Scalable Injury-Risk Screening in Baseball Pitching From Broadcast Video

This paper presents a scalable monocular video pipeline that recovers clinically relevant biomechanical metrics from broadcast baseball footage with high accuracy, enabling effective injury-risk screening for thousands of pitchers without the need for expensive stadium-based motion capture systems.

The Gist

Imagine you are a baseball coach standing on the sidelines. You want to know if your pitcher is about to blow out their elbow, but you don't have a million-dollar laboratory with high-speed cameras and sensors strapped to their body. In the past, you'd have to guess based on how the pitch looks.

This paper introduces a "magic lens" that turns a regular TV broadcast of a baseball game into a high-tech medical scanner. Here is how it works, broken down into simple concepts:

1. The Problem: The "Black Box" of Injury

In professional baseball, teams use expensive stadium cameras (costing up to $500,000) to track exactly how a pitcher moves their joints. This data helps predict injuries like "Tommy John" surgery. But for high school kids, college players, or independent leagues, that tech is impossible to afford. They are flying blind, trying to prevent injuries without the data to back it up.

2. The Solution: A "Digital X-Ray" from a Single Camera

The researchers built a computer program that can watch a standard TV broadcast of a pitcher and "see" their skeleton in 3D. Think of it like a digital puppeteer. The computer watches the video, finds the pitcher's joints (shoulders, elbows, hips), and builds a 3D stick-figure model of them moving in real-time.

3. The Challenge: Fixing the "Jittery" Video

Regular video is messy. It gets blurry when the arm moves fast, it compresses (like a low-quality YouTube video), and the pitcher's body often blocks its own limbs. If you just took the raw video and calculated angles, the numbers would be all over the place—like a shaky hand trying to draw a straight line.

The Fix (The "Refinement Stack"):
The authors created a cleaning process they call the Biomechanics Refinement Stack. Imagine you are fixing a shaky video of a dancer:

• The "Drift" Fix: Sometimes the computer thinks the pitcher is floating away. They added a "gravity anchor" (Pelvis-to-Global Lifting) to make sure the pitcher stays planted on the ground where they belong.
• The "Rigid Bone" Fix: In real life, your arm bones don't stretch or shrink. But computer guesses might make an arm look like it's growing an inch every second. The program forces the bones to stay the same length, like a rigid skeleton.
• The "Symmetry" Fix: If the computer gets confused and thinks the left arm is longer than the right, the program fixes it to match human anatomy.

4. The Result: 18 "Vital Signs" for Pitchers

Once the video is cleaned up, the system measures 18 specific numbers that doctors and coaches care about. These are like the pitcher's "vital signs":

• How bent is the knee when the foot hits the ground?
• How far is the shoulder twisted back?
• How much is the torso leaning forward?

They tested this on 13 professional pitchers. The computer's measurements were almost identical to the expensive stadium cameras (within 1 degree of error). It's like using a smartphone camera to measure a room and getting the same result as a laser measure.

5. The Payoff: The "Injury Crystal Ball"

The real magic happens when they feed these 18 numbers into an AI "doctor."

• They looked at data from 7,348 pitchers.
• They taught the AI to spot the patterns that happen before a pitcher gets hurt.
• The AI learned that it's not just the average movement that matters, but the extreme movements (the top 10% of the craziest pitches).

The Score: The system predicts the risk of a major arm injury with 81% to 82% accuracy. That is a huge leap from just guessing.

Why This Matters

Think of this system as democratizing safety.

• Before: Only rich professional teams could afford the "medical scanner" to check for injuries.
• Now: Any coach with a smartphone and a video of a game can run this analysis.

It allows a high school coach to say, "Hey, your pitcher's elbow is twisting 10 degrees more than usual on his hardest throws. Let's rest him for a week," before the injury happens. It turns a simple video into a life-saving tool, making professional-level safety accessible to everyone.

Technical Summary

1. Problem Statement

Injury prediction in baseball pitching relies heavily on precise biomechanical signals (e.g., joint flexion, trunk orientation, hip-shoulder separation). Currently, the "gold standard" for measuring these metrics involves expensive, stadium-installed multi-camera optical tracking systems (e.g., Hawk-Eye), costing between $150,000 and $500,000 per venue. This creates a critical accessibility gap: amateur leagues, college programs, and independent training facilities lack the infrastructure to obtain reliable kinematic feedback for early injury-risk screening.

The core challenge addressed by this paper is: Can accurate, clinically actionable biomechanical metrics be recovered from monocular broadcast video (single camera view) without dedicated hardware, despite challenges like motion blur, compression artifacts, and extreme pitching deformations?

2. Methodology

The authors propose a monocular video pipeline that transforms raw broadcast footage into stable, global 3D kinematics suitable for injury modeling. The pipeline consists of three main stages:

A. Pelvis-to-Global Lifting Module (PGLM)

Standard 3D pose estimators (like DreamPose3D) often output poses relative to the pelvis, lacking global translation.

