Using Multiple Neurophysiological Features to Construct the Predictive Models of Lower Limb Motor Preparation Onset

Motor preparation (MP) is a state of brain activity that occurs for a short period of time before an exercise is performed, and it has also been shown that this brain activity is associated with neurological rehabilitation. However, it is not easy to detect this phenomenon in the lower limbs using EEG, so this study aims to construct a prediction model using multiple combinations of feature extraction, feature selection, and classifiers, and to find the optimal combination of results. In this study, 34 healthy subjects were collected to construct and test the unilateral leg MP onset detection model. PSD, PCC, and CSP were used to extract the features and find the best model combination for each of them. PSD was found to perform best when using mRMR for feature selection and a soft-voting classifier containing multiple classifier algorithms, with MP onset accuracy of 81.14±1.11% and 81.01±1.37% for the left and right legs, respectively; PCC performs best when using LinearSVM without feature selection, with MP onset accuracies of 85.46±1.24% and 85.5±1.44% for the left and right legs, respectively; CSP performs best when using LDA with feature selection, with MP onset accuracy of 81.77±1.6% and 83.29±1.57% for the left and right legs, respectively. The accuracy of all three features was more than 80%, which means that these three neurophysiological features have their physiological significance. In addition, we also analyzed the subjects' motor preparation data by MRCP and found that MRCP tended to be more pronounced in the senior subjects when they were in motor preparing. Perhaps this can be studied in depth and used as an EEG indicator in the future to be extended to the clinic.