Reducing interdataset covariate shift in sleep EEG of traumatic brain injured humans and mice using Transfer Euclidean Alignment

While the analysis of sleep electroencephalography (EEG) offers distinct advantages over other methods in clinical applications, high variability across subjects poses a significant challenge in deploying machine learning (ML) models for real-world classification tasks. In such instances, ML models that exhibit exceptional performance on a dataset may not necessarily demonstrate similar proficiency when applied to a different dataset performing the same task. The scarcity of high-quality biomedical data further exacerbates this challenge, making it difficult to comprehensively evaluate the model’s generalizability. In this paper, we introduce Transfer Euclidean Alignment — a transfer learning (TL) technique designed to tackle the problem of the dearth of human biomedical data in training deep learning (DL) models. We test the robustness of this TL technique across various ML models as well as the EEGNet-based DL model, by evaluating them on different datasets, including human and mouse data, for the binary classification task of detecting individuals with versus without traumatic brain injury (TBI). This work represents the first demonstration of the use of TL in the context of utilizing a mouse model as the source data to improve the performance of the target human dataset. By demonstrating notable improvements with an average increase of 14.42% for intraspecies datasets and 5.46% for interspecies datasets, our findings underscore the importance of TL in boosting the performance of ML and DL models trained on diverse datasets.