Abstract
Repetitive Transcranial Magnetic Stimulation (rTMS) is an evidence-based treatment for depression, but patient response is inconsistent, with approximately one-third of patients showing minimal improvement. Predicting outcomes using Deep Neural Networks (DNNs) is often hindered by "data sparsity" and imbalanced datasets in psychiatry. This paper introduces Diversity Enhancing Conditional Generative Adversarial Networks (DE-CGAN), a framework designed to identify "difficult-to-classify" patient profiles and generate targeted synthetic data to fill these gaps. By focusing on diversity enhancement rather than simple class balancing, DE-CGAN achieves a prediction accuracy of 0.9360, significantly outperforming baseline models.
Introduction
While rTMS is a vital tool in treating depression, its efficacy varies widely among patients. Deep learning models trained on fMRI connectivity data offer a path toward personalized medicine, but they frequently struggle with small, imbalanced datasets. Previous research has shown that standard models consistently mislabel patients with specific connectivity patterns, such as those between the Subgenual Anterior Cingulate Cortex and the Occipital Cortex. Traditional oversampling methods like SMOTE or basic GANs often fail because they prioritize balancing class labels (responders vs. non-responders) rather than addressing the underlying sparsity in "hard-to-classify" regions of the feature space.
Methods: The DE-CGAN Framework
The proposed DE-CGAN framework shifts the focus from class balancing to diversity enhancement. The process involves four primary stages:
- Error Identification: A baseline DNN identifies examples that are frequently misclassified.
- Targeted Generation: A Conditional GAN (CGAN), conditioned on class labels, is trained specifically on these "difficult" examples.
- Diversity Boosting: The CGAN generates synthetic "patients" ($D_{synth}$) in the sparse regions where the model previously failed.
- Hybrid Training: A final classifier is trained on a hybrid dataset ($D_{hybrid} = D_{real} \cup D_{synth}$), where $\alpha$ represents the proportion of synthetic data.
To ensure rigorous validation, the study employs the Train Synthetic Test Real (TSTR) framework, ensuring the model is always tested against actual, held-out patient data.
Datasets and Connectivity Features
The study utilized fMRI connectivity features from a dataset of rTMS patients. The features consist of six functional connectivity measures between the Subgenual Anterior Cingulate Cortex and various other brain regions.
The target for the model is a binary classification of treatment response. The performance was benchmarked against state-of-the-art models and standard GAN-based augmentation techniques.
Key Results and Insights
The researchers identified a "sweet spot" for synthetic data integration. While adding synthetic data improves performance, excessive amounts can lead to "mode collapse" or skew the model away from the true population distribution.
- Optimal Performance: At $\alpha = 0.10$ (10% synthetic data), the DE-CGAN model achieved an Accuracy of 0.9360 and an F1-Score of 0.9414.
- Comparative Advantage: This outperformed the baseline real-data model (Accuracy: 0.9227) and standard CGAN augmentation (Accuracy: 0.9134).
- Statistical Significance: A two-tail proportion test confirmed that the DE-CGAN model at 10% synthetic data found optimal solutions significantly more frequently than benchmarks ($p < 0.05$).
Beyond raw metrics, DE-CGAN promotes fairness by oversampling underrepresented fMRI profiles, making the resulting AI more robust for patients who do not fit the "average" clinical profile.
Conclusion and Future Work
DE-CGAN demonstrates that targeting model errors through diversity-enhanced synthetic data is more effective than simple class balancing for psychiatric outcome prediction. However, limitations remain, including the risk of mode collapse and the relatively small size of the underlying real-world dataset. Future work will focus on validating DE-CGAN across larger, multi-site clinical trials and exploring more advanced generative architectures to further improve synthetic data diversity. These tools may eventually allow psychiatrists to explore variable relationships in "synthetic patients" to better inform real-world clinical decisions.
Summary updated