Classification of neurodevelopmental disorders and typical development using deep learning and a portable patch-type electroencephalography device

Abstract

Autism spectrum disorder (ASD) and attention-deficit/hyperactivity disorder (ADHD) are common neurodevelopmental disorders (NDDs) in children and often co-occur (ASD + ADHD), complicating the diagnosis. The diagnostic process is lengthy and subjective, relying heavily on expert knowledge, which limits accessibility. Electroencephalography (EEG) offers potential as a biomarker but requires skilled technicians for measurements and can be stressful for children with NDDs. This study aimed to provide a more accessible diagnostic support tool. We used a portable EEG device with a low participant burden and a deep learning model to distinguish between the typical development group (TD) and the NDD group, comprising children with ASD, ADHD, and ASD + ADHD. Resting-state EEG data were recorded for 5 min using the portable HARU-2 device, which features three channels placed on the forehead of 163 participants (87 TD, 76 NDD). A deep learning model combining a one-dimensional convolutional neural network and a transformer encoder was developed to analyze the EEG data. In 5-fold cross-validation, the model achieved an area under the curve (AUC) of 0.713 and a balanced accuracy (bACC) of 67.5% for classifying NDD and TD. Exploratory evaluation of out-of-fold predictions stratified by clinical phenotype showed an AUC of 0.793 and bACC of 75.0% for ASD vs. TD, 0.817 and 79.5% for ADHD vs. TD, and 0.667 and 62.7% for ASD + ADHD vs. TD. These findings suggest that portable EEG devices combined with deep learning models may serve as accessible adjunctive tools for NDD screening in children.