An Efficient Convolutional Neural Network Model for Classifying Positive Affective States from EEG Data Considering...

The recognition of affective states from EEG recordings considering the sleep condition can be an important supporting tool in therapy and in the treatment of psychological or psychiatric disorders.

Since sleep deprivation changes EEG characteristics, a two-stage framework is developed for affective state classification. First, a classification between "Normal Sleep" (NS) and "Sleep-Deprived" (SD) is conducted using resting-state EEG data. The data are then fed into one of two newly proposed, lightweight Convolutional Neural Networks (CNNs) to recognize four affective states.

In the first stage, CNN-BiLSTM achieves 95.96% accuracy under window-level splitting and 70.46% under subject-independent 5-fold cross-validation, while in the second stage the proposed models achieve 93.40% (NS) and 90.95% (SD) under window-level splitting, and 38.98% (NS) and 38.97% (SD) under cross-validation. The lower subject-independent accuracies reflect the class imbalance and the clustering of scores near the class boundaries, which make cross-subject four-class generalization challenging.

As existing studies mostly classify momentary emotions without addressing sleep conditions, direct comparison with prior work is difficult; we therefore benchmarked our framework against several established baseline models, including SqueezeNet, EEGNet, ShallowConvNet, Support Vector Machine, and Random Forest.

This study suggests that considering physiological states can improve multi-class classification of affective states from time-varying neurophysiological signals such as EEG. The framework performs strongly under window-level evaluation, whereas its performance under the stricter subject-independent setting is more limited and should be interpreted accordingly. It can nonetheless contribute to more comprehensive research in affective neuroscience.