Transfer Learning for Subject-Independent Sleep Deprivation Detection from Resting-State EEG

Abstract

Sleep deprivation (SD) impairs cognition and heightens safety risks, yet reliable electroencephalography (EEG)-based detection remains challenging in low-data settings. We evaluated transfer learning with a compact Convolutional Neural Network (CNN) (EEGNetv4) to classify SD versus well-rested wakefulness using an open-source EEG dataset containing eyes-open resting-state data from 71 healthy young adults. EEGNetv4 was initialized with publicly available weights pretrained on an Event-Related Potential (ERP) dataset. Shape-compatible layers were transferred and frozen, with the remaining layers trained on the target data. Baselines comprised EEGNetv4, a bidirectional Long Short-Term Memory (LSTM), and a Transformer model, each trained without pretraining. Five-fold subject-independent cross-validation was used to evaluate model performance. EEGNetv4 with transfer learning achieved the highest mean accuracy (70.79% ± 4.17), outperforming EEGNetv4 trained from scratch (65.75% ± 5.48), the Transformer (63.35% ± 2.78), and the LSTM (61.70% ± 3.20). These findings suggest that leveraging pretrained EEG representations can enhance subject-generalizable SD classification in small-sample contexts, supporting transfer learning as a pragmatic strategy for neurophysiological applications.