Machine Learning Does Not Improve Humeral Torsion Prediction Compared to Regression in Baseball Pitchers

Background Humeral torsion is an important osseous adaptation in throwing athletes that can contribute to arm injuries. Currently there are no cheap and easy to use clinical tools to measure humeral torsion, inhibiting clinical assessment. Models with low error and “good” calibration slope may be helpful for prediction. Hypothesis/Purpose To develop prediction models using a range of machine learning methods to predict humeral torsion in professional baseball pitchers and compare these models to a previously developed regression-based prediction model. Study Design Prospective cohort Methods An eleven-year professional baseball cohort was recruited from 2009-2019. Age, arm dominance, injury history, and continent of origin were collected as well as preseason shoulder external and internal rotation, horizontal adduction passive range of motion, and humeral torsion were collected each season. Regression and machine learning models were developed to predict humeral torsion followed by internal validation with 10-fold cross validation. Root mean square error (RMSE), which is reported in degrees (°) and calibration slope (agreement of predicted and actual outcome; best = 1.00) were assessed. Results Four hundred and seven pitchers (Age: 23.2 +/-2.4 years, body mass index: 25.1 +/-2.3 km/m2, Left-Handed: 17%) participated. Regression model RMSE was 12° and calibration was 1.00 (95% CI: 0.94, 1.06). Random Forest RMSE was 9° and calibration was 1.33 (95% CI: 1.29, 1.37). Gradient boosting machine RMSE was 9° and calibration was 1.09 (95% CI: 1.04, 1.14). Support vector machine RMSE was 10° and calibration was 1.13 (95% CI: 1.08, 1.18). Artificial neural network RMSE was 15° and calibration was 1.03 (95% CI: 0.97, 1.09). Conclusion This is the first study to show that machine learning models do not improve baseball humeral torsion prediction compared to a traditional regression model. While machine learning models demonstrated improved RMSE compared to the regression, the machine learning models displayed poorer calibration compared to regression. Based on these results it is recommended to use a simple equation from a statistical model which can be quickly and efficiently integrated within a clinical setting. Levels of Evidence 2