Robust Detection of Brain Stimulation Artifacts in iEEG Using Autoencoder-Generated Signals and ResNet Classification
Background: Intracranial EEG (iEEG) is crucial for understanding brain function, but stimulation-induced noise complicates data interpretation. Traditional artifact detection methods require manual user input or struggle with noise variability, especially with limited labeled data. Objective: We developed a supervised method to automatically detect stimulation-induced noise in human iEEG recordings using synthetic data generated by Variational Autoencoders (VAEs) to train a ResNet-18 classifier. Methods: Multi-lead iEEG data were collected, preprocessed, and used to train VAEs for generating synthetic clean and noisy signals. The ResNet-18 model was trained on images of spectra generated from these synthetic signals and validated on real iEEG data from five participants. Results: The classifier, trained exclusively on synthetic data, demonstrated high accuracy, precision, and recall when applied to real iEEG recordings, with AUC values greater than 0.99 across all participants. Conclusion: We present a novel approach to effectively detect stimulation-induced noise in iEEG, offering a robust solution for improving data interpretation in scenarios with limited labeled data. Additionally, the pre-trained ResNet-18 model is available for the community to use, facilitating further research and application in similar datasets.