Objective classification and scoring of movement deficiencies in patients with anterior cruciate ligament reconstruction

Abstract

Motion analysis systems are widely employed to identify movement deficiencies-e.g. patterns that potentially increase the risk of injury or inhibit performance. However, findings across studies are often conflicting in respect to what a movement deficiency is or the magnitude of association to a specific injury. This study tests the information content within movement data using a data driven framework that was taught to classify movement data into the classes: NORM, ACLOP and ACLNO OP, without the input of expert knowledge. The NORM class was presented by 62 subjects (124 NORM limbs), while 156 subjects with ACL reconstruction represented the ACLOP and ACLNO OP class (156 limbs each class). Movement data from jumping, hopping and change of direction exercises were examined, using a variety of machine learning techniques. A stratified shuffle split cross-validation was used to obtain a measure of expected accuracy for each step within the analysis. Classification accuracies (from best performing classifiers) ranged from 52 to 81%, using up to 5 features. The exercise with the highest classification accuracy was the double leg drop jump (DLDJ; 81%), the highest classification accuracy when considering only the NORM class was observed in the single leg hop (81%), while the DLDJ demonstrated the highest classification accuracy when considering only for the ACLOP and ACLNO OP class (84%). These classification accuracies demonstrate that biomechanical data contains valuable information and that it is possible to differentiate normal from rehabilitating movement patterns. Further, findings highlight that a few features contain most of the information, that it is important to seek to understand what a classification model has learned, that symmetry measures are important, that exercises capture different qualities and that not all subjects within a normative cohort utilise 'true' normative movement patterns (only 27 to 71%).

Fig 1. Illustration of the work-flow during the features engineering.

Fig 2. Illustration of the workflow to identify the most suitable learning technique.

Fig 3. Illustration of the workflow to identify key features toward classification.

Fig 4. Illustration of the results of the SLHop model: The model accuracy is displayed on the top right, while the confusion matrix of a 2-feature model is displayed below (right button).

The three graphs on the left display the confusion pattern (button) and selection frequency of each trail within the ACL^OP, ACL^{NO OP} and NORM (top).

Fig 5. Illustration of the 2 most important SLDJ model features (CoM resultant velocity at take off values and vertical CoM velocity at impact) plotted against each other on the left and the CoM velocity at impact for the logic cases on the right.

Fully coloured bubbles represent trials with ‘true’ group pattern, while faded coloured bubbles represent trails that did not.