Decoding functional changes in the brain following ischemic stroke: a multimodal feature approach integrating fNIRS...

Ischemic stroke (IS) induces significant structural and functional remodelling of brain networks. While functional near-infrared spectroscopy (fNIRS) offers high temporal resolution for monitoring cerebral hemodynamics, its potential for capturing complex dysfunction patterns remains underutilized. This study investigated task-evoked and resting-state fNIRS responses to decode neurovascular deficits in IS patients compared to normal controls (NCs).

A total of 160 participants (115 IS patients and 45 NCs) were enrolled and underwent fNIRS recording across three paradigms: resting state, verbal fluency task (VFT), and right-hand gripping task (RHGT). Hemodynamic features including oxyhemoglobin (HbO), deoxy-hemoglobin (HbR), and total hemoglobin (HbT) were extracted for functional connectivity analysis (resting state) and cortical activation assessment (task-based). Three traditional machine learning classifiers (Gaussian Naïve Bayes (GNB), Support Vector Machine (SVM), and Logistic Regression (LR)) and a Transformer-based deep learning model were constructed to classify IS patients from NCs. SHAP analysis was performed to identify the most discriminative features. Statistical comparisons between groups were conducted using multiple linear regression models, adjusting for age, sex, brain lesions, and bilateral hand functional status.

During resting-state recording, IS patients exhibited widespread reductions in functional connectivity within sensorimotor regions, particularly involving bilateral primary motor cortex (M1), premotor and supplementary motor cortex (PMC), and primary somatosensory cortex (PSC), with the most pronounced alterations observed in HbO- and HbT-derived signals (p < 0.05). During the VFT, IS patients showed significant hemodynamic abnormalities in the left dorsolateral prefrontal cortex (DLPFC) and the right Brodmann Area 45 (BA45), characterized by elevated HbR concentrations (p < 0.05). Additionally, a significant increase in HbT was detected at channel 16, predominantly localized in the left frontopolar area (FPA) (p < 0.05). During theRHGT, IS patients exhibited a significant decrease in HbO at channel 26, predominantly localized in the left DLPFC (p < 0.05), along with non-significant downward trends in HbR and HbT (p > 0.05). The Transformer model consistently outperformed all conventional classifiers across three paradigms, achieving peak performance with raw gripping-task data (AUC = 0.86, accuracy = 86%, sensitivity = 100%, specificity = 78%). SHAP analysis identified sensorimotor and prefrontal regions as the most discriminative features for IS classification.

This study revealed distinct hemodynamic abnormalities in IS patients across resting-state, VFT, and RHGT paradigms. The Transformer model consistently outperformed traditional classifiers, with SHAP analysis confirming sensorimotor and prefrontal regions as key discriminative features. These findings support fNIRS combined with deep learning as a promising approach for non-invasive post-stroke functional assessment and rehabilitation planning.