Early Detection of Injuries in MLB Pitchers From Video

A. J. Piergiovanni, Michael S. Ryoo

Abstract

Injuries represent a significant financial and competitive burden in professional sports. In Major League Baseball (MLB), pitcher injuries alone account for hundreds of millions of dollars in annual costs. This paper introduces a novel approach to injury detection and prediction using standard broadcast video. By employing 3D convolutional neural networks (I3D) trained on optical flow data, we evaluate the ability to identify subtle mechanical shifts that precede an injury. Our experiments on a dataset of 20 injured pitchers from the 2017 season demonstrate that the model can successfully distinguish between "healthy" and "injured" states, with peak predictive power appearing 10 to 30 pitches before a player is officially sidelined.

Introduction

Professional pitchers are masters of repeatability, making the "tells" of an impending injury—such as microscopic changes in stride length or arm slot—nearly invisible to the human eye. Traditional injury prevention relies on wearable sensors or subjective coaching observations, which are often unavailable or inconsistent. This research explores whether deep learning can detect these kinematic shifts using existing broadcast footage, providing a non-invasive "canary in the coal mine" for player health.

Methodology

The core challenge in using broadcast video for biomechanical analysis is avoiding model bias. Standard RGB video contains extraneous information like team jerseys, stadium lighting, and scoreboards, which can lead a model to overfit on specific players or environments rather than actual motion.

Input Representation and Architecture

To mitigate this, the researchers utilized optical flow as the primary input. Optical flow isolates pixel motion, forcing the model to focus strictly on the pitcher's mechanics. The architecture used is the Inflated 3D ConvNet (I3D), a standard for spatio-temporal video recognition, processing 10-second clips at 60 fps.

Sequence of a pitcher's delivery showing raw RGB frames (top) and the corresponding optical flow representation (bottom).

Labeling Strategy

Because overuse injuries develop over time, the researchers adopted a binary classification task. Pitches thrown immediately before a pitcher was placed on the Disabled List (DL) were labeled as "injured," while pitches from earlier in the season served as the "healthy" baseline.

Datasets and Experimental Setup

The study used a custom dataset from the 2017 MLB season, comprising 5,479 pitches from 20 different pitchers (12 left-handed, 8 right-handed). The dataset covered 10 different injury types, from UCL tears to hamstring strains. The model was pre-trained on the Kinetics dataset to establish a foundational understanding of human motion before fine-tuning on the baseball-specific data.

Key Results and Insights

  • Performance: For individual pitchers, the model achieved an average accuracy of 0.93 and an F1 score of 0.75. While performance remained strong when testing on seen pitchers (F1 of 0.74), generalization to entirely "unseen" pitchers dropped significantly (F1 ~0.35), highlighting the unique nature of individual deliveries.
  • Predictive Lead Time: The model's accuracy peaked when analyzing the 10–30 pitches preceding a DL stint. Beyond 50 pitches, the signal weakened, suggesting that mechanical breakdown occurs relatively rapidly.
  • Injury Specificity: The model was most effective at identifying injuries involving large, full-body compensations (e.g., back or hamstring strains) but struggled with localized issues like finger blisters.
  • Generalization via Mirroring: Horizontally flipping video of left-handed pitchers allowed models trained on right-handers to perform better, suggesting that "injury signatures" share symmetrical biomechanical traits.

Conclusion

This work demonstrates the potential of computer vision to provide early warnings for athletic injuries using existing infrastructure. While the current model faces challenges in generalizing to new pitchers and distinguishing between pitch types, it establishes a framework for non-invasive biomechanical monitoring. Future work will focus on expanding the dataset and improving interpretability to provide coaches with specific actionable insights.

Summary updated

Abstract

Injuries are a major cost in sports. Teams spend millions of dollars every year on players who are hurt and unable to play, resulting in lost games, decreased fan interest and additional wages for replacement players. Modern convolutional neural networks have been successfully applied to many video recognition tasks. In this paper, we introduce the problem of injury detection/prediction in MLB pitchers and experimentally evaluate the ability of such convolutional models to detect and predict injuries in pitches only from video data. We conduct experiments on a large dataset of TV broadcast MLB videos of 20 different pitchers who were injured during the 2017 season. We experimentally evaluate the model's performance on each individual pitcher, how well it generalizes to new pitchers, how it performs for various injuries, and how early it can predict or detect an injury.
BibTeX
@inproceedings{piergiovanni2019early,
  author = {A. J. Piergiovanni and Michael S. Ryoo},
  title = {{Early Detection of Injuries in MLB Pitchers From Video}},
  booktitle = {Workshops @ CVPR},
  year = {2019},
  url = {https://openreview.net/forum?id=_MEb0jendB},
}
Figure