DE-CGAN: Boosting rTMS Treatment Prediction with Diversity Enhancing Conditional Generative Adversarial Networks

TL;DR

This paper tackles the problem of unstable rTMS outcome prediction due to underrepresented fMRI connectivity patterns by introducing DE-CGAN, a diversity-enhancing conditional GAN that oversamples misclassified regions with conditioned synthetic examples. The method builds a diversity-enhanced training set by identifying frequently mislabelled real samples and generating labeled synthetic variants to balance and expand the feature space. Empirical results using a Train Synthetic Test Real framework show that a modest synthetic proportion (notably α ≈ 0.10) improves held-out performance beyond traditional augmentation and CGAN baselines, supporting the value of synthetic patients for robust psychiatry AI. The findings have practical implications for improving generalisation in small, privacy-conscious datasets and for enabling deeper exploration of connectivity-behavior relationships in rTMS response.

Abstract

Repetitive Transcranial Magnetic Stimulation (rTMS) is a well-supported, evidence-based treatment for depression. However, patterns of response to this treatment are inconsistent. Emerging evidence suggests that artificial intelligence can predict rTMS treatment outcomes for most patients using fMRI connectivity features. While these models can reliably predict treatment outcomes for many patients for some underrepresented fMRI connectivity measures DNN models are unable to reliably predict treatment outcomes. As such we propose a novel method, Diversity Enhancing Conditional General Adversarial Network (DE-CGAN) for oversampling these underrepresented examples. DE-CGAN creates synthetic examples in difficult-to-classify regions by first identifying these data points and then creating conditioned synthetic examples to enhance data diversity. Through empirical experiments we show that a classification model trained using a diversity enhanced training set outperforms traditional data augmentation techniques and existing benchmark results. This work shows that increasing the diversity of a training dataset can improve classification model performance. Furthermore, this work provides evidence for the utility of synthetic patients providing larger more robust datasets for both AI researchers and psychiatrists to explore variable relationships.

Figures (3)

• Figure 1: DE-CGAN conceptual model (Created with BioRender.com)
• Figure 2: Box plot showing the distribution of accuracies obtained using various algorithms.
• Figure 3: Bar chart showing the frequency of optimal solutions by algorithm.