A cross-task fNIRS framework for multi-class mental health classification using adaptive stacking of temporal and network topological features

Background: Early detection of sub-health mental states in college students is critical for timely intervention but remains challenging due to subtle neural signals. This study aimed to classify distinct mental health states (healthy, sub-healthy, abnormal) by integrating functional near-infrared spectroscopy (fNIRS) signal processing with machine learning framework. Method: A total of 292 participants completed resting-state (Rest), verbal fluency (VFT), and emotional Stroop (PStroop, NStroop) tasks during fNIRS acquisition. Temporal and network topological features were extracted from whole-brain signals. Following preprocessing and feature selection, an adaptive stacking ensemble model using a one-vs-all strategy was developed and optimized via 10-fold nested cross-validation. A multi-layer perceptron ensemble served as the prediction assessment. SHapley Additive exPlanations (SHAP) were applied to enhance model interpretability. Results: In single-task classification, the PStroop task achieved a balanced accuracy of 0.92. For multi-task classification, the Rest & VFT and Rest & PStroop combination yield the best performance, with a balanced accuracy of 0.94. SHAP revealed task-specific potential biomarkers (e.g., prefrontal hyperactivity in sub-healthy states). Limitations: The reliance on single modality fNIRS data and sampling from a single site may limit the generalizability. Future work should validate findings using multimodal and multi-site datasets. Conclusion: This study demonstrates that integration of fNIRS with adaptive machine learning approaches provides a scalable and interpretable framework for mental health classification, enhancing the detection of sub-health states—a key target for preventive mental healthcare.