DST-GAN: Dempster-Shafer based fusion for multimodal oversampling of healthcare data

Multimodal healthcare data often have small population sizes and are imbalanced both within modalities (class imbalance) and across modalities (dimensionality imbalance), leading to biased fusion where minority classes are underrepresented and high-dimensional modalities dominate. We propose a Generative Adversarial Network (GAN) with a differentiable component based on Dempster-Shafer Theory (DST). Our DST-GAN is a DST-based multimodal generative framework that fuses modalities to form an equitable joint representation and generates high-quality unimodal samples from it. On five real-world datasets, DST-GAN improves unimodal and multimodal fusion classification over four strong baselines, demonstrating more balanced modality contributions as well as better and more reliable downstream performance.