An evaluation of clustering techniques to classify dexterous manipulation of individuals with and without dysfunction

Abstract

The rehabilitation of manipulation ability in orthopedic (e.g., thumb carpometacarpal osteoarthritis-CMC OA) and neurological (e.g., Parkinson's disease-PD) conditions depends critically on our ability to detect dysfunction and quantify its evolution and response to treatment. The Strength-Dexterity (SD) test is a validated indicator of dynamic dexterous manipulation function, but its ability to categorize clinical populations has not been tested. We 1) used the SD test to compare manipulation ability among patients with OA and PD and healthy age-matched elderly control subjects; and 2) compared and evaluated the ability of different clustering techniques to classify subjects into clinical or control groups and calculate their respective cluster centroids. We considered five clustering methods (three hard and two fuzzy): K-means, K-medoids, Gaussian expectation-maximization (GEM), Subtractive, and Fuzzy C-means clustering. We found the centroids of the SD test scores differed significantly between the clinical and control groups. Of the five methods considered, the GEM clustering algorithm most accurately classified SD test performance between these two groups.