FNIRNet: A Spatiotemporal Statistical Feature-Based Convolutional Neural Network for Low-Channel fNIRS Brain Function Classification

Abstract

We propose functional near-infrared spectroscopy network (FNIRNet), a lightweight deep learning framework that combines spatiotemporal representations with a statistical feature module to directly process low-channel functional near-infrared spectroscopy signals for brain function classification. A clinically guided and SHAP-validated channel selection strategy identifies 8 critical channels, reducing hardware dependence while preserving discriminative information. On a depression dataset, FNIRNet achieves 92% accuracy and an F1-score of 0.95, outperforming machine-learning baselines including support vector machine, k-nearest neighbors, random forest, and deep models such as electroencephalography network and convolutional neural network–long short-term memory under low-channel input. With 8 channels, FNIRNet also improves computational efficiency and supports practical deployment. The model generalizes to a public stroke dataset, indicating robustness across probe configurations and hemoglobin signals. FNIRNet offers a low-cost, high-performance, and deployable solution for large-scale mental-health screening in nonspecialist settings, and its performance is validated using the area under the receiver operating characteristic curve metric.