Researching the best feature vector cluster selection for eye gaze prediction using EOG data

Many studies have been carried out on machine learning using electrooculography (EOG) to facilitate the living conditions of people with disabilities. In this study, it is aimed to compare the effect of different feature selection methods on estimation errors by using various data engineering techniques while estimating eye angular displacements in EOG data. After filtering by removing the blinks in the vertical and horizontal components of the EOG signals in the data set, both channels were processed separately and 23 (root mean square, mean, absolute maximum, kurtosis, skewness, coefficient of variation, geometric mean, harmonic mean, variance, median, average energy, standart deviation, standart error, shape factor, interquatile range, %50 trimmed mean, %25 trimmed mean, maximum, mean absolute deviation, minimum, central moment, singular value decomposition and crest) features were extracted in the time domain, 10 (dominant power, standart deviation, energy, power, energy entropy, kurtosis, skewness, median and mean) features in the frequency domain, and 36 features (activity, mobility, complexity, kurtosis, skewness and energy) in the time-frequency domain. Discrete fourier transform is used in the frequency domain and discrete wavelet transform is used in the time-frequency domain. By using the obtained data holdout and k-fold methods, features were selected with Fsrtest, Fsrnca, Fsrmrmr, RReliefF and Sequentials feature methods. The selected features were inserted into the machine learning algorithms Gaussian Process Regression(GPR), Support Vector Machine Regression(SVR), Regression Tree and Regression Tree Ensembles. The lowest error rates obtained, individually, in the horizontal EOG component, the RMSE and angular error values in the RTE algorithm with the rerelieff feature selection method in the time-frequency plane with the best results are 0.96, 0.77, respectively. In the vertical EOG component, on the other hand, the RMSE and angular error values in the RTE algorithm are 1.73 and 1.35, respectively, with the feature selection method fsrnca in the time domain. In the mean values, the best results were obtained in the RTE algorithm, with the RMSE and angular error values in the horizontal EOG component in the time-frequency plane with 2.15±0.73, 1.39±0.41 fsrnca feature selection method, respectively. In the vertical EOG component, the RMSE and angular error values in the time-frequency plane were obtained with the fsrtest feature selection method, with 3.67±0.87, 2.61±0.64, respectively, in the GPR algorithm.