Machine learning classification of fNIRS signals in patients with major depressive disorder and healthy controls.

Objective: To determine whether combining resting-state and the verbal fluency task (VFT) prefrontal fNIRS features with machine learning can distinguish patients with MDD from healthy controls (HCs), and to clarify the relative contributions of functional connectivity (FC) versus amplitude features to classification. Methods: We enrolled 98 patients with MDD and 51 HCs, matched on age, sex, education, and handedness. Participants completed a resting-state fNIRS recording and a VFT using a multichannel prefrontal-temporal montage. After preprocessing, oxyhemoglobin (HbO)-based amplitude features and Pearson correlation-based FC features were extracted from two conditions. Five classifiers-logistic regression, k-nearest neighbors, support vector machine (SVM), partial least squares, and linear discriminant analysis-were evaluated using nested stratified cross-validation. Model performance was quantified using out-of-fold (OOF) predicted probabilities from the outer loop and summarized by accuracy, sensitivity, specificity, and area under the ROC curve (AUC). The 95% confidence intervals (95% CIs) for AUC were estimated using DeLong's method. The classification threshold was determined using the Youden index on the OOF ROC curve. SHAP analysis was then applied to the best-performing model to quantify the contribution of each feature to the predicted probability of MDD. Results: Demographic variables did not differ significantly between groups. Across all classifiers, the SVM achieved the best discrimination between MDD and HCs, with OOF AUC of 0.867 (95% CI: 0.805-0.928) and an accuracy of 0.839 (95% CI: 0.770-0.894). At the Youden-optimal threshold (0.482), the SVM model yielded a sensitivity of 0.918 (95% CI: 0.845-0.964) and a specificity of 0.686 (95% CI: 0.541-0.809). SHAP analysis indicated that the top-ranked predictors were predominantly FC features from both resting-state and VFT conditions, particularly IPFC-TPC connectivity at rest and FPC-TPC and IPFC-DLPFC connectivity during VFT, whereas regional HbO amplitude measures contributed less. Conclusion: The present study suggests that network-level dysconnectivity, which fNIRS captures, plays a central role in distinguishing MDD from HCs, rather than isolated regional hypoactivation. With its high sensitivity, low cost, and SHAP-based interpretability, the SVM model provides an interpretable framework for investigating fNIRS-based neurophysiological patterns associated with MDD and may serve as a potential adjunctive approach for future screening applications after external validation. Multi-center external validation is required before routine implementation.