Consistency verification and interpretation of explainable AI for predicting annual home runs of professional baseball players from sensor data

Abstract

Abstract This study aimed to verify and interpret a model for predicting the number of home runs per year using sensor data from professional baseball players during batting practice. A machine learning model was constructed using Random Forest from the bat kinematics and bat mass data of 41 professional baseball players collected by a bat-mounted sensor. Partial Dependence analysis and Feature Importance analysis by SHAP (SHapley Additive exPlanations) were used to explain the model’s predictions. The predictive model showed that the bat speed, bat mass, and rotational acceleration are particularly important. The results indicated that a bat speed of 33.3 m/s and rotational acceleration exceeding 157 m/s2 exhibited a trend toward a rapid increase in the number of predicted home runs per year. The mass of the bat suggests that an optimum value exists at 0.91 kg. These results suggest that batters who are expected to hit a large number of home runs each year increase the acceleration at the beginning of their swing to produce high bat speed in a short period of time and achieve bat speeds of 33.3 m/s or more with a bat that is somewhat heavier.