Wavelet Oracle, suggesting its learned representation is fundamentally more efficient for cardiac data than fixed mathematical bases.
Robustness and Anomaly Detection
The model proved resilient to real-world noise. When tested with signals containing 20% added white noise, VAEs trained on clean data successfully recovered the underlying heartbeat manifold. Furthermore, the VAE served as an effective anomaly detector. By training exclusively on "normal" heartbeats, the reconstruction error was used as a metric for identifying irregularities. The VAE achieved 69% accuracy in anomaly detection, significantly surpassing the 55% accuracy of wavelet-based methods.
Conclusion
CNN-VAEs provide a powerful framework for the dimensionality reduction of complex time series. By learning data-driven representations, they achieve higher compression efficiency and better reconstruction accuracy than traditional fixed-basis methods. While the current model shows promise for anomaly detection and noise robustness, future work will focus on improving the interpretability of the latent space through disentanglement and increasing diagnostic accuracy for clinical applications.
Summary updated