Pitch-classifier model for professional pitchers utilizing 3D motion capture and machine learning algorithms

Joseph E Manzi; Brittany Dowling; Spencer Krichevsky; Nicholas L S Roberts; Suleiman Y Sudah; Jay Moran; Frank R Chen; Theodore Quan; Kyle W Morse; Joshua S Dines

doi:10.1016/j.jor.2023.12.007

Pitch-classifier model for professional pitchers utilizing 3D motion capture and machine learning algorithms

J Orthop. 2023 Dec 20:49:140-147. doi: 10.1016/j.jor.2023.12.007. eCollection 2024 Mar.

Authors

Affiliations

¹ Department of Orthopaedic Surgery, Northwell Health, New York, NY, USA.
² Sports Performance Center, Midwest Orthopaedics at Rush, Chicago, IL, USA.
³ Stony Brook University, Department of Biomedical Informatics, Stony Brook, NY, USA.
⁴ Department of Orthopaedic Surgery, Monmouth Medical Center, Monmouth, NJ, USA.
⁵ Yale School of Medicine, New Haven, CT, USA.
⁶ Department of Anesthesiology, Hospital of the University of Pennsylvania, Philadelphia, PA, USA.
⁷ George Washington University School of Medicine, Washington, DC 20037, USA.
⁸ Sports Medicine Institute Hospital for Special Surgery, New York, NY, USA.

PMID: 38682007
PMCID: PMC11043625 (available on 2025-03-01)
DOI: 10.1016/j.jor.2023.12.007

Abstract

Introduction: A pitcher's ability to achieve pitch location precision after a complex series of motions is of paramount importance. Kinematics have been used in analyzing performance benefits like ball velocity, as well as injury risk profile; however, prior utilization of such data for pitch location metrics is limited.

Objective: To develop a pitch classifier model utilizing machine learning algorithms to explore the potential relationships between kinematic variables and a pitcher's ability to throw a strike or ball.

Methods: This was a descriptive laboratory study involving professional baseball pitchers (n = 318) performing pitching tests under the setting of 3D motion-capture (480 Hz). Main outcome measures included accuracy, sensitivity, specificity, F1 score, positive predictive value (PPV), and negative predictive value (NPV) of the random forest model.

Results: The optimized random forest model resulted in an accuracy of 70.0 %, sensitivity of 70.3 %, specificity of 48.5 %, F1 equal to 80.6 %, PPV of 94.3 %, and a NPV of 11.6 %. Classification accuracy for predicting strikes and balls achieved an area under the curve of 0.67. Kinematics that derived the highest % increase in mean square error included: trunk flexion excursion(4.06 %), pelvis obliquity at foot contact(4.03 %), and trunk rotation at hand separation(3.94 %). Pitchers who threw strikes had significantly less trunk rotation at hand separation(p = 0.004) and less trunk flexion at ball release(p = 0.003) compared to balls. The positive predictive value for determining a strike was within an acceptable range, while the negative predictive value suggests if a pitch was determined as a ball, the model was not adequate in its prediction.

Conclusions: Kinematic measures of pelvis and trunk were crucial determinants for the pitch classifier sequence, suggesting pitcher kinematics at the proximal body segments may be useful in determining final pitch location.

Keywords: Kinematics; Motion-capture; Pitch location; Predictive model; Strike zone.