Introduction
Machine learning in healthcare is often bottlenecked by the scarcity and privacy constraints of medical biosignal data, such as ECG and EEG recordings. While Generative Adversarial Networks (GANs) offer a path for data augmentation, applying them to time-series data presents unique challenges. Standard GANs are typically designed for static images and lack the temporal awareness needed for signals. Furthermore, they often act as "black boxes," providing little control over the specific physical characteristics of the generated output.
This research introduces a conditional recurrent GAN capable of generating realistic, multi-class biosignals within a single model. By integrating Canonical Correlation Analysis (CCA) into the latent space analysis, the authors provide a mechanism to intentionally control features like amplitude and frequency without requiring complex, domain-specific mathematical models.
Methodology
The core of the proposed approach is a conditional recurrent GAN designed specifically for sequential data.
Conditional LSTM-GAN Architecture
Unlike standard GANs that might require separate models for different classes (e.g., normal vs. abnormal heartbeats), this architecture is conditional. It accepts a class label $c$ alongside a random noise vector $z$. Both the Generator and Discriminator utilize Long Short-Term Memory (LSTM) networks to capture temporal dependencies.
Training and Latent Space Control
To ensure stable training and prevent mode collapse, the authors employ an "unrolled" training update rule. The most significant technical innovation, however, is the use of Canonical Correlation Analysis (CCA) for latent variable analysis. After training, physical features (such as maximum amplitude or mean frequency) are extracted from the generated signals. CCA is then used to identify linear correlations between the input noise vectors and these output features. This allows researchers to find "canonical loadings"—specific directions in the noise space that correspond to physical traits—enabling predictable manipulation of the generated signals.
Datasets and Evaluation
The model was evaluated using three real-world datasets from the UEA & UCR repository:
- ECG200: 200 single heartbeats (normal vs. myocardial infarction).
- TwoLeadECG: 1,162 heartbeats from two different ECG leads.
- EEG (Epileptic Seizure Recognition): 11,500 brain activity samples (normal vs. seizure).
Key Results and Insights
The proposed method demonstrated high-quality signal generation, with Dynamic Time Warping (DTW) distances comparable to or better than traditional augmentation methods like noise addition or Hidden Markov Models.
Key findings include:
- Controllability: By adjusting the input noise along the canonical loading directions identified via CCA, the authors successfully forced the Generator to produce signals with specific traits, such as higher peaks or faster frequencies.
- Efficiency: Using a single conditional model reduces the parameter count by a factor of $C$ (the number of classes) compared to training independent GANs for each class.
- Smooth Transitions: The model can generate "interpolated" signals by blending class labels, providing a smooth transition between different physical states.
Limitations and Conclusion
While effective, the model has limitations. LSTMs tend to act as low-pass filters, making it difficult to capture very high-frequency components. Additionally, GAN training remains computationally intensive and lacks clear indicators of convergence. Finally, the controllability via CCA is limited to features that vary naturally within the training data. Despite these challenges, this domain-agnostic approach provides a powerful tool for controllable biosignal synthesis, reducing the need for hand-crafted mathematical models in medical data augmentation.
Summary updated