Pitch-Tracking Metrics as a Predictor of Future Shoulder and Elbow Injuries in Major League Baseball Pitchers: A Machine-Learning and Game-Theory Based Analysis

Abstract

Background: Understanding interactions between multiple risk factors for shoulder and elbow injuries in Major League Baseball (MLB) pitchers is important to identify potential avenues by which risk can be reduced while minimizing impact on player performance.

Purpose: To apply a novel game theory-based approach to develop a machine-learning model predictive of next-season shoulder and elbow injuries in MLB pitchers and use this model to understand interdependencies and interaction effects between the most important risk factors.

Study design: Case-control study; Level of evidence, 3.

Methods: Pitcher demographics, workload measures, pitch-tracking metrics, and injury data between 2017 and 2022 were used to construct a database of MLB pitcher-years, where each item in the database corresponded to a pitcher's information and metrics for that year. An extreme gradient boosting machine-learning model was trained to predict next-season shoulder and elbow injuries utilizing Shapley additive explanation values to quantify feature importance as well as interdependencies and interaction effects between predictive variables.

Results: A total of 3808 pitcher-years were included in this analysis; 606 (15.9%) of these involved a shoulder or elbow injury resulting in placement on the MLB injured list. Of the >65 candidate features (including workload, demographic, and pitch-tracking metrics), the most important contributors to predicting shoulder/elbow injury were increased: pitch velocity (all pitch types), utilization of sliders (SLs), fastball (FB) spin rate, FB horizontal movement, and player age. The strongest game theory interaction effects were that higher FB velocity did not alter a younger pitcher's predicted risk of shoulder/elbow injury versus older pitchers, risk of shoulder/elbow injury increased with the number of high-velocity pitches thrown (regardless of pitch type and in an additive fashion), and FB velocity <95 mph (<152.9 kph) demonstrated strong negative interaction effects with higher SL percentage, suggesting that the overall predicted risk of injury for pitchers throwing a high number of SLs could be attenuated by lower FB velocity.

Conclusion: Pitch-tracking metrics were substantially more predictive of future injury than player demographics and workload metrics. There were many significant game theory interdependencies of injury risk. Notably, the increased risk of injury that was conferred by throwing with a high velocity was even further magnified if the pitchers were also older, threw a high percentage of SLs, and/or threw a greater number of pitches.

Keywords: Major League Baseball; artificial intelligence; machine learning; pitching metrics; shoulder and elbow injuries.

Figure 1.

Mean feature importance rankings as determined using 3 separate algorithms: recursive feature elimination with a random forest algorithm, linear model coefficient ranking with ridge regression, and Gini importance (mean decrease in impurity) of a random forest. Selected (estimated best) features are assigned rank 1, while nonimportant features have ranks close to 0. FB, fastball; CB, curveball; CH, changeup; CT, cutter; SI, sinker; SL, slider; SP, splitter; %, frequency; v, velocity; -X, horizontal movement; -Z, vertical movement.

Figure 2.

Performance of the XGBoost model created with only the top 10 features on internal validation using cross-validation. The shaded area indicates the standard deviation. The area under the receiver operating characteristic curve (AUC) decreased from 0.66 to 0.61.

Figure 3.

Feature importance (Shapley additive explanation [SHAP] plots) for the XGBoost model showing the relative contribution of each feature to model predictions. Each point represents a single pitcher on which the probability of experiencing shoulder or elbow injury was predicted. The y-axis lists variables selected into the model in order of importance from top to bottom. SHAP values are listed on the x-axis and correspond to the change in log-odds attributed to a feature's value (and hence the change in likelihood of experiencing injury). A higher absolute value indicates greater importance for generating a prediction. Gradient colors correspond to the original value for a particular feature. Positive SHAP values contribute to increased likelihood of experiencing a shoulder or elbow injury, while negative SHAP values correspond to decreased likelihood. FB, fastball; CB, curveball; CH, changeup; CT, cutter; SI, sinker; SL, slider; SP, splitter; %, frequency; v, velocity; -X, horizontal movement; -Z, vertical movement.

Figure 4.

Shapley additive explanation (SHAP) dependence plots demonstrating how the importance of multiple predictive features varies depending on a pitcher's fastball velocity (vFB): (A) age, (B) curveball velocity (vCB), and (C) changeup velocity (vCH). Interaction effects between each variable are displayed as vertical dispersion. Each dot represents a single pitcher in the study.

Figure 5.

Shapley additive explanation (SHAP) dependence plot showing the effect of age on the impact of fastball velocity. Interaction effects between each variable are displayed as vertical dispersion. Each dot represents a single pitcher in the study. vFB, fastball velocity.

Figure 6.

Shapley additive explanation (SHAP) interaction effects between important predictors of shoulder and elbow injuries in pitchers. Variable interaction effects are quantified by SHAP values and can shift the model output beyond what the individual risk factor would predict alone. vFB, fastball velocity; SL%, slider frequency.