Deep learning-based classification of cortical and subcortical ischemic stroke infarction using fNIRS signals

Distinguishing cortical from subcortical ischemic stroke may help guide management. We developed DBT-Net, which combines a DHR residual module, BiLSTM, and Transformer, to classify stroke subtypes from fNIRS signals. The model was evaluated by patient-wise stratified five-fold cross-validation in 175 participants under verbal fluency, right-hand grasp, and resting-state conditions. Accuracy was 0.886 ± 0.055, 0.913 ± 0.062, and 0.878 ± 0.046, with AUCs of 0.920 ± 0.072, 0.912 ± 0.069, and 0.900 ± 0.034, respectively. SHAP analysis highlighted cortical channels contributing most to the predictions.