Importance of the Features of Event-Related Potentials Used for a Machine Learning-Based Model Applied to...

In this study, a method for assessing the human state and brain-machine interface (BMI) has been developed using event-related potentials (ERPs). Most of these algorithms are classified based on the ERP characteristics. To observe the characteristics of ERPs, an averaging method using electroencephalography (EEG) signals cut out by time-locking to the event for each condition is required. To date, several classification methods using only single-trial EEG signals have been studied. In some cases, the machine learning models were used for the classifications; however, the relationship between the constructed model and the characteristics of ERPs remains unclear. In this study, the LightGBM model was constructed for each individual to classify a single-trial waveform and visualize the relationship between these features and the characteristics of ERPs. The features used in the model were the average values and standard deviation of the EEG amplitude with a time width of 10 ms. The best area under the curve (AUC) score was 0.92, but, in some cases, the AUC scores were low. Large individual differences in AUC scores were observed. In each case, on checking the importance of the features, high importance was shown at the 10-ms time width section, where a large difference was observed in ERP waveforms between the target and the non-target. Since the model constructed in this study was found to reflect the characteristics of ERP, as the next step, we would like to try to improve the discrimination performance by using stimuli that the participants can concentrate on with interest.