[Paper Review] Biosignal Generation and Latent Variable Analysis with Recurrent Generative Adversarial Networks

Shota Harada, Hideaki Hayashi+1 more|arXiv (Cornell University)|May 17, 2019

Time Series Analysis and Forecasting34 references4 citations

TL;DR

This paper proposes a conditional recurrent GAN framework for biosignal generation that enables controllable, high-fidelity synthesis of time-series biosignals (e.g., ECG, EEG) using class labels as conditioning inputs. By applying canonical correlation analysis (CCA) to latent variables, the method reveals and exploits the input-output relationship, allowing precise control over generated signal characteristics such as amplitude and frequency, outperforming prior GAN-based approaches in both data quality and controllability without requiring domain-specific knowledge.

ABSTRACT

The effectiveness of biosignal generation and data augmentation with biosignal generative models based on generative adversarial networks (GANs), which are a type of deep learning technique, was demonstrated in our previous paper. GAN-based generative models only learn the projection between a random distribution as input data and the distribution of training data.Therefore, the relationship between input and generated data is unclear, and the characteristics of the data generated from this model cannot be controlled. This study proposes a method for generating time-series data based on GANs and explores their ability to generate biosignals with certain classes and characteristics. Moreover, in the proposed method, latent variables are analyzed using canonical correlation analysis (CCA) to represent the relationship between input and generated data as canonical loadings. Using these loadings, we can control the characteristics of the data generated by the proposed method. The influence of class labels on generated data is analyzed by feeding the data interpolated between two class labels into the generator of the proposed GANs. The CCA of the latent variables is shown to be an effective method of controlling the generated data characteristics. We are able to model the distribution of the time-series data without requiring domain-dependent knowledge using the proposed method. Furthermore, it is possible to control the characteristics of these data by analyzing the model trained using the proposed method. To the best of our knowledge, this work is the first to generate biosignals using GANs while controlling the characteristics of the generated data.

Motivation & Objective

• To develop a single-model, conditional GAN framework that generates multiple classes of biosignals without requiring domain-specific knowledge.
• To address the lack of interpretability and controllability in standard GANs, where input–output relationships are opaque.
• To enable quantitative evaluation of generated biosignals and demonstrate control over specific signal characteristics such as amplitude and frequency.
• To analyze the latent space using CCA to map input conditions (e.g., class labels) to meaningful changes in generated signal features.
• To reduce training cost and model complexity by replacing multiple independent models per class with a single conditional model.

Proposed method

• A conditional GAN architecture is designed with long short-term memory (LSTM) units in both generator and discriminator networks to model temporal dependencies in biosignals.
• Class labels are used as conditioning inputs to both the generator and discriminator, enabling conditional generation of multiple biosignal classes from a single model.
• Canonical correlation analysis (CCA) is applied to latent variables to uncover the statistical relationship between input conditions and generated signal characteristics.
• The CCA results are used to generate controlled input vectors that induce desired changes in output signal features, such as mean frequency or maximum amplitude.
• The model is trained end-to-end using adversarial loss and reconstruction loss, with no need for domain-specific signal priors.
• Input interpolation between class labels is used to analyze the smoothness and continuity of generated signal transitions.

Experimental results

Research questions

• RQ1 Can a single conditional GAN model generate diverse biosignal classes (e.g., ECG, EEG) with high fidelity while reducing model complexity compared to training separate models per class?
• RQ2 How can the latent space of a GAN be analyzed to reveal the relationship between input conditions (e.g., class labels) and generated signal characteristics?
• RQ3 To what extent can the characteristics of generated biosignals be controlled using CCA-derived canonical loadings without explicit supervision on those features?
• RQ4 Does the proposed method achieve comparable or better generation quality than prior GAN-based biosignal generation methods, especially in terms of quantitative similarity to real data?
• RQ5 Can the model generate realistic signal transitions by interpolating between class labels in the latent space, indicating smooth and meaningful interpolation?

Key findings

• The proposed method achieved high-quality biosignal generation with DTW-based similarity scores between real and generated data that were comparable to or better than those of prior methods, including the authors’ own previous independent-class models.
• The number of model parameters was reduced to 1/C of the previous method (where C is the number of classes), significantly lowering training cost and complexity.
• CCA analysis revealed strong canonical correlations between input latent variables and key signal features such as maximum amplitude and mean frequency, confirming a meaningful input–output relationship.
• By manipulating input vectors based on CCA loadings, the model successfully generated signals with intended changes in target features—e.g., increased mean frequency or reduced amplitude—demonstrating effective controllability.
• Interpolation between class labels produced smooth, physiologically plausible transitions in generated biosignals, indicating the model captures meaningful latent structure.
• The method successfully controlled signal characteristics not explicitly used during training (e.g., frequency and amplitude) through latent space analysis, proving the utility of CCA for interpretability and control.