Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Conditional Generative Data Augmentation for Clinical Audio Datasets

Matthias Seibold, Armando Hoch, Mazda Farshad, Nassir Navab, Philipp Fürnstahl

TL;DR

This paper addresses the data bottleneck in clinical audio by introducing a conditional Wasserstein GAN with Gradient Penalty applied to log-mel spectrograms (cWGAN-GP) to generate realistic, class-conditioned samples. The method is evaluated on a novel THA surgical audio dataset and compared against classical augmentations using a ResNet-18 classifier in a 5-fold cross-validation setup. Results show that cWGAN-GP augmentation, especially when doubling the sample count, yields the highest mean Macro F1-score ($95.60\% \pm 1.26\%$), outperforming noise addition, pitch shifts, time stretching, and SpecAugment, with a $+1.70\%$ improvement over no augmentation. The study demonstrates the potential of in-distribution synthetic data to improve robustness for safety-critical medical AI and discusses practical considerations and future extensions to broader datasets and augmentation strategies.

Abstract

In this work, we propose a novel data augmentation method for clinical audio datasets based on a conditional Wasserstein Generative Adversarial Network with Gradient Penalty (cWGAN-GP), operating on log-mel spectrograms. To validate our method, we created a clinical audio dataset which was recorded in a real-world operating room during Total Hip Arthroplasty (THA) procedures and contains typical sounds which resemble the different phases of the intervention. We demonstrate the capability of the proposed method to generate realistic class-conditioned samples from the dataset distribution and show that training with the generated augmented samples outperforms classical audio augmentation methods in terms of classification performance. The performance was evaluated using a ResNet-18 classifier which shows a mean Macro F1-score improvement of 1.70% in a 5-fold cross validation experiment using the proposed augmentation method. Because clinical data is often expensive to acquire, the development of realistic and high-quality data augmentation methods is crucial to improve the robustness and generalization capabilities of learning-based algorithms which is especially important for safety-critical medical applications. Therefore, the proposed data augmentation method is an important step towards improving the data bottleneck for clinical audio-based machine learning systems.

Conditional Generative Data Augmentation for Clinical Audio Datasets

TL;DR

This paper addresses the data bottleneck in clinical audio by introducing a conditional Wasserstein GAN with Gradient Penalty applied to log-mel spectrograms (cWGAN-GP) to generate realistic, class-conditioned samples. The method is evaluated on a novel THA surgical audio dataset and compared against classical augmentations using a ResNet-18 classifier in a 5-fold cross-validation setup. Results show that cWGAN-GP augmentation, especially when doubling the sample count, yields the highest mean Macro F1-score (), outperforming noise addition, pitch shifts, time stretching, and SpecAugment, with a improvement over no augmentation. The study demonstrates the potential of in-distribution synthetic data to improve robustness for safety-critical medical AI and discusses practical considerations and future extensions to broader datasets and augmentation strategies.

Abstract

In this work, we propose a novel data augmentation method for clinical audio datasets based on a conditional Wasserstein Generative Adversarial Network with Gradient Penalty (cWGAN-GP), operating on log-mel spectrograms. To validate our method, we created a clinical audio dataset which was recorded in a real-world operating room during Total Hip Arthroplasty (THA) procedures and contains typical sounds which resemble the different phases of the intervention. We demonstrate the capability of the proposed method to generate realistic class-conditioned samples from the dataset distribution and show that training with the generated augmented samples outperforms classical audio augmentation methods in terms of classification performance. The performance was evaluated using a ResNet-18 classifier which shows a mean Macro F1-score improvement of 1.70% in a 5-fold cross validation experiment using the proposed augmentation method. Because clinical data is often expensive to acquire, the development of realistic and high-quality data augmentation methods is crucial to improve the robustness and generalization capabilities of learning-based algorithms which is especially important for safety-critical medical applications. Therefore, the proposed data augmentation method is an important step towards improving the data bottleneck for clinical audio-based machine learning systems.
Paper Structure (9 sections, 3 equations, 3 figures, 1 table)

This paper contains 9 sections, 3 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Materials and Method
  3. Novel Surgical Audio Dataset
  4. Data Preprocessing and Baseline Augmentations
  5. Conditional Generative Data Augmentation Method
  6. Classification Model
  7. Results and Evaluation
  8. Discussion
  9. Conclusion

Figures (3)

  • Figure 1: The classes of the novel clinical dataset resemble the phases of a THA procedure. Occurrences with drawn through lines indicate intensive usage of the respective surgical action, dashed lines correspond to sporadic usage.
  • Figure 2: The architecture of the proposed model including output sizes of each layer. The input for the generator is a noise vector of size 1x128 and a class condition. The generator outputs a spectrogram which is fed to the discriminator together with the class condition. The discriminator (critic) outputs a scalar realness score.
  • Figure 3: The top row shows log-mel spectrograms of random samples for each class present in the acquired dataset, the bottom row shows log-mel spectrograms generated by the proposed model for each class, respectively.