EEG-based Cognitive Performance Discrimination Between Athletes and Non-Athletes

In the world of sports, winning in a sports competition depends on various factors, and different sports may require different strategies and skills such as physical and health conditions, technical skills, team support, environment, mental abilities, and cognitive performance. The ability to make quick decisions, stay focused under pressure, and react to unexpected situations can be the deciding factor in many sports competitions. Compared to non-athletes or non-specialized trained persons, athletes often require a higher level of cognitive performance.

This article explores the discrimination between athletes and non-athletes, extending beyond the sporting task, and delves into the neuroscience of cognitive abilities — Attention. By analysing brain signals acquired while performing the cognitive task, we can understand the changes in brain mechanisms, and the use of features computed from brain signals enhances abilities to understand, predict, and advance the improvement of cognitive performance in a wide range of applications, from sports optimization to detection of cognitive stages.

The dataset used in the following analysis was obtained by Pei, X. (2022).

Pei, X., Xu, G., Zhou, Y., et al. A simultaneous electroencephalography and eye-tracking dataset in elite athletes during alertness and concentration tasks. Sci Data 9, 465 (2022). doi:10.1038/s41597-022-01575-0

Athletes and non-athletes were recruited and required to perform the cognitive task — Alertness Behaviour Task (ABT) for 10 minutes. Electroencephalography (EEG) signals were recorded throughout. Prior to the ABT task, 4-minute baseline recordings (2-minute eyes opened; 2-minute eyes closed) were conducted to record participants’ current cognitive stage.

Figure: Illustration shows experimental setting, participants respond to the ABT task on the laptop as soon as observe the skipped line (presented in a dotted line, red arrow). No response is needed when the line has a single jump (presented in a solid line, dark blue arrows). EEG signals were recorded throughout the task.

The Alertness Behaviour Task (ABT) was modified from the Mackworth clock task. Pei, X. (2022) modified the task to use a solid line as a stimulus, instead of a red dot. The solid line moved around the circle step-by-step (single jump) in a clockwise direction. Participants were instructed to press spacebar as soon as they observed the solid line skipped one step (double jump). Each step was considered as one trial. There were 990 trials in total. The number of double jump trials was approximately 10% of the total number of trials. Only double jump trials in which participants responded correctly were used for analysis.

The Electroencephalography (EEG) signals were acquired at a 1000 Hz sampling frequency, 64 channels (Neuroscan, Australia). 19 participants from athlete group and 19 participants from non-athlete group were randomly selected for the analysis.

EEG signals were pre-processed by following steps:

1. Re-sampling: down-sampling from 1000 Hz to 256 Hz
2. Filtering: bandpass filter at 0.1–40 Hz
3. Re-referencing: common average reference
4. Epoch extraction: at 0–1 second after event onset
5. Independent Component Analysis (ICA) decomposition and rejection: remove remaining artefacts i.e., eye-blinking, muscle tensions, channel noises.

The power spectral density (PSD) or band power of five frequency bands were calculated from filtered signals. Ranges of the five frequency bands are listed below:

1. Delta band at 0.1–4 Hz
2. Theta band at 4–8 Hz
3. Alpha band at 8–12 Hz
4. Beta band at 12–30 Hz
5. Gamma band at 30–40 Hz

The k-Nearest Neighbours model was chosen to perform classification between athlete and non-athlete groups, and band powers were used as features for the classifier model. Data of each group was split into two sets; 80% or 30 participants for the train set, and 20% or 8 participants for the test set. The ratio of number of athlete to number of non-athelete in each dataset was set at 50:50. The Stratified k-fold was adopted to split the train set into separate folds for training the model, and GridSearch cross-validation was applied to tune the k parameter (k = 1–30). The model’s feature selection was based on two factors; 1) attention-related frequency bands (alpha, beta, and gamma bands) and 2) channels in brain metrics that demonstrate statistically significant differences. The significant channels reported by Pei, X. (2022) were F1, Fz, F2, FC1, FCz, FC2, C1, Cz, C2, CP3, CP1, CPz, P5, P3, and P1.

Figure: Schematic scalp map shows channel locations that are significantly different in P300-ERP between athlete and non-athlete groups. Channel locations marked in blue colour represent the frontal region, yellow colour represent the central region, and pink colour represent the parietal region.

The ABT task was evaluated only for correct responses. Double jump onset and participants responded correctly. The task responses were grouped in 10 windows, 1 minute for each window or approximately 99 trials. The plot below shows that both athlete and non-athlete groups have a decreasing trend over time. Performance of the non-athlete group tends to decrease linearly whereas the athlete group decreased and recovered in cycle.

Figure: Graphs show the number of correct responses over 10 minutes. The solid line represents mean values, and shade represents standard error of mean. There is no statistical difference between two groups.

By assessing the difference in frequency band power between athlete and non-athlete groups, the topographic maps revealed that the athlete group had greater band power in the frontal region in alpha, beta, and gamma bands, and in the central region in theta and alpha bands. In contrast, delta band power at the temporal, parietal, and occipital regions has greater power in the non-athlete group.

Figure: Topographic scalp maps show the mean power difference (Athlete- Non-Athlete group) at each channel location. The warm tone colour shows greater band power in the athlete group, while the cool tone colour shows greater band power in the non-athlete group.

In observation of the classification accuracy achieved by different feature selection methods, the classifier trained by using frontal channels at both all 3 frequency bands and only alpha band stood out as the top two performers, consistently delivering the highest accuracy rates (99.11%, SD=0.008 and 99.25%, SD=0.007, respectively). While the classifier trained by using central, parietal channels, or all 3 regions at both all 3 frequency bands and only alpha band also provided high performance, they showed a slightly lower accuracy compared to the classifier trained by frontal channels. The classifier trained by using all 64 channels provided the lowest accuracy among all feature selection methods. This could imply that alpha band at frontal region play an important role in discriminating between these two groups.

Table: Classification performance for different feature selection methods

This project aimed to assess the frequency band power in different regions and apply those to a classification model discriminating between athlete and non-athlete groups. The classification performance results indicated that frontal alpha band power is crucial in classification as it obtained the highest accuracy of 99.25 %.

Surprisingly, athlete group had greater alpha power than non-athlete group. The greater alpha power indicates lower level of attention/concentration, which contradicts the hypothesis that athletes demonstrate higher levels of attention capabilities compared to non-athletes. However, the task responses showed that there is no difference in performance (correctness) between athlete and non-athlete groups. This can imply that due to the athletes’ specialized training and conditioning, they likely achieve the same performance as non-athletes with less attentional resources.

In summary, alpha band power can serve as a biomarker for attention discrimination between athletes and non-athletes. The application includes but is not limited to sports training purposes i.e., indicating athletes’ performance level or skills progression. Lastly, it would be interesting to see how other band power and/or brain indexes such as engagement index take part in this setting.

References

1. Pei, X., Xu, G., Zhou, Y. et al. A simultaneous electroencephalography and eye-tracking dataset in elite athletes during alertness and concentration tasks. Sci Data 9, 465 (2022). doi:10.1038/s41597–022–01575–0
2. Original Mackworth Clock Test (available at, https://www.psytoolkit.org/experiment-library/mackworth.html)

This article is part of a project presented to Brain Code Camp 2023, BrainCode101 Community, Thailand.

The presented results and discusssions are considered as a preliminary study and broad outlook to the area and have not yet gone through comprehensive scientific review.