Machine learning models effectively distinguish attention-deficit/hyperactivity disorder using event-related potentials

Abstract

Accurate diagnosis of Attention-Deficit/Hyperactivity Disorder (ADHD) is a significant challenge. Misdiagnosis has significant negative medical side effects. Due to the complex nature of this disorder, there is no computational expert system for diagnosis. Recently, automatic diagnosis of ADHD by machine learning analysis of brain signals has received an increased attention. This paper aimed to achieve an accurate model to discriminate between ADHD patients and healthy controls by pattern discovery. Event-Related Potentials (ERP) data were collected from ADHD patients and healthy controls. After pre-processing, ERP signals were decomposed and features were calculated for different frequency bands. The classification was carried out based on each feature using seven machine learning algorithms. Important features were then selected and combined. To find specific patterns for each model, the classification was repeated using the proposed patterns. Results indicated that the combination of complementary features can significantly improve the performance of the predictive models. The newly developed features, defined based on band power, were able to provide the best classification using the Generalized Linear Model, Logistic Regression, and Deep Learning with the average accuracy and Receiver operating characteristic curve > %99.85 and > 0.999, respectively. High and low frequencies (Beta, Delta) performed better than the mid, frequencies in the discrimination of ADHD from control. Altogether, this study developed a machine learning expert system that minimises misdiagnosis of ADHD and is beneficial for the evaluation of treatment efficacy.

Keywords: Attention deficit hyperactivity disorder; Band power; Classification; Event-related potentials; Frequency bands; Machine learning.

Fig. 1

Preliminary classification results for 4 frequency bands; heat map showing the accuracy of the prediction models for Attention Deficit/Hyperactivity Disorder. The color scale indicates the value of the accuracy, the intensity increases from red to yellow. Each column represents one model, and each row represents a feature. In each of the frequency bands, 182 times the execution has been performed, so that at each run, only one of the algorithms and a single feature is used

Fig. 2

ROC curves obtained from seven machine learning algorithms in four frequency bands when proposed features in Tables 1, 2, 3 and 4 are used. Generalized Linear Model, Logistic Regression, and Deep Learning display a perfect discrimination in all four frequencies

Fig. 3

Pediction mean accuracy and Prediction mean error for frequency bands