Hyperparameter Optimization of Deep Learning Models for EEG-Based Vigilance Detection

ElectroEncephaloGraphy (EEG) signals have a nonlinear and complex nature and require the design of sophisticated methods for their analysis. Thus, Deep Learning (DL) models, which have enabled the automatic extraction of complex data features at high levels of abstraction, play a growing role in the field of medical science to help diagnose various diseases, and have been successfully used to predict the vigilance states of individuals. However, the performance of these models is highly sensitive to the choice of the hyper-parameters that define the structure of the network and the learning process. When targeting an application, tuning the hyper-parameters of deep neural networks is a tedious and time-consuming process. This explains the necessity of automating the calibration of these hyper-parameters. In this paper, we perform hyper-parameters optimization using two popular methods: Tree Parzen Estimator (TPE) and Bayesian optimisation (BO) to predict vigilance states of individuals based on their EEG signal. The performance of the methods is evaluated on the vigilance states classification. Compared with empirical optimization, the accuracy is improved from 0.84 to 0.93 with TPE and from 0.84 to 0.97 with Bayesian optimization using the 1D-UNet-LSTM deep learning model. Obtained results show that the combination of the 1D-UNet encoder and LSTM offers an excellent compromise between the performance and network size (thus training duration), which allows a more efficient hyper-parameter optimization.