Advancing passive BCIs: a feasibility study of two temporal derivative features and effect size-based feature selection in continuous online EEG-based machine error detection

Abstract

The emerging integration of Brain-Computer Interfaces (BCIs) in human-robot collaboration holds promise for dynamic adaptive interaction. The use of electroencephalogram (EEG)-measured error-related potentials (ErrPs) for online error detection in assistive devices offers a practical method for improving the reliability of such devices. However, continuous online error detection faces challenges such as developing efficient and lightweight classification techniques for quick predictions, reducing false alarms from artifacts, and dealing with the non-stationarity of EEG signals. Further research is essential to address the complexities of continuous classification in online sessions. With this study, we demonstrated a comprehensive approach for continuous online EEG-based machine error detection, which emerged as the winner of a competition at the 32nd International Joint Conference on Artificial Intelligence. The competition consisted of two stages: an offline stage for model development using pre-recorded, labeled EEG data, and an online stage 3 months after the offline stage, where these models were tested live on continuously streamed EEG data to detect errors in orthosis movements in real time. Our approach incorporates two temporal-derivative features with an effect size-based feature selection technique for model training, together with a lightweight noise filtering method for online sessions without recalibration of the model. The model trained in the offline stage not only resulted in a high average cross-validation accuracy of 89.9% across all participants, but also demonstrated remarkable performance during the online session 3 months after the initial data collection without further calibration, maintaining a low overall false alarm rate of 1.7% and swift response capabilities. Our research makes two significant contributions to the field. Firstly, it demonstrates the feasibility of integrating two temporal derivative features with an effect size-based feature selection strategy, particularly in online EEG-based BCIs. Secondly, our work introduces an innovative approach designed for continuous online error prediction, which includes a straightforward noise rejection technique to reduce false alarms. This study serves as a feasibility investigation into a methodology for seamless error detection that promises to transform practical applications in the domain of neuroadaptive technology and human-robot interaction.

Keywords: Brain Computer Interface; artifact rejection; error detection; feature extraction; human-robot interaction; signal processing; time-derivative features.

Figure 1

Experimental setup and procedure. (A) Participant prepared with EEG and EMG electrodes wearing orthosis on their right arm (B) Visualization of the different steps in the experimental procedure. Reprinted from Kueper et al. (2024).

Figure 2

Derivation process of temporal derivative features. Top: the derivation of temporal difference (Diff) features, capturing variations over 160 ms intervals. Bottom: the derivation of temporal dynamics (Dyn) features, capturing variations over 400 ms intervals. Subtractive operations are indicated by lines. The temporal difference values in the bottom plot differ from those in the top plot and are only used to calculate the temporal dynamics values. Bold fonts indicate features that were used for classification.

Figure 3

Feature selection process flow. Cohen's d was computed for each feature and then used to select common channels as spatial filters and for the final selection of 150 features used in the model.

Figure 4

Feature effect size topographies. Averaged Cohen's d topographies for the different features in the offline training (N = 8 participants, with 48 errors and 192 non-error epochs per participant). Higher absolute values indicate a higher discriminability between the two classes (error/no error), with positive values corresponding to a positive deflection in the ERP amplitude. See Section 4.1 for details.

Figure 5

Online simulation ERPs and prediction scores. ERPs and prediction scores after error onsets in the online simulation analysis (N = 8 participants, two datasets with each N_error = 6 per participant). The figure presents the grand mean with red line, individual means with gray lines and the standard error of the individual means as shaded areas in each subplot. The thicker gray line highlighted within these represents the participant of the online prediction stage. (Top-left) ERPs at Cz. (Top-right) ERPs at CPz. (Bottom-left) ERPs at Pz. (Bottom-right) the prediction scores of the SVM classifier. The decision threshold for online error prediction is indicated by the dashed line. The yellow box highlights the window used for classification, aligned with the yellow line in the bottom-right plot, showing the peak prediction score for the data window (0–800) ms post-error, in correspondence with the window it was trained on.

Figure 6

Grand averaged online simulation ErrP topographies. Grand averaged ErrP topographies in the online simulation of eight participants with 2*6 errors each (N = 96).

Figure 7

Window length analysis. The figure compares the classification accuracies of the proposed and reference methods for window lengths of 50 and 80 ms. The x-axis, logarithmized for clarity, represents the number of features selected in the proposed method. For context, the reference method uses 1,024 temporal average features when using 50 ms windows and 640 features when using 80 ms windows.

Figure 8

Impact of feature types on average balanced accuracy (BA). The figure shows the classification accuracy achieved with different numbers of selected features under five different conditions: using only temporal average features, temporal difference features, temporal dynamics features, a combination of temporal average and temporal difference features, and a combination of temporal average, temporal difference, and temporal dynamics features. Feature selection was guided by the ranked Cohen's d values. For comparison, the average BA of the reference method, which uses 1,024 temporal average features, is shown as a horizontal line. The gray zone indicates the ideal range for selecting the number of features, specifically between 200 and 300, which is associated with achieving the highest BA. The x-axis, representing the number of features, is logarithmized.