pDOC-STANet: A spatiotemporal attention network for automatic assessment of prolonged disorders of consciousness using fNIRS

Abstract This study addresses the difficulty and subjectivity of assessing consciousness levels in patients with prolonged disorders of consciousness (pDOC) by proposing an automated classification method based on functional near-infrared spectroscopy (fNIRS). A spatiotemporal attention-based feature-fusion network, termed pDOC-STANet, was developed. The temporal branch extracts multiscale features using an improved InceptionTime architecture with squeeze-and-excitation modules, whereas the image branch represents each trial as channel-wise Gramian Angular Difference Field (GADF) maps and processes the resulting 105 × 330 × 330 tensor with a modified ResNet18. An attention-fusion module adaptively integrates the two representations. The model was evaluated using stratified, subject-independent five-fold cross-validation, with all trials from each participant kept in the same fold. pDOC-STANet achieved an accuracy of 88.72%, outperforming traditional machine-learning methods and single-branch deep-learning baselines. Ablation experiments confirmed the contributions of GADF encoding, dual-branch modeling, and attention-based fusion. LIME and SHAP analyses further indicated that the model relied on discriminative channels in bilateral frontotemporal and prefrontal regions, with HbR- and HbT-related features making substantial contributions. These results suggest that pDOC-STANet is a promising framework for fNIRS-based consciousness assessment, while external multi-center validation remains necessary before broad clinical deployment.