Pitcher Injury-Risk Model

The Problem

Major League Baseball teams face a critical operational challenge: predicting which pitchers are at highest risk of injury before it happens. Injuries are costly—both financially (medical, lost production, roster turnover) and strategically (mid-season rotations collapse without notice). While pitch-count management has become standard, quantifying individual injury risk requires integrating multiple data streams: biomechanics, workload, age, prior injury history, and season-level stress.

Traditional approaches use rule-of-thumb thresholds (e.g., “don’t throw over 100 pitches”) but miss the nuance: a 26-year-old in his first heavy workload is not the same risk as a 34-year-old in his tenth year. The question is: Can we predict injury with enough precision to shift roster construction, training prioritization, and bullpen usage?

Why It Matters

Pitcher injuries are:

• Financially significant — an ace on the IL costs $1–2M in replacement value per month.
• Predictable to a degree — injury isn’t random; it accumulates with workload, age, and prior history.
• Actionable — if you can flag high-risk scenarios in advance, you can adjust load, recommend training interventions, or adjust trade strategy.

For sports analytics, this is a canonical use case: real data, clear outcome, measurable business impact. A model that improves roster health by even 5–10% is worth significant investment.

My Approach

Data Sources

• Statcast (2015–2024): Every pitch thrown in MLB, including velocity, movement, release mechanics.
• Injury history: Retrosheet + manual research for all pitcher DL/IL stints spanning 15+ years.
• Workload metrics: Pitches thrown, innings pitched, rest days between appearances, season totals.
• Biometric proxies: Age, prior surgery count, career length, velocity decline year-over-year.

Methodology

1. Feature engineering: 12 primary features capturing usage intensity, age cohort, prior injury, and velocity anomalies.
2. Class imbalance handling: Injury is rare (~3–4% of pitcher-seasons); used stratified CV and class weighting (SMOTE considered but rejected due to leakage risk).
3. Model selection: Logistic regression with L2 regularization (interpretability priority), then gradient boosting (XGBoost) for robustness.
4. Evaluation: Stratified 5-fold cross-validation; primary metric AUC-ROC; secondary metrics precision/recall tradeoff for high-risk flagging.
5. Validation: Holdout 2024 season cohort; prospective tracking planned.

Key Decisions

• Why logistic regression first? Baseball scouts and front-office staff need to understand why a pitcher is flagged. A 0.5 coefficient on “age” is interpretable; a tree split is not.
• Why not neural nets? Data is relatively small (2,847 pitchers); overfitting risk is real. Simpler models generalize better and are more trusted.
• Why stratified CV? Injury is class-imbalanced; naive CV underestimates performance on the rare class.

Results

Model Performance

Chart interpretation: The interactive chart above shows predicted injury risk across pitcher ages (22–36) stratified by workload intensity. Three lines represent different usage levels: high usage (peaked at 13.1% risk for 36-year-olds), medium usage (peaked at 9.9%), and low usage (peaked at 6.7%). Notice the inflection at age 30 under high usage—this is where cumulative stress and age intersect.

Chart data summary for screen readers: This visualization displays predicted injury risk percentages by pitcher age and workload. High-usage pitchers show injury risk starting at 3.2% at age 22 and climbing to 13.1% at age 36. Medium and low usage show proportionally lower trajectories. All three usage categories demonstrate a consistent increase in injury risk with age.

Validation Results

• Training AUC-ROC: 0.79
• Cross-validation AUC-ROC: 0.78 (±0.04)
• Holdout 2024 AUC-ROC: 0.76
• High-risk precision (top 10%): 64% (true injury rate among flagged pitchers)

The slight gap between CV and holdout reflects real distribution shift (2024 had unusual injury patterns) but is within acceptable bounds.

What the Model Captures

1. Age effect: Injury risk accelerates after age 30, especially under heavy use.
2. Workload dependency: 150+ innings + 50+ appearances nearly doubles risk vs. lighter loads.
3. Prior injury: A single prior IL stint increases risk by ~35%; multiple priors, ~70%.
4. Velocity loss: Year-over-year velocity decline of >2 mph correlates with +20% risk.

What It Misses

• Acute biomechanical changes (form breakdown mid-season): Statcast measures outcome, not mechanics.
• Off-field factors (training quality, sleep, nutrition): Unavailable in this dataset.
• Individual pitcher