Identification of temporal variations in mental workload using locally-linear-embedding-based EEG feature reduction and support-vector-machine-based clustering and classification techniques

• Recognition of mental workload is crucial to improve the safety of human–machine system.
• A simulated process control task was used to elicit different levels of mental workload.
• Mental workload was assessed by using multichannel EEG signals.
• Locally linear embedding was used to reduce the dimensionality of EEG power features.
• Support vector clustering and support vector data description methods were combined to classify EEG data.

Identifying the abnormal changes of mental workload (MWL) over time is quite crucial for preventing the accidents due to cognitive overload and inattention of human operators in safety-critical human–machine systems. It is known that various neuroimaging technologies can be used to identify the MWL variations. In order to classify MWL into a few discrete levels using representative MWL indicators and small-sized training samples, a novel EEG-based approach by combining locally linear embedding (LLE), support vector clustering (SVC) and support vector data description (SVDD) techniques is proposed and evaluated by using the experimentally measured data. The MWL indicators from different cortical regions are first elicited by using the LLE technique. Then, the SVC approach is used to find the clusters of these MWL indicators and thereby to detect MWL variations. It is shown that the clusters can be interpreted as the binary class MWL. Furthermore, a trained binary SVDD classifier is shown to be capable of detecting slight variations of those indicators. By combining the two schemes, a SVC–SVDD framework is proposed, where the clear-cut (smaller) cluster is detected by SVC first and then a subsequent SVDD model is utilized to divide the overlapped (larger) cluster into two classes. Finally, three-class MWL levels (low, normal and high) can be identified automatically. The experimental data analysis results are compared with those of several existing methods. It has been demonstrated that the proposed framework can lead to acceptable computational accuracy and has the advantages of both unsupervised and supervised training strategies.