Effect of machine learning prediction on surgical decision making for shoulder arthroplasty: a multi-surgeon study

Abstract

Background Machine learning algorithms for surgical decision making in shoulder arthroplasty has not been reported. Though there are recommendations based on available literature regarding the selection of anatomic versus reverse shoulder replacement, there are no clear guidelines on how this decision should be made. Our aim was to assess the viability of machine learning for this application by evaluating the agreement of the algorithm’s recommendation on type of arthroplasty versus that recommended by six shoulder surgeons. Methods There were 84 cases of patients with glenohumeral osteoarthritis planned using a three-dimensional CT-based software. Half of the cases were planned without any ML algorithm-based recommendation available, and half were planned with the recommendation available, and kappa coefficients were calculated to determine agreement. Results In 78% of cases, the software’s surgical recommendation on arthroplasty type completely aligned with that of the surgeons. The Cohen’s kappa coefficients for surgeons’ versus software’s recommendations were 0.56 and 0.61 for rounds 1 and 2, respectively, while the Fleiss kappa coefficients (inter-surgeon agreement) were 0.87 and 0.77. Discussion Machine learning for the application of guiding surgeons on which type of shoulder arthroplasty to select has demonstrated viability in this study, with further research needed to refine this system.