Classification of EEG features for prediction of working memory load
The objective of this research was to compare classification methods aimed at predicting working memory (WM) load. Electroencephalogram (EEG) data was collected from physicians while performing basic WM tasks and simulated medical scenarios. Data processing was performed to remove noise from the signal used for analysis (e.g., muscle activity, eye-blinks). The data from basic WM tasks was used to develop and test the four classification models (LASSO regression, support vector machines (SVM), nearest shrunken centroids (NSC), and iterated supervised principal components (ISPC) to predict a WM state indicative of physicians’ optimal performance. The naïve misclassification rate was 19.74 %; LASSO and SVM outperformed this threshold: 18.10 and 12.21 % respectively). Both classification models had relatively high-specificity (LASSO: 97.2 %; SVM: 99.8 %); but relatively low-sensitivity LASSO: 20.7 %; SVM: 39.6 %). Results from simulated medical scenarios suggest that physicians were approximately 83 % of the time in the WM state that is likely indicative of optimal performance.
Abrantes, A., Comitz, E., & Mazur, L. (2016). Classification of EEG features for prediction of working memory load. In Advances in the human side of service engineering (pp. 115-126). ( Advances in Intelligent Systems and Computing ). Springer. DOI: https://doi.org/10.1007/978-3-319-41947-3_12