Multiple Regression Model for Cognitive Function Evaluation Using P300 Based Spelling-Brain–Computer Interface

Abstract

Early diagnosis is important in the treatment of dementia; however, many dementia patients are resist seeking medical attention. In our laboratory, we are developing a dementia-screening device using a P300 based spelling-brain–computer interface (spelling-BCI) for early diagnosis of dementia. We believe that by estimating the results of neuropsychological examinations through measurements using the spelling-BCI, it would be possible to realize tests similar to those performed by medical specialists. According to a previous study, a multiple regression equation with Mini-Mental State Examination (MMSE), the Japanese version of the Montreal Cognitive Assessment (MoCA-J), and Frontal Assessment Battery (FAB) scores are objective variables, and the features obtained using the BCI and age are explanatory variables. In multiple regression analysis, variable selection was performed via the forward–backward stepwise selection method and the data exceeding the 95% confidence interval of the estimation error were excluded. As a result, the measurement data were removed using 95% confidence intervals for the estimation errors so that the multiple regression equation was identified using only about 30 of the more than 200 measurement data. In this study, we evaluated the estimation model of neuropsychological examinations using errors and investigated the effect of excluding outliers. As a result, in the estimation model of each neuropsychological examination, there was little change in the root mean squared errors before and after outlier removal. Therefore, the effect of outlier removal on the estimation model was insignificant.