Deep learning-assisted non-invasive pediatric tic disorder diagnosis using EEG features extracted by residual neural networks

Abstract

Early diagnosis of pediatric tic disorders (TD) is crucial for effective therapeutic intervention and management, which can significantly improve neurological development and psychological well-being from childhood through adulthood. However, current pediatric TD diagnostic methodologies suffer from low specificity and sensitivity, as they rely primarily on the subjective expertise of clinicians. Herein, we demonstrated a non-invasive approach for deep learning-assisted diagnosis of pediatric TD. A residual neural network model was developed to predict TD using electroencephalogram (EEG) signals. The optimized model analyzed preprocessed EEG data to generate diagnostic reports indicating the probability of TD occurrence, thus providing deep learning-assisted support for clinical decisions. The clinical features of EEG signals in pediatric TD are elucidated through extensive analysis. Predictive accuracy of EEG decreases over time, with short-term EEG indicating that right hemisphere EEG activity is a predominant clinical feature of TD. A computer-based application was developed and implemented to calculate the probability of TD based on individual EEG patterns, thereby assisting clinicians with diagnostic decision-making in real-world scenarios. This work not only proposes a non-invasive and accurate approach for TD diagnosis but also contributes to the early intervention and long-term management of neurological and psychological health in affected individuals.