Establishing Diagnosis Models for Autism Spectrum Disorder With Auditory Brainstem Response Using Machine Learning Methods.

This study aimed to develop and validate models built using machine learning methods for the diagnosis of autism spectrum disorder (ASD). A total of 851 children with suspected ASD were enrolled. Related features were acquired using the Autism Behavior Checklist and Auditory Brainstem Response (ABR) outcomes. The Shapley additive explanation (SHAP) values were used to explain the feature impact used in the models with the best performance. In addition, different methods were used to depict group differences. Our study found that significant results in ABR outcomes among children with different phenotypes mainly depend on stimulation intensity. The predominant finding is that signal conduction from the cochlear nerve to the inferior colliculus in children with ASD may be associated with autistic behaviors in a way that displays region-specific (e.g., cochlear nerve, olivary) and intensity-dependent (e.g., 45, 65 dB) complexity. The XGBoost models outperformed the other candidates, with the highest prediction accuracy (85.9%) and AUROC (0.779). Wave III and V were identified as the features with largest impact. These findings highlight atypical ABR patterns in ASD and their association with autistic behaviors, contributing to the growing literature on understanding alterations in bioelectrical performance in children with ASD.