Severity level diagnosis of Parkinson’s disease by ensemble K-nearest neighbor under imbalanced data

The reliable and explainable diagnosis of severity level for Parkinson’s disease (PD) is significant for the therapy. Nonetheless, there are little data for severe PD patients but abundant data for slight PD patients, and this imbalanced distribution reduces the accuracy of diagnosis. Besides, the intrinsic differences for different severity levels are still unclear due to the individual differences and similarity of gait. To figure out the gait differences toward the development of PD severity level, gait features like time and force features as well as their coefficient of variance and asymmetry index have been extracted and compared. To overcome the imbalance influence during the severity level diagnosis, an ensemble K-nearest neighbor (EnKNN) is proposed. The K-nearest neighbor algorithm is applied to construct the base classifiers with extracted features, then the weight of each base classifier is calculated by the G-mean score and the F-measure. Finally, base classifiers are integrated by weight voting. Results show that the proposed EnKNN can achieve an average accuracy of 95.02% (0.44%) for PD severity level diagnosis overwhelming the imbalanced distribution of data. Additionally, some gait features exhibit distinct change with the increase of PD severity level which helps to a reliable and explainable diagnosis.