Post-Traumatic Stress Disorder Recognition Using the fNIRS Dual-Attention Network

Post-traumatic stress disorder (PTSD) is a prevalent psychological disorder that can impair daily functioning and mental health. Current clinical assessment mainly relies on structured interviews and symptom scales, which may introduce subjectivity and inter-rater variability. Functional near-infrared spectroscopy (fNIRS) provides a noninvasive approach for measuring task-evoked hemodynamic responses and may offer objective physiological information for auxiliary PTSD recognition.

This study proposes an end-to-end deep learning framework, the fNIRS Dual Attention Network (FDANet), for PTSD recognition. Ten PTSD patients and ten healthy controls (HC) were recruited, and prefrontal hemodynamic signals, including oxyhemoglobin (HbO), deoxyhemoglobin (HbR), and total hemoglobin (HbT), were recorded during fear and neutral facial-expression stimuli. The three hemodynamic signals were concatenated along the channel dimension to form the model input. FDANet combines Transformer-based temporal modeling, one-dimensional temporal-channel attention, and convolutional feature extraction to learn discriminative representations from multichannel fNIRS signals.

The model was evaluated using subject-wise nested cross-validation with five outer folds. At the trial level, FDANet achieved an accuracy of 79.4%. At the subject level, FDANet achieved an accuracy of 95.0%.

Compared with the evaluated traditional machine-learning and deep-learning baseline methods, FDANet achieved higher classification accuracy at both trial and subject levels.

These findings indicate that task-evoked prefrontal fNIRS signals contain discriminative information for PTSD recognition. Given the portability and noninvasive measurement characteristics of fNIRS, FDANet may support the future development of objective auxiliary screening tools for PTSD, although further validation with larger and independent samples is required.