Cross-Time-Window Transfer Benchmark for fNIRS-Based Autism Classification

A recent investigation published on arXiv (2608.07567) tackles the issue of temporal distribution shifts in functional near-infrared spectroscopy (fNIRS) for classifying autism spectrum disorder (ASD). The authors define this challenge as a cross-time-window transfer problem, proposing a method that alters the window length (ranging from 2.5 to 10 seconds) and offset during biological motion trials. They utilize topographic map representations of fNIRS data to evaluate three vision architectures against two zero-shot baselines and eight adaptation techniques through leave-one-subject-out cross-validation involving 124 participants. Notable results reveal that zero-shot cross-window accuracy hovers around chance levels (54–69%), while subject-specific fine-tuning can enhance accuracy to 90–96%. The findings underscore the significance of considering individual variations in hemodynamic delay and neurovascular coupling, which influence optimal observation windows. This research sets a benchmark for future fNIRS-based ASD classification efforts, highlighting the necessity of adaptation strategies for managing temporal shifts. The paper can be accessed on arXiv under the identifier 2608.07567.