fNIRS Foundation Model for Few-Shot based fNIRS Classification

Functional near-infrared spectroscopy (fNIRS) is a non-invasive technique with significant potential for applications in brain-computer interfaces (BCIs) including mental health diagnostics and cognitive state monitoring. However, the reliance on large labeled datasets for high-performing classification methods poses a critical challenge, given the time-consuming and resource-intensive nature of fNIRS data collection. To address this, we propose a novel foundation model for fNIRS data based on a self-supervised masked autoencoder framework. The proposed method enables efficient pre-training on unlabeled data, reducing the dependence on labeled datasets while maintaining robust performance for downstream tasks. Experimental results demonstrate that the proposed model achieves performance comparable to supervised learning approaches while requiring only one-third of the labeled training data. It consistently outperforms state-of-the-art self-supervised models in both linear probing and fine-tuning settings. Moreover, ablation studies show that a larger masking size aligns with the low-frequency nature of fNIRS signals, enabling the model to capture broader patterns and further enhance classification accuracy. These findings validate the proposed method as an effective and scalable solution for fNIRS-based classification tasks.