Making Sense of Shoulder Exercise: Measuring the Accuracy of an Artificial Intelligence Model to Classify Shoulder...

This study aimed to compare the accuracy of machine learning classification for three commonly prescribed shoulder exercises in people with and without rotator cuff tendinopathy. Eighteen participants with rotator cuff tendinopathy (mean age 54.2, SD 13.2; 50% female), followed by eighteen matched controls completed a laboratory-based shoulder strength testing protocol. Three exercises were performed (shoulder press, lateral raise and bent over row) while wearing three inertial measurement (IMU) sensors (Axivity, Ax6 - 3 axis accelerometry and gyroscope at 100 Hz and 1000°/sec respectively) positioned on the wrist arm and trunk. Data were analysed and accuracy was compared between common machine learning algorithms for those with rotator cuff tendinopathy and healthy matched controls in a subject-dependent and subject-independent model. The best accuracy scores for the subject-dependent results were achieved by a random forest algorithm; 96.12% (3-sensor combined system) for those with rotator cuff tendinopathy. For the subject-independent results best accuracy scores were achieved by a convolutional neural network algorithm; 94.55% (3-sensors) for the healthy controls without shoulder pain. K-fold cross validation confusion matrix results by exercise type for the entire cohort show 97% accuracy (shoulder press), 95.5% (lateral raise) and 90.7% (bent over row) (3-sensors, CNN subject-independent analysis). Machine learning classification of 3 different shoulder exercises in people with rotator cuff tendinopathy and matched healthy controls demonstrate most accurate results using a CNN algorithm for subject-independent analysis and a RF algorithm for subject-dependent analysis. Results were similar for both those with rotator cuff tendinopathy and their matched healthy controls.