Abstract
BackgroundThe Wolf Motor Function Test (WMFT) is a well-recognized measure for assessing upper extremity motor function in stroke rehabilitation. However, prolonged administration time limits the WMFT in clinical use.ObjectiveThis study aimed to reduce the number of WMFT tasks using machine learning and explore its measurement structure and psychometric properties, using data from 3 stroke rehabilitation trials that together engaged 543 participants with a wide range of motor impairment during subacute and chronic recovery phases.MethodsWMFT performance time data were converted to rates and outliers were eliminated using multivariate normality tests. Random forest regression with the elbow method was employed to determine the optimal number of items in the WMFT. Further, a machine learning technique with cross-validation and bootstrapping was used to select items. We used confirmatory factor analysis to determine the measurement structure of the original and shortened version of WMFT. Psychometric properties of the shortened version were also assessed.ResultsMachine learning-based item reduction identified 4 items (Hand to Table, Hand to Box, Extend Elbow Without Weight, and Lift Can) as representative tasks. Factor analysis revealed a 2-factor structure for both original and shorten versions, comprising non-manipulative/transport and manipulative/dexterity factors. WMFT-4 showed strong convergent validity with WMFT-15 (R = 0.98, P < .001) and moderate cross-domain validity with the Fugl-Meyer Assessment of Upper Extremity (FMA-UE) (R = 0.523, P < .001), comparable to the original WMFT-15 (R = 0.526, P < .001).ConclusionThe streamlined WMFT-4 enhances the feasibility of the WMFT for both clinical and research settings while maintaining its original measurement characteristics.