Prediction of Pilot's Reaction Time Based on EEG Signals

The main hypothesis of this work is that the time of delay in reaction to an unexpected event can be predicted on the basis of the brain activity recorded prior to that event. Such mental activity can be represented by electroencephalographic data. To test this hypothesis, we conducted a novel experiment involving 19 participants that took part in a 2-h long session of simulated aircraft flights. An EEG signal processing pipeline is proposed that consists of signal preprocessing, extracting bandpass features, and using regression to predict the reaction times. The prediction algorithms that are used in this study are the Least Absolute Shrinkage Operator and its Least Angle Regression modification, as well as Kernel Ridge and Radial Basis Support Vector Machine regression. The average Mean Absolute Error obtained across the 19 subjects was 114 ms. The present study demonstrates, for the first time, that it is possible to predict reaction times on the basis of EEG data. The presented solution can serve as a foundation for a system that can, in the future, increase the safety of air traffic.

it to unit variance. This procedure is then applied to each feature vector independently by computing the 24 relevant statistics on the basis of samples from the training set. Mean and standard deviation are then stored 25 for use on an independent test data. 26 If we denote the mean value of training samples from feature f by µ(X f ) and their standard deviation as 27 σ(X f ), then for a sample x f i (i = 1, . . . , M f , M f -is a number of samples of feature f ), the described 28 transformation can be calculated using the following equation: In Eq. 2, p(x f , y) is the joint probability function of X f and Y , and p(x f ) and p(y) are the marginal 40 probability distribution functions of X f and Y respectively. In this study, x f is a vector of 10 bandpower 41 features obtained for a single event, while y is the corresponding time of delay in reaction to that event.

42
The MI criterion is known for being capable of capturing any kind of dependency between variables.

43
Use of MI-based feature selection methods have been proven to yield highly satisfactory results in many 44 approaches to EEG signal processing (Binias et al., 2016(Binias et al., , 2018.

46
F-test statistics can be used as a criterion for ranking features. This approach utilizes univariate linear regression for testing the individual effect of the regression variables. To extract this information, the first 48 step requires that the correlation between the vector of regressors X f ∈ R M f and the vector of targets 49 Y ∈ R M f is computed, according to the following equation: The R 2 f is then converted to an F-score to obtain the final result. If we denote the number of observations 51 as M f and the degrees of freedom as p f , then the relation between the F-score F f and R 2 f is expressed as It must be noted that the F-test expresses only a linear dependency between variables. In this study, x f is 54 a vector of 10 bandpower features obtained for a single event, while y is the corresponding time of delay in 55 reaction to that event.

90
In this study, a Radial Basis Function (RBF) was used for kernel transformation. The RBF for a feature 91 vector X f ∈ R M is defined as presented in Eq.7.

101
• e is the vector of all ones,

102
• C > 0 is the upper bound,

104
• Q ij ≡ K(X i , X j ) = φ(X i ) T φ(X j ) is the kernel function.