Is Notable EEG Feature Extracted Over Time-Dependent Cognitive Load Variation During Intelligence Tests?

Abstract

In this study, we collected electroencephalography (EEG) data during an intelligence test to understand the tendency of neurophysiologic change according to task difficulty or mental fatigue. Four healthy subjects were recruited for the study. Subjects solved problem sets (Raven’s APM Set II problems) without a time limit in random order for measuring their intellectual quotient level. We measured EEG activity as participants performed the task. We used spectral power as a feature and introduced XGBoost as the predictor of cognitive load. When we trained the network of XGBoost using the feature of EEG labeled in problem order (ordered by difficulty), the root mean squared error (RMSE) from the test data was significantly larger (12.5 ± 1.3) than the same measure from a regressor trained by a feature aligned by time (9.6 ± 1.4, p < 0.001 from unpaired student’s t-test). Moreover, we found a stronger correlation from the prediction result of a time-dependent feature learning network (0.7 ± 0.1) compared to the prediction of a difficulty-dependent feature learning network (0.39 ± 0.1, p < 0.001 from unpaired students’ t-test). In summary, we found a better predictive performance in networks trained with time-dependent features compared to networks with difficulty-dependent features. We propose that these results may explain EEG feature variability biased by mental state changes during intellectual tasks.