Explainable machine learning characterizes temporally structured EEG differences in pragmatic processing between children with autism and typically developing peers

Pragmatic impairment is a core feature of autism spectrum disorder (ASD), yet its neural mechanisms remain poorly understood within a systematic explanatory framework. Traditional research focusing on isolated ERP components has been limited in capturing the temporal organization of pragmatic comprehension. Guided by a multi-stage framework of pragmatic processing, this study combined theory-driven feature construction with explainable machine learning to examine within-dataset group-discriminative EEG patterns during a context-sentence matching task. EEG data were collected from 103 children with ASD and 107 typically developing children (aged 4–10 years). Neural features were constructed from overlapping time-window × ROI measures, and classification analysis was performed using L2-regularized logistic regression within a nested cross-validation framework, followed by Haufe-transformed feature interpretation, bootstrap stability assessment, and complementary temporal and analytical-specification controls. The classifier achieved modest but statistically above-chance group discrimination (AUC = 0.652, empirical p = 0.001), indicating limited group-level separability within the current sample. Haufe-transformed pattern analysis showed a non-uniform descriptive distribution of model-associated EEG information across temporal ranges and scalp ROIs, with the largest mean absolute Haufe-pattern magnitude in the early range (0–300 ms) and over the temporal scalp ROI, middle-range concentration over the frontocentral scalp ROI, and predominantly temporal-scalp contributions in the late range. Exploratory brain-behavior analyses within the ASD group indicated opposite-direction associations of early- and late-temporal features with receptive vocabulary ability. Repeated cross-validated evaluation showed only limited improvement in within-dataset out-of-fold performance when the EEG features were added beyond age, and these findings should not be interpreted as direct behavioral evidence of pragmatic competence. These findings provide a within-sample description of how modest group-discriminative EEG information was distributed across time and scalp ROIs during the task and motivate independent replication and models that explicitly estimate temporal dependencies.