TinyCNNDeep: Lightweight CNN classifies sleep deprivation and eye states with 83.7% accuracy

Sleep deprivation impairs cognitive function, but jointly identifying sleep condition and eye state from EEG remains underexplored. Researchers (Vo et al.) address this four-class problem with TinyCNNDeep, a lightweight convolutional neural network that fuses residual learning with a Squeeze-and-Excitation (SE) attention module. The pipeline converts short multi-channel EEG segments from five physiologically relevant channels—Fp1, Fp2, O1, Oz, O2—into 224x224 grayscale images via per-channel Z-score normalization, min-max scaling, and center padding. This image-based representation allows 2D convolutions to jointly model inter-channel and temporal structure, while the SE attention emphasizes informative features across channels.

On a 35-subject dataset recorded under normal-sleep and sleep-deprivation sessions, TinyCNNDeep attains a subject-wise mean accuracy of 83.69%, outperforming the strongest baseline (Random Forest with combined time-frequency features, 47.66%) by 36.03 percentage points. Critically, three established EEG architectures—EEGNet, ShallowConvNet, and DeepConvNet—operate near chance on this task. Per-subject analysis reveals inter-subject variability, and confusion-matrix inspection shows that residual misclassifications concentrate between eyes-closed states across both sleep conditions. These results demonstrate that an image-based EEG representation paired with residual feature extraction and channel attention provides an accurate, computationally efficient framework for multiclass sleep-related EEG classification under a minimal electrode configuration.