Predictive analytics in motor imagery brain-computer interfaces using deep learning techniques

Abstract

Deep learning (DL) has emerged as an outstanding processing tool for predictive analytics and decision-making in electroencephalogram (EEG)-based motor imagery brain-computer interface (MIBCI) systems in recent years. DL approaches provide an effective way to extra information from the EEG data and produce accurate predictive outcomes. However, there are still challenges in building high-quality DL models for MIBCI developments. First, overfitting is a common issue in modelling motor imagery (MI) data. MI data collection is labour-intensive and time-consuming, so only limited data can usually be obtained, easily leading to an overfitting issue while using the DL method. The low signal-to-noise ratio (SNR) of the EEG data is another significant obstacle to establishing well-performing DL methods since much noise is recorded on the scalp during the non-invasive data collection. In addition, there is a significant difference in EEG data between individuals, making it difficult to apply DL approaches to different people without major modifications. Finally, the long re-training period for the target user of the MIBCI system is also a challenging issue, which seriously degrades the user experience. To address the above issues, this thesis develops multiple advanced DL approaches for MIBCI systems. In the first approach, a new objective function, a combination of smoothed categorical cross-entropy (with label smoothing) and center loss, is proposed for accurate MI classification. It decreases the risk of the overfitting issue due to small sample size and simultaneously ensures the discriminative power of the deep features. In the second approach, a correntropy-based center loss (CCL) is introduced to assist the training of DL methods. It addresses the negative effect of low SNR of EEG data and increases the discriminative ability of deep features. In the third approach, a shallow inception domain adaptation network is proposed to leverage the MI data from multiple subjects to train DL models. A novel combined loss function is proposed and optimized to reduce the marginal and class conditional discrepancies of deep features. The proposed method extracts informative features from highly varied EEG data from different subjects and achieves great performance in the MI classification task. In the fourth approach, a novel learning scheme called temporal episode relation learning (TERL) is introduced to develop practical MIBCI systems. TERL encodes the temporal pattern between trials in each episode during the training to improve the classification performance. The model optimized by the TERL scheme can be directly applied to a new user without the need of retraining, which tackles the issue of long-retraining time and improves user experience. These approaches are evaluated by experiments on well-recognized public MI benchmarks. Experimental results show that they have better performance than the existing methods and have the potential to be widely used in real-world MIBCI systems in the future.