Generative AI Enables the Detection of Autism Using EEG Signals

In disease detection, generative models for data augmentation offer a potential solution to the challenges posed by limited high-quality electroencephalogram (EEG) data. The study proposes a temporal-spatial feature-aware denoising diffusion probabilistic model (DDPM), termed TF-DDPM, as an EEG time-series augmentation framework for autism research. The module for predicting noise is CCA-UNet based on the channel correlation-based attention (CCA) mechanism, which considers the spatial and temporal correlation between channels, and uses depthwise separable convolution instead of traditional convolution, thereby suppressing the interference from irrelevant channels. Visualization and binary classification results on synthetic signals indicate that proposed method generates higher quality synthetic data compared to Generative Adversarial Networks (GAN) and DDPM.