Developing a reliable and practical multi-domain model to facilitate the diagnosis of ADHD in older preschool children

Abstract

Abstract A multi-method, multi-informant approach is emphasized for the evaluations of attention-deficit/hyperactivity disorders (ADHD) in preschool children due to the complexity and challenges of diagnosis at this stage. Most artificial intelligence (AI) studies on the automated detection of ADHD used a single type of data. This study aims to create a reliable multimodal AI-detection system for facilitating the diagnosis of ADHD among older preschool children. 78 older preschool children were recruited; 43 (mean age: 68.07 ± 6.19 months) of them were diagnosed with ADHD and 35 (mean age: 67.40 ± 5.44 months) of them were with typical development (TD). Machine learning (ML) and deep learning (DL) methods were adopted to develop three individual predictive models by using electroencephalography (EEG) data recording with a wearable wireless device, scores of the computerized attention assessment via Conners’ Kiddie Continuous Performance Test (K-CPT), and ratings of the ADHD-related symptom scales; finally, one ensemble model was merged. The results suggest that teacher ratings, K-CPT reaction time, and occipital high-frequency EEG band power values are significant features in identifying older preschool children with ADHD, and the ensemble model can achieve an accuracy of 0.974. The present study can respond to the three issues in most ADHD-related AI studies: the utility of wearable technologies, databases derived from different types of ADHD diagnostic instruments, and appropriate interpretability of the models. This established multimodal system can be reliable and practical in discriminating ADHD from TD and further facilitate the clinical diagnosis of preschool ADHD.