EEG-based user performance prediction using random forest in a dynamic learning environment

Itelligent tutoring systems (ITS) can be improved by considering real time information from phsysiological signals. For example, performance could be predicted based on the user’s affective state. In this chapter a feature extraction approach is presented to try to predict when a user makes a mistake while interacting in a dynamic learning environment (DLE). Electroencephalogram (EEG) signals from participants were recorded and used as inputs in a random forest. Three approaches were followed: in the first one we used the affective states provided by a commercial headset, in the second approach we additionally considered the affective state rate of change and finally we used the 14 channels raw information generated by the EEG after applyinga Fast Fourier Transformation (FFT) to calculate the Power Spectral Density (PSD). Results show that random forest provides high accuracy in the three approaches that were followed. Further, key results show an extensive analysis of feature importance which helped identify the relevant affective states, bandwidths and sensors. Lessons learned can be incorporated into the desing of intelligent tutoring sytems.