Editorial Commentary: Calling Balls and Strikes: For Artificial Intelligence Models to Accurately Predict Rotator Cuff Retear, We Must Carefully Define the Strike Zone.

Abstract

Early retear is a devastating complication following arthroscopic rotator cuff repair. The ability to predict early retear in patients undergoing rotator cuff repair may allow for interventions targeted at mitigating the risk of this complication, thereby improving surgical outcomes. Transfer learning, wherein labeled datasets are used to repurpose pretrained convolutional neural networks, represents an important approach for potentially improving the accuracy of these predictions through using artificial intelligence and deep learning. Leveraging deep learning provides an alternative approach to using tabular data for predictions by extracting important and subtle data from imaging (such as arthroscopic photographs) to provide insight into how the appearance of a surgical construct may predict outcomes. However, the value of such pipelines depends highly on preprocessing and generalizability of the training datasets. When labeling datasets, characteristics such as tendon integrity may be partially subjective, so surgeon-versus-surgeon (interrater) reliability or surgeon-versus-model analysis is crucial towards reducing bias. Clinical adoption of these pipelines requires generalizability of models, and convolutional neural networks trained on single surgeon datasets introduces bias variability. Clinical adoption requires trust and fairness as well as governance by artificial intelligence regulatory committees or "umpires;" and, if we are to trust DL models to call balls and strikes with respect to rotator cuff integrity in the real-world, it is important to have a clearly defined strike-zone.