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

Abstract

Autism Spectrum Disorder (ASD) is a neurodevelopmental condition whose early diagnosis remains difficult: behavioural assessments are subjective and often applied late. The brain’s stimulus-locked electrical responses, recorded with electroencephalography (EEG), are a candidate neurophysiological biomarker, but it is not well established whether single-trial ERP epochs combined with modern deep learning can support subject-level ASD screening on the small datasets typical of clinical EEG research. This thesis investigates that question using EEG recordings collected from 92 children aged 7–12 (42 ASD, 50 neurotypical) during an auditory stimulus paradigm with a 16 channel OpenBCI system. A complete pipeline is built and reported: a preprocessing stack (band-pass 0.5–40Hz, notch, common-average reference, optional Gaussian smoothing, baseline correction, peak-to-peak artifact rejection) yielding 7,316 single trial stimulus-locked epochs, and a Tiny Recursive Model (TRM) with ∼ 22,000 parameters that combines self-attention with recursive depth and demographic feature combination. The TRM operates directly on the single-trial epochs and is evaluated against five classical baselines (Logistic Regression, Linear SVM, SVM RBF, Random Forest, XGBoost) trained on the same input under identical 5-fold subject-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 200–500ms post-stimulus window over parietal channels, consistent with the latencies of the N2 and P3a/P3b components. However, a complementary group-level analysis of peak amplitudes and latencies extracted from per-subject averaged ERPs finds no significant differences in the ERP components on the group level. This contradiction makes the reasoning of the model difficult to accept. The TRM is best interpreted as performing demographically-conditioned ERP classification, since it also make use of age and gender as added contextual information.