Brain-computer interface (BCI) has been widely used to predict the intention of users in motor imagery-based (MI-based) task. Although the overall MI classification accuracy has been largely enhanced from previous efforts, applying MI-BCI to the so-called BCI-illiterate subjects remains as an unsolved problem. This study proposed a physiological approach for improving MI-BCI performance, by measuring the baseline attention level estimated by coefficient F from the electroencephalogram (EEG) band-activities. In this endeavor, a total of 9 MI-EEG recordings were retrieved from an open BCI dataset. A measure of attention level was calculated for each trial to select high attention trials. High attention trial-based machine learning model showed higher MI classification performance (median accuracy = 62.50% (interquartile range (IQR) = 55.21-82.29%)) than the conventional approach (median accuracy = 57.64% (IQR = 54.17-62.50%)) with statistical significance (Wilcoxon rank sum test, p = 0.037). This study found that machine learning models trained from high attention trials yield improved classification accuracy to the models derived from total trial regardless of both BCI illiterate and literate.