Infrared depth sensor based automated classification of motor dysfunction in multiple sclerosis – a proof-of-concept study

Background: The Expanded Disability Status Scale (EDSS) is the most frequently used scale to rate disability in multiple sclerosis (MS) and an important outcome measure in clinical trials and daily care. However, it has a relatively low inter-rater and intra-rater reliability. To improve assessment of disability in MS patients, we developed an automated image analysis algorithm using advanced machine learning techniques, which classified movements recorded non-invasively with an infrared depth sensor (Kinect).

Objectives: To test classification performance of the automated image analysis algorithm on depth sensor recordings of six defined movements. To understand the dominant features in the movements identified by the automated image analysis algorithm in a medical context.

Methods: Six pre-defined movements from standardized neurological assessments, covering upper and lower extremities and trunk, as well as a movement typical of activities of daily living (drinking from a cup), were recorded in 72 MS patients and 102 healthy volunteers. For all patients a standardized EDSS-assessment, Nine-Hole-Peg-Test, and the Symbol Digit Modalities Test were performed and documented. Movement information was extracted directly from the depth sensor recordings, pre-processed and analysed by the bespoke automated image analysis algorithm. The built-in skeleton algorithm was not used. We evaluated the outcome with Dice scores, conservative estimates for classification accuracy that penalize misclassification.

Results: We were able to quantify the motor dysfunction of 72 MS patients in six pre-defined and recorded movements that cover the upper body, trunk, and lower body. Each movement was correctly classified with a Dice score of at least 80%, supporting the validity of the approach. Average sensitivity/specificity scores were 80-90%. The bespoke automated image analysis algorithm identified movement properties with highest discriminative power, which may help facilitate training of neurologists. Conclusions: The infrared depth sensor recording technique with bespoke automated image analysis algorithm enables a reliable and sensitive quantitative assessment of motor dysfunction in MS patients. With a finer classification to capture different degrees of movement disability in future, this approach may improve the evaluation of disability and disease progression in MS.