Validation of an Assistant System for Motion Analysis in Equipment-Based Exercise Therapy

Background: Due to demographic change, an increasing number of orthopedic diseases is expected. This complicates the care situation in rehabilitation. Therefore, movement execution cannot be adequately controlled by the therapist (Lösch et al., 2018). In exercise therapy, strength exercises are an important component. For example, exercises on the cable pulley make it possible to train activities of daily living, but because of the degrees of freedom, they have a high potential for incorrect movement execution. Sensor-based assistance systems may support therapists and patients, but for an effective use in exercise therapy (Vergbrugghe et al., 2018), an assistance system must reliably detect movement execution.

Aim: The aim of this study was to compare the fault detection of a developed assistance system with a traditional visual fault detection of an exercise therapy expert, to check the usability in exercise therapy.

Method: 14 older adults (69.4±4.4 years) completed the exercise hip abduction on a cable pulley. During the exercise execution, the movement was recorded by a marker and contactless sensor system and movement quality was analyzed. This was done by a red blue-green sensor (RGB; Kinect 1.0), which was rule-based trained on the following error patterns: bent knee – BK, tilting upper body – UB, wrong plane – WP, hip rotated outwards – HO. The following rules have been set for the error detection: BK: measured knee angle165°; UB: angle between shoulder-center and left ankle, with pivot hip-center160°; WP: distance of the right ankle to the correct movement plane380mm; HO: realized by the foot position. After that, the movement was visually assessed by an expert. For the comparison between the assistant system and the expert, a four-fold table was used. In this process, the evaluation of the error patterns within the repetitions was based on this. Cohens κ was used for analyzing degree of agreement occurring by chance.

Results: Sensitivity was defined as the degree of the matching error detection between the expert and the assistant system. For BK the sensitivity was 61.4% (Cohens κ=0.213; p=0.001), for UB 76.2% (Cohens κ=0.451; p0.001), for WP 75.7% (Cohens κ=0.022; p=0.395), for HO 25.4% (Cohens κ=-0.007; p=0.422).

Discussion and Conclusion: The present data indicate that there is an insufficient agreement between the expert and the assistant system. Furthermore, the error detection works well for BK and UP, but not as well for WP and HO. Possible reasons can be currently unimplemented filtering of detected error patterns within a repetition, the manual rule creation and faulty localized joint angles. This is due to the measuring principle of the RGB-sensor.