Multi-view Clustering with mvReliefF for Parkinson's Disease Patients Subgroup Detection

摘要

Parkinson's disease is a chronic neurodegenerative disease affecting people worldwide. Parkinson's disease patients experience motor symptoms and many other symptoms that affect the quality of their lives. Discovering groups of patients with similar symptoms from different symptom groups can improve the understanding of this incurable disease and advance the development of personalized treatment of Parkinson's disease patients. This paper proposes a multi-view clustering approach to discover groups of patients experiencing similar symptoms from different symptom groups (views). For that we modified ReliefF feature ranking algorithm to characterize subsets of most informative symptoms that maximize the similarity between the detected patient groups, described by symptoms from different views (i.e. different symptom groups). The adapted mvReliefF algorithm calculates the weight of features based on the values of their neighbors over multiple views. The current approach works for two views simultaneously, but can be extended to multiple views. The results of the experiments show that the proposed methodology, applied on a pair of data sets from the PPMI data collection, successfully identified lists of most important symptoms that divide patients into groups, ordered by the severity of patients' symptoms.