Deep Learning Tools for Autism Screen ing using ERPs as Biomarkers - Tiny Recursive Model for Autism Spectrum Disorder Screening

Sammanfattning

Autism Spectrum Disorder (ASD) is a neurodevelopmental condition whose earlydiagnosis remains difficult: behavioural assessments are subjective and often appliedlate. The brain’s stimulus-locked electrical responses, recorded with electroencephalography (EEG), are a candidate neurophysiological biomarker, but it is notwell established whether single-trial ERP epochs combined with modern deep learningcan support subject-level ASD screening on the small datasets typical of clinicalEEG research.This thesis investigates that question using EEG recordings collected from 92 childrenaged 7–12 (42 ASD, 50 neurotypical) during an auditory stimulus paradigm with a 16channel OpenBCI system. A complete pipeline is built and reported: a preprocessingstack (band-pass 0.5–40Hz, notch, common-average reference, optional Gaussiansmoothing, baseline correction, peak-to-peak artifact rejection) yielding 7,316 singletrial stimulus-locked epochs, and a Tiny Recursive Model (TRM) with ∼ 22,000parameters that combines self-attention with recursive depth and demographicfeature combination. The TRM operates directly on the single-trial epochs and isevaluated against five classical baselines (Logistic Regression, Linear SVM, SVMRBF, Random Forest, XGBoost) trained on the same input under identical 5-foldsubject-grouped stratified cross-validation.The TRM achieves a subject-level AUC of 0.867, exceeding the strongest classical baseline. Saliency analysis localizes the model’s attention to the 200–500mspost-stimulus window over parietal channels, consistent with the latencies of theN2 and P3a/P3b components. However, a complementary group-level analysis ofpeak amplitudes and latencies extracted from per-subject averaged ERPs finds nosignificant differences in the ERP components on the group level. This contradictionmakes the reasoning of the model difficult to accept. The TRM is best interpretedas performing demographically-conditioned ERP classification, since it also makeuse of age and gender as added contextual information.