• Approach: The authors introduce a lightweight sequence module using a BERT-style transformer encoder-decoder.
• Mechanism: Instead of regressing absolute translation directly (which causes drift), the model predicts changes in pelvis velocity relative to the first frame.
• Drift Control: To prevent error accumulation over time, the model runs on overlapping windows. It estimates trajectories for $n$ frames but only commits the first $K=10$ frames to the final sequence before sliding the window forward. This "predict-commit-recompute" strategy ensures temporal continuity while limiting long-horizon drift.

B. Biomechanics Refinement Stack (BRS)

Raw neural pose estimates often suffer from jitter, inconsistent bone lengths, and spurious asymmetries due to motion blur and occlusion. The BRS enforces physical plausibility through four steps:

1. Bone-Length Enforcement: A reference skeleton is estimated from the first 30 frames (median joint positions). For every subsequent frame, the system traverses the kinematic tree (rooted at the pelvis) and rescales child joints to maintain constant bone lengths, effectively performing a forward-reaching pass similar to the FABRIK algorithm.
2. Constrained Inverse Kinematics (IK): The system optimizes joint angles to minimize reprojection error against the length-corrected joints, subject to pitcher-specific joint limits ($\theta_{min} \le \theta \le \theta_{max}$). This ensures the pose remains within physiological ranges.
3. Smoothing: Temporal filtering is applied to reduce high-frequency jitter.
4. Symmetry Constraints: Left and right bone lengths (e.g., upper arms) are tied to a shared average length, and the pelvis is centered between the hips to reduce artifacts caused by partial occlusion.

C. Metric Computation & Injury Screening

• Handedness Estimation: Before computing metrics, the system classifies the pitcher as left or right-handed using ankle displacement and pelvis rotation vectors.
• Metric Extraction: 18 clinically relevant metrics are computed (e.g., knee/elbow flexion, shin angles, hip-shoulder separation, trunk tilt, Center of Gravity).
• Feature Engineering: Metrics are extracted at key phases (Foot Plant, Max External Rotation, Ball Release). These are aggregated into statistics (mean, std, range, P90, coefficient of variation) and combined with workload data and medical history.
• Prediction Model: An ensemble model (averaging Gradient Boosting, Random Forest, and L1-regularized Logistic Regression) is trained to predict Tommy John surgery risk and significant arm injuries.

3. Key Contributions

1. PGLM & BRS: The introduction of a drift-controlled global lifting module and a multi-stage refinement stack that stabilizes broadcast pose sequences into physically plausible global kinematics.
2. Scalable Injury Screening: A clinically actionable pipeline that uses pose-derived metrics to screen for injury risk without stadium hardware, democratizing access to biomechanical analysis.
3. Validation at Scale: Recovery of 18 metrics from monocular video, validated against professional tracking data on 13 pitchers, and generalized to a dataset of 119,561 professional pitches.

4. Results

Quantitative Validation (13 Professional Pitchers, 156 Pitches)

The system was validated against a professional multi-camera tracking database.

• Accuracy: 16 out of 18 metrics achieved sub-degree agreement (Mean Absolute Error, MAE < 1°).
  • Lower body angles (knee flexion, shin angles) and elbow flexion achieved errors < 0.5°.
  • Trunk/pelvis metrics (rotation, tilt, hip-shoulder separation) had MAE < 1° with correlations > 0.996.
  • Center of Gravity (COG) coordinates matched within 0.02–0.50 ft.
• Limitations: Shoulder abduction was the only metric failing the 1° threshold (MAE = 6.2° for throwing arm, 21.4° for glove side). The authors attribute this to the fundamental difficulty of markerless estimation for joints hidden beneath soft tissue/clothing, where small localization errors propagate into large angular deviations.

Injury Prediction Performance (7,348 Pitchers)

The automated screening model was tested on a large cohort including 167 recent Tommy John cases.

• Tommy John Prediction: AUC of 0.811.
• Significant Arm Injury Prediction: AUC of 0.825.
• Feature Importance: Biomechanical variability (range and coefficient of variation) and extreme values (P90) were found to be stronger predictors than mean values. Medical history remained the single strongest predictor, but the inclusion of kinematic data significantly improved performance over static thresholds.

5. Significance

This work bridges the gap between high-fidelity biomechanical research and practical, scalable application.

• Democratization: It proves that broadcast video can serve as a viable alternative to expensive motion capture systems for injury-risk modeling, making advanced biomechanics accessible to amateur and collegiate levels.
• Clinical Utility: The system enables proactive risk stratification, allowing coaches and medical staff to identify pitchers with "extreme" mechanics (e.g., high variability in hip-shoulder separation) and intervene before injury occurs.
• Technical Achievement: By combining deep learning pose estimation with rigorous, hand-crafted kinematic constraints (IK, bone-length enforcement), the authors overcame the noise and artifacts inherent in broadcast footage to achieve sub-degree accuracy comparable to professional lab setups.