300 Section - Development and Validation of an Interpretable Multitask Model to Predict Motor Function Outcomes and Neural Mechanisms in Stroke Patients: A Retrospective Cohort Study

Objectives : Stroke is an acute cerebrovascular disease caused by an obstruction of blood supply to the brain, characterized by high disability and mortality rates. This study aims to use low-cost functional near-infrared spectroscopy (fNIRS) technology to explore brain function abnormalities in stroke patients from multiple feature dimensions and to investigate their potential applications in disease classification and clinical function prediction.

Design: This study collected fNIRS data from 76 healthy controls (HC) and 343 stroke patients (including 69 cases of periventricular, 79 cases of cortical, and 195 cases of deep infarct), covering the bilateral frontal, parietal, and occipital regions. Based on the preprocessed signals and functional matrices, we extracted three types of brain functional features: ALFF, ReHo, and FCS.

Results: Functional abnormalities in the parietal lobe were the most pronounced among stroke patients between groups, suggesting that this region may be a key node in stroke-related network damage. We further compared the results of three binary classification models (SVM, GLM, and Random Forest) in distinguishing subtypes, with the Random Forest model performing best across all tasks. For the three stroke subtypes, the classification AUCs against healthy controls were: HC vs. Cortical Stroke: AUC = 0.93; HC vs. Deep Infarct: AUC = 0.90; HC vs. Periventricular Stroke: AUC = 0.84. We further used fNIRS features to predict multiple clinical indicators (such as Barthel Index, Brunnstrom stages, etc.), and the results showed that fNIRS features were significantly correlated with patients' daily living abilities (correlation coefficients for multiple indicators all r > 0.95, p < 0.001), demonstrating the significant potential of fNIRS in stroke functional assessment.

Conclusion: The multi-feature fusion method based on fNIRS can not only effectively distinguish strokes and their subtypes, but also provide reliable predictions of clinical functional status, laying an important foundation for the development of a low-cost, bedside-applicable stroke auxiliary diagnostic system.