12.3.5 - Identifying Subscapularis Tendon Tears: Machine Learning Can Improve Diagnostic Accuracy
- A. Pareek (Rochester, US)
- A. Pareek (Rochester, US)
- M. Nieboer (Rochester, US)
- D. Kilinc (Rochester, US)
- N. Rhodes (Rochester, US)
- C. Tiegs-Heigen (Rochester, US)
- K. Pasupathy (Rochester, US)
- C. Camp (Rochester, US)
- J. Sanchez-Sotelo (Rochester, US)
Abstract
Purpose
The purpose of this study was to use machine learning to develop a predictive model of subscapularis tears and delineate findings most helpful in identifying subscapularis tears preoperatively.
Methods and Materials
Between 2010 and 2020, 202 consecutive shoulders (57% male, mean age 58.8 years) underwent arthroscopic cuff repair by a single surgeon with particular attention to the arthroscopic diagnosis of subscapularis tears. Overall, 60% of patients had intraoperatively verified partial or full thickness tears of the subscapularis (50% of these were upper third). All radiographs and MRIs were reviewed independently in a blinded manner by two board-certified musculoskeletal radiologists. Patient demographics, physical exam, and imaging (tear, biceps status, fatty atrophy, and other) were included for model creation. An eXtreme Gradient Boosting (XGBoost) algorithm was selected due to its predictive power and efficiency. We used nested cross-validation to optimize hyperparameters with 10-fold validation.
Results
The first model was built using only imaging parameters and second model was built using imaging in addition to patient and physical exam parameters. The top five features included metrics related presence of tearing as evidenced on MRI (Figure 1). The overall area under the curve (AUC) of the model was 84.5% (Figure 2), which was deemed to be a very good model. The model’s accuracy did not increase with the addition of patient or physical exam parameters. In addition, isolating the model to only the top five features maintained the AUC of the model to 84%.
Conclusion
Machine learning was successful in predicting subscapularis tears in close to 85% of the patients and this accuracy did not decrease by isolating the model to the top features. The addition of physical exam or patient characteristics does not make a significant difference in the predictive ability of this particular model.