Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Classification of Autism in Preschoolers Using Neural Encoding of Speech
Poster Session D, Thursday, October 1, 4:30 - 6:30 pm, Wangari Maathai
Hoyee Hirai1, Shaoqi Pan1,2, Y.Y. Ho1, Sitong Zhang3, Eric C. H. Poon3, Xin Qi1,3, Li Wang4, Patrick C. M. Wong1,3; 1Brain and Mind Institute, The Chinese University of Hong Kong, Shatin, Hong Kong SAR, China, 2Department of Psychology, The Chinese University of Hong Kong, Hong Kong SAR, China, 3Department of Linguistics and Modern Languages, The Chinese University of Hong Kong, Shatin, Hong Kong SAR, China, 4Department of Educational Psychology, Faculty of Education, The Chinese University of Hong Kong, Shatin, Hong Kong SAR, China
Introduction: Autistic individuals often show differences in speech and prosodic perception, which may associate with anomalies in gamma oscillations. However, most prior evidence is based on group-level analyses in school-aged or older populations; it remains unclear whether these findings generalize to individual-level detection in younger children, given greater developmental variability and increased electroencephalography (EEG) measurement noise in early childhood. Neural encoding, frequency-following responses (FFR) in particular, may provide a sensitive approach for individual-level detection in young childhood. This study aimed to examine whether an EEG-based machine learning classifier can accurately distinguish autistic from typically developing (TD) preschoolers. Methods: We recruited 149 (92 autistic and 57 TD) Cantonese-learning children aged 2 to 5 years. The mean age was 38.0 months; 79% were male. For autistic children, Autism Diagnostic Observation Schedule, Second Edition was administered to confirm autistic conditions. TD children with no reported developmental conditions were screened using the Social Communication Questionnaire Current version, with scores below 15 indicating non-autistic status. Child (age, sex at birth, gestational age, and birth weight) and parent (age, education, occupation, income, and Broader Autism Phenotype Questionnaire) characteristics were collected. Group comparisons showed no significant differences in child age, sex at birth, gestational age, or household income. Autistic children had significantly lower birth weight and older parents; however, effect sizes were small and consistent with the literature. During EEG recording, children sat on caregivers’ laps watching a silent movie of their choice or sleeping. Auditory stimuli consisting of a random sequence of syllables in three tones (/ga2/, /ga3/, /ga4/) were presented binaurally. Continuous EEG was collected at 20kHz sampling rate from Cz, M1, M2 (CPz reference, Fpz ground) using a Neuroscan SynAmps2 System. EEG neural encoding features, including early latency response (FFR) and long-latency response were extracted for each tone. Various machine learning models, including support vector machine (SVM), logistic regression, random forest, decision tree, extreme gradient boosting (XGBoost) and K-nearest neighbors (KNN), were applied with nested 5-fold cross-validation to classify autistic status (autistic vs TD) based on (1) child and parent characteristics, (2) EEG neural encoding features and (3) both combined. Classification performance was evaluated using area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV). Results: Machine learning models using only child and parent characteristics showed classification performance close to chance level (AUC= 0.51-0.56). In contrast, models using only EEG features achieved significantly better classification performance (AUC= 0.64-0.83). Adding child and parent characteristics to EEG-based models did not improve performance (AUC= 0.63-0.82), suggesting limited additional predictive gain beyond EEG neural encoding features. Among all models, XGBoost using only EEG features demonstrated the best performance, with an AUC of 0.83 (95%CI: 0.82–0.83), sensitivity of 77.2% (95%CI: 75.9%–77.9%), specificity of 74.7% (95%CI: 73.0%–75.2%), PPV of 83.8% (95%CI: 83.0%–84.2%), and NPV of 67.7% (95%CI: 67.3%–69.1%). Conclusion: EEG-based machine learning models demonstrated reliable individual-level classification of autistic from TD preschoolers. Differences in neural encoding of speech between autistic and TD children are detectable as early as the preschool years.
Topic Areas: Disorders: Developmental, Computational Approaches