Time-frequency deep metric learning of resting-state fNIRS signals for staging Alzheimer's disease

This study proposes a deep metric learning framework designed to embed resting-state functional near-infrared spectroscopy (fNIRS) signals into a feature space using a continuous wavelet transform layer in the time-frequency domain for classifying stages of Alzheimer's disease (AD). Early AD pathology, particularly working memory impairment, is closely linked to changes in frontal-lobe networks. Classifying AD patients based on the low-frequency components (0.008-0.15 Hz) of resting-state fNIRS signals presents a significant research challenge. We collected resting-state fNIRS data from individuals with AD, mild cognitive impairment (MCI), and healthy controls (HCs). Due to the limited size of the dataset and the challenges in quantifying individual signal characteristics, conventional data-intensive deep learning models show constrained performance and generalizability. To address these limitations, our proposed architecture generates robust signal embeddings and evaluates inter-sample similarity using a learned distance metric. Experimental results demonstrate high classification accuracy between AD and HCs in the time- frequency domain. Additionally, our findings indicate that i) frequency-domain metric learning effectively captures brain-signal complexity, and ii) differences in inter-regional activation play a crucial role in the progression of neurodegeneration.