Time Domain Classification for Brain-Computer Interface Based Problems
Poster No:
1461
Submission Type:
Abstract Submission
Authors:
Aiden Rushbrooke1, Jason Lines1, Thomas Kinna1, Anthony Bagnall2, Saber Sami1
Institutions:
1University of East Anglia, Norwich, Norfolk, 2University of Southampton, Southampton, Hampshire
First Author:
Co-Author(s):
Introduction:
Electroencephalographic (EEG) data classification is a ubiquitous task across various disciplines. This task is of particular significance in Brain Computer Interfacing (BCI) enabling direct communication between the brain and external devices.
Despite the prevalence of EEG classification many of the current popular approaches use standard classifiers rather than algorithms designed to explicitly exploit information within the time domain.
These time series classification (TSC) methods have been shown to perform better than standard approaches such as deep learning on a range of time series machine learning tasks and problem domains[1]. We explore the application of TSC algorithms to BCI based EEG classification problems and highlight their potential to enhance performance and interpretability compared to conventional classification techniques.
Despite the prevalence of EEG classification many of the current popular approaches use standard classifiers rather than algorithms designed to explicitly exploit information within the time domain.
These time series classification (TSC) methods have been shown to perform better than standard approaches such as deep learning on a range of time series machine learning tasks and problem domains[1]. We explore the application of TSC algorithms to BCI based EEG classification problems and highlight their potential to enhance performance and interpretability compared to conventional classification techniques.
Methods:
Classifier Selection:
For our experimental TSC framework we used the recent Random Convolutional Kernel Transform (ROCKET) algorithm[2]. A large number of random convolutions are used to transform the data, which is then used to train a ridge regression classifier. The main advantage of ROCKET methods are that they are fast and achieve accuracy comparable to state of the art approaches. To enhance BCI efficiency, we explored an adaptation of ROCKET known as MINI-ROCKET[3]. This was selected for its increased speed and deterministic nature.
In order to provide a comparison between TSC and traditional methods we employed Support Vector Machine (SVM), Random Forest, and K-Nearest Neighbors (KNN), chosen for their wide applicability in EEG data classification[4].
Datasets:
Five BCI datasets were used to compare performance. The first was a novel dataset for a BCI benchmark experiment where 28 participants were asked to press a button when a stimulus appears on a screen.
The data was recorded at 1000Hz with 32 EEG channels. Each participant completed 40 button presses, spaced 1.5 second apart. Each participant also recorded resting state data. To form the BCI experiment 1 second segments around each press were taken, starting 200ms before the initial stimulus. This dataset, whilst not currently publicly available, will be released on timeseriesclassification.com in the future. The other 4 datasets were formed from a larger existing BCI motor data collection[5][6] (see table 1).
Experimental Design:
A leave-one-out strategy was used to reduce the risk of bias caused by a train test split. The model is trained on all but one participant, who is used to evaluate performance. This is then repeated on a new model for each participant, and average accuracy calculated. The experiments were run using the Aeon toolkit.
For our experimental TSC framework we used the recent Random Convolutional Kernel Transform (ROCKET) algorithm[2]. A large number of random convolutions are used to transform the data, which is then used to train a ridge regression classifier. The main advantage of ROCKET methods are that they are fast and achieve accuracy comparable to state of the art approaches. To enhance BCI efficiency, we explored an adaptation of ROCKET known as MINI-ROCKET[3]. This was selected for its increased speed and deterministic nature.
In order to provide a comparison between TSC and traditional methods we employed Support Vector Machine (SVM), Random Forest, and K-Nearest Neighbors (KNN), chosen for their wide applicability in EEG data classification[4].
Datasets:
Five BCI datasets were used to compare performance. The first was a novel dataset for a BCI benchmark experiment where 28 participants were asked to press a button when a stimulus appears on a screen.
The data was recorded at 1000Hz with 32 EEG channels. Each participant completed 40 button presses, spaced 1.5 second apart. Each participant also recorded resting state data. To form the BCI experiment 1 second segments around each press were taken, starting 200ms before the initial stimulus. This dataset, whilst not currently publicly available, will be released on timeseriesclassification.com in the future. The other 4 datasets were formed from a larger existing BCI motor data collection[5][6] (see table 1).
Experimental Design:
A leave-one-out strategy was used to reduce the risk of bias caused by a train test split. The model is trained on all but one participant, who is used to evaluate performance. This is then repeated on a new model for each participant, and average accuracy calculated. The experiments were run using the Aeon toolkit.
Results:
The results of the experiments are displayed in the tables below, with average accuracy shown in Table 1 and average time taken to fit a model in Table 2. For all classifiers except for SVM, prediction speed was around 1ms. Overall Mini-ROCKET was the best performing classifier in terms of accuracy. Unsurprisingly KNN, whilst being the fastest classifier used, performed much worse than the rest. A time-accuracy plot for the ButtonPress dataset is shown in Figure 2a. From this we can clearly see that the two best performing are MINI-ROCKET and Random Forest. We also tested reducing the training size of the ButtonPress dataset with the two best classifiers, the results of which are shown in Figure 2b. The results show even with a low number of instances, MINI-ROCKET performs well, and outperforms Random Forest at all points.
·Results Tables
Conclusions:
Our experiments show that TSC methods have potential when applied to BCI problems, achieving high accuracy whilst needing little pre-processing or prior knowledge. The speed of models such as those based on ROCKET could be used for real time BCI classification experiments. Further work would be to expand both the range of datasets and models to evaluate the generalisable nature of TSC methods to real-time BCI adaptations and test their potential for clinical usage.
Modeling and Analysis Methods:
Classification and Predictive Modeling 1
Motor Behavior:
Brain Machine Interface 2
Keywords:
Computing
Data analysis
Electroencephaolography (EEG)
Other - Classification
1|2Indicates the priority used for review
Provide references using author date format
series classification bake off: a review and experimental evaluation of recent
algorithmic advances. Data Mining and Knowledge Discovery, 31(3):606–
660.
[2] Angus Dempster, Fran ̧cois Petitjean, and Geoffrey Webb(2020). ROCKET: Ex-
ceptionally fast and accurate time series classification using random con-
volutional kernels. Data Mining and Knowledge Discovery, 34:1454–1495.
[3] Angus Dempster, D Schmidt, and Geoffrey Webb (2021). Minirocket: A very fast
(almost) deterministic transform for time series classification. In proceedings
of the 27th ACM SIGKDD International Conference on Knowledge Discov-
ery and Data Mining.
[4] Fabien Lotte, Laurent Bougrain, Andrzej Cichocki, Maureen Clerc, Marco
Congedo, Alain Rakotomamonjy, and Florian Yger (2018). A review of classifica-
tion algorithms for eeg-based brain-computer interfaces: A 10-year update.
Journal of Neural Engineering, 15, 02.
[5] G. Schalk, D.J. McFarland, T. Hinterberger, N. Birbaumer, and J.R. Wol-
paw (2004). Bci2000: a general-purpose brain-computer interface (bci) system.
IEEE Transactions on Biomedical Engineering, 51(6):1034–1043.
[6] Gerwin Schalk, Dennis J McFarland, Thilo Hinterberger, Niels Birbaumer,
and Jonathan R Wolpaw (2022). “eeg motor movement/imagery dataset”.