Machine learning algorithms for predicting scapular kinematics

Abstract The goal of this study was to develop and validate a non-invasive approach to estimate scapular kinematics in individual patients. We hypothesized that machine learning algorithms could be developed using motion capture data to accurately estimate dynamic scapula orientation based on measured humeral orientations and acromion process positions. The accuracy of the algorithms was evaluated against a gold standard of biplane fluoroscopy using a 2D to 3D fluoroscopy/model matching process. Individualized neural networks were developed for nine healthy adult shoulders. These models were used to predict scapulothoracic kinematics, and the predicted kinematics were compared to kinematics obtained using biplane fluoroscopy to determine the accuracy of the machine learning algorithms. Results showed correlations between predicted kinematics and validation kinematics. Estimated kinematics were within 10 of validation kinematics. We concluded that individualized machine learning algorithms show promise for providing accurate, non-invasive measurements of scapulothoracic kinematics.