An Error Detection System for Brain-Computer Interfaces
What is a Brain-Computer Interface?
Brain-Computer Interfaces (BCIs) are systems that detect and interpret electrical signals that are constantly being emitted by the brain. These miniscule signals measured in microvolts, are then converted into computer inputs. Upon capture, these brain signals undergo a computationally intensive process to extract important information in real time. A very typical system first presents the relevant stimuli on a screen to the user, which elicit corresponding brain potentials in specific areas of the brain. Metallic discs are typically used to detect these potentials which are then processed by applying filters and various signal processing algorithms to output the user's intentions. Finally, the computer interface then translates this information into commands that the computer can understand.
Electroencephalography (EEG) is typically utilised to measure brain electrical activity through metallic disks that capture electrical voltages. Gel-based electrodes use a conductive electrolyte gel to optimise signal quality by establishing a good contact between the electrode and the scalp of the user. This method is recommended for long duration sessions but they require a long time to set up and require another person to set it up for the user.
BCI and their errors
Brain-computer interfaces (BCIs) have been in existence since the 1970s. Over time, numerous advancements have been made to improve their practicality. These systems facilitate computer control through brain activity. While BCI systems are primarily tailored for individuals with mobility impairments, such as those afflicted with motor neuron degenerative conditions like Amyotrophic Lateral Sclerosis (ALS), they are accessible to a wider user base. It is imperative for these systems to have reliable performance to sustain user engagement. Despite improvements in reliability over the years, instances of low accuracies and low information transfer rates persist [1]. Higher classification accuracy correlates with increased adoption of BCI technology in everyday life.
Since the recorded EEG signals are very low in magnitude, they are more susceptible to environmental interferences and user changes [2]. Due to this, the system typically misclassifies the user’s intent and therefore the computer selects an input that does not correspond to the user’s intent. Diverse BCI types have emerged, each utilising distinct signals for computer control however, certain systems still exhibit suboptimal accuracies, relying heavily on user’s familiarity with BCI systems. Due to this, frequent misclassifications may deter users from using the system.
The automatic detection of an erroneous signal holds great practical significance and can be applied across different types of BCI’s including P300 event-related potentials, motor imagery and SSVEP, to enhance their accuracy and reliability [3].
SSVEP, or Steady-State Visually Evoked Potential, is a type of brain response that occurs when a person looks at a visual stimulus that is flickering at a constant frequency. This response is a steady-state because it remains stable over time when the stimulus is consistently presented. SSVEPs are commonly used in neuroscience research and applications such as brain-computer interfaces (BCIs) due to their distinct and easily recognizable frequency components.
Given that these systems produce errors, this error detection process could be used in all circumstances and the outcome will result in an increase in information transfer rates and a better motivation to use the system. In this project, the ErrP will be applied on an SSVEP-based BCI system that can be used for a motorized bed application.
ErrP, or Error-Related Potential, is a type of event-related potential (ERP) in the brain that occurs in response to errors or perceived mistakes. These brain signals are typically recorded using electroencephalography (EEG) and are characterized by distinct changes in the EEG waveform following the detection of an error.
Why is ErrP detection required?
BCI systems have a certain level of accuracy that may vary depending on the user's response, past experience with BCI systems and the environment. An error occurs when the feedback given by the BCI does not match the user’s intent, resulting in an unwanted control action. This may be frustrating to the user particularly if it occurs regularly as the user would have to cancel or correct the unwanted action causing unwanted delays. Due to this, error detection systems have been designed to detect these errors and cancel the unwanted action automatically, thus making the system easier to run whilst increasing the information transfer rate (ITR).
ErrPs are a set of specific potentials that are found in an EEG signal whenever a person makes or perceives an error [4]. Therefore if a person makes an error or watches a system do an error, this potential is expected to be present. These error potentials are located in the fronto-central and centro-parietal regions and are typically made up of two main components. The first component is a positive potential which peaks at 200ms and the second component peaks at 320ms after an erroneous response. Two negative peaks are present as well and occur at around 250ms and 450ms after an incorrect response [5]. A typical ErrP waveform could be seen Figure 1.
Proposed method
As was seen in most papers, the ErrP signal is typically detected by using the ErrP time features to classify between signals that contain an ErrP and those that do not contain an ErrP. This method gave an average accuracy of around 73.18%. It is proposed that ErrP classification may be improved if SSVEP features are used in combination with ErrP features as this method was previously proven for a P300-based BCI. Since the ErrP is detected using time-domain features whereas the SSVEP features are in the frequency domain, probability values were combined using an ensemble classifier to detect signals with ErrP and signals without an ErrP.
Data gathering
To capture the EEG data, subjects wear an EEG cap and electrodes are placed at specific positions on the scalp to capture the brain signals. The subject is then asked to attend to a number of flickering stimulus by the use of a cue and the EEG data is recorded. Specifically, one of eight boxes is highlighted at a time and the subject is requested to attend to the flickering stimulus for a number of seconds. This process is repeated to ensure that each box is selected for a number of times, ensuring that enough data is captured for further processing and analysis. The interface used was originally designed to control a motorised bed application [6] but in this project the scope is to collect the EEG data. The setup could be seen in Figure 2.
Results
In the analysis carried out in this work, the combination of the posterior probabilities resulting from an ErrP classifier and an SSVEP strength detector gave the highest ErrP detection results. Specifically, the incorrect SSVEP detection carried out by the system was detected with an accuracy of 78% which was 9.76% higher than when using the ErrP time domain signal alone. These results compare well to those obtained by Aniana Cruz [3], in which the highest accuracy achieved for ErrP detection for a P300-based BCI system was 85.1%. Therefore, from this research, it was concluded that the ensemble classifier increases the accuracy of the incorrect SSVEP detection when compared to the accuracy obtained by the ErrP classifier alone.
Further developments
In future work, various factors of this work could be addressed. The BCI system could be tested in an online manner such that the system automatically detects an error and cancels the corresponding control function. This would make the system easier to use and closer to a real-life application.
Conclusion
The error detection algorithm is a stepping-stone in improving the overall performance of BCI systems and could be used in various systems to detect errors in an automatic manner, without requiring the user to carry out another task to cancel any incorrectly classified actions. By detecting errors, the system could be made more efficient and accurate and hence more people will be enticed to use it.
References
[1] A. Kübler, E. M. Holz, E. W. Sellers, and T. M. Vaughan, “Toward independent home use of brain-computer interfaces: A decision algorithm for selection of potential end-users,” Archives of Physical Medicine and Rehabilitation, vol. 96, no. 3, 2015.
[2] P. W. Ferrez and J. del R. Millan, “Error-related EEG potentials generated during simulated brain–computer interaction,” IEEE Transactions on Biomedical Engineering, vol. 55, no. 3, pp. 923–929, 2008.
[3] A. Cruz, G. Pires, and U. J. Nunes, “Double ERRP detection for automatic error correction in an ERP-based BCI speller,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 26, no. 1, pp. 26–36, 2018.
[4] M. Spüler and C. Niethammer, “Error-related potentials during continuous feedback: Using EEG to detect errors of different type and severity,” Frontiers in Human Neuroscience, vol. 9, 2015.
[5] R. Chavarriaga and J. del Millan, “Learning from EEG error-related potentials in noninvasive brain-computer interfaces,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 18, no. 4, pp. 381–388, 2010.
[6] R. Zerafa, “Towards SSVEP-Based BCI Applications for Real-World Environments,” University of Malta, PhD dissertation, 2022
This article reflects the MSc work of Fabian Camilleri who was supervised by Prof. Tracey Camilleri and Prof. Kenneth Camilleri.
Comments