Classification of electroencephalography (EEG) signals corresponding to imagined speech production is important for the development of a direct-speech brain-computer interface (DS-BCI). Deep learning (DL) has been utilized with great success across several domains. However, it remains an open question whether DL methods provide significant advances over traditional machine learning (ML) approaches for classification of imagined speech. Furthermore, hyperparameter (HP) optimization has been neglected in DL-EEG studies, resulting in the significance of its effects remaining uncertain. In this study, we aim to improve classification of imagined speech EEG by employing DL methods while also statistically evaluating the impact of HP optimization on classifier performance. We trained three distinct convolutional neural networks (CNN) on imagined speech EEG using a nested cross-validation approach to HP optimization. Each of the CNNs evaluated was designed specifically for EEG decoding. An imagined speech EEG dataset consisting of both words and vowels facilitated training on both sets independently. CNN results were compared with three benchmark ML methods: Support Vector Machine, Random Forest and regularized Linear Discriminant Analysis. Intra- and inter-subject methods of HP optimization were tested and the effects of HPs statistically analyzed. Accuracies obtained by the CNNs were significantly greater than the benchmark methods when trained on both datasets (words: 24.97%, p < 1 × 10-7, chance: 16.67%; vowels: 30.00%, p < 1 × 10-7, chance: 20%). The effects of varying HP values, and interactions between HPs and the CNNs were both statistically significant. The results of HP optimization demonstrate how critical it is for training CNNs to decode imagined speech.
Keywords: brain–computer interface (BCI); convolutional neural networks (CNN); deep learning; electroencephalography (EEG); hyperparameter optimization; imagined speech; machine learning.