Explainable Artificial Intelligence Approaches for Autism Spectrum Disorder Classification via Hyperspectral Imaging Analysis of Red Blood Cells

Objective biological markers for Autism Spectrum Disorder (ASD) remain an important research goal because clinical assessment is behavior-based and ASD is heterogeneous. This study evaluated whether eight red-blood-cell (RBC) hyperspectral endmember distributions contain participant-level ASDversus-neurotypical classification signal under leakage-controlled validation. The cohort comprised 58 male children (27 ASD and 31 neurotypically developing), each with nine RBC hyperspectral acquisitions (522 acquisitions in total). Six classifiers were evaluated using 58-fold outer leave-one-subject-out cross-validation, with five-fold participant-grouped inner hyperparameter selection. The nine held-out probabilities were averaged to obtain one prediction per participant. XGBoost achieved the highest observed fixed-threshold accuracy (0.828; 95% CI 0.724–0.914), F1-score (0.800; 0.667–0.906), Matthews correlation coefficient (0.657), and lowest Brier score (0.155). Support Vector Machine achieved the highest AUROC (0.879; 0.784–0.952), average precision (AP; 0.891), and specificity (0.968), but lower sensitivity (0.630). Exact paired McNemar tests found no significant participant-level accuracy difference between XGBoost and the other models (all p ≥ 0.109). Out-of-fold TreeSHAP identified Spectrum 1, Spectrum 2, Spectrum 5, and Spectrum 4 as the leading XGBoost features. These model-specific associations are hypothesis-generating and do not establish molecular or causal specificity. The findings support further investigation of RBC hyperspectral profiles, but the small, male-only, single-cohort design and absence of external validation preclude claims of clinical generalizability or diagnostic readiness.