AnesthesiaGAN: Frequency-Aware EEG Augmentation for Improved Anesthesia Depth Prediction

The prediction of anesthesia depth is crucial in clinical anesthesia, as it directly affects the safety of surgery and postoperative rehabilitation quality. Deep learning models trained on electroencephalogram (EEG) have wide applications in predicting anesthesia depth, but there are still difficulties in obtaining anesthesia data and insufficient data volume. Therefore, this article proposes the frequency-domain conditional generative adversarial network AnesthetiaGAN. The generator, conditioned on the bispectral index (BIS), produces EEG signals reflecting specific anesthesia depths. A multilayer convolutional module within the discriminator extracts frequency-domain features, guided by adversarial training with a gradient penalty. Signal similarity is assessed using Wasserstein distance (WD), Kullback–Leibler divergence (KLD), and Fréchet Inception Distance (FID). BIS prediction models are trained on original and AnesthesiaGAN-augmented datasets to compare accuracy. AnesthesiaGAN-generated signals outperform baselines, achieving lower WD (0.072 vs. 0.0904), KLD (0.106 vs. 0.110), and FID (36.95 vs. 38.19) in the time domain, and an 81.8% KLD reduction (15.54 vs. 85.19) in the frequency domain. Augmenting training data with AnesthesiaGAN reduces BIS prediction mean absolute error from 0.083 to 0.055 (33.7% improvement). AnesthesiaGAN preserves time-frequency fidelity, enhancing synthetic EEG authenticity and BIS prediction accuracy. This algorithm can effectively expand anesthesia state data samples, and meet downstream task requirements.