Hyperparameter Tuning of a Deep Learning EEG-based Neural Network for the Diagnosis of ADHD

In order to implement an automatic-computer-aided system for the diagnosis of the Attention-Deficit Hyperactivity Disorder (ADHD), a Deep Learning Multihead Convolutional Based model (EEG-MHCNet) was developed in previous work. To obtain the best model performance, five hyperparameters were studied. This procedure led us to train the model with all hyperparameter combination values. As a result, 1920 executions were performed and took 15.5 days to complete the training process. Finally, the model obtained a better f1-score than the state-of-art ones. To improve these results, more values should be tested. The main problem is the high computational cost of executing such a considerable number of combinations. In this paper, a null hypothesis testing procedure has been applied to find which hyperparameter values have statistical confidence in the model’s performance. By doing this procedure, some values could be discarded thus reducing the number of combinations and the execution of training time. Results show that the number of executions can be reduced from 1920 to 320 without loss of performance with a 95% of confidence level, thus enabling us to considerably reduce the computational training cost from 15.5 days to 1.3 days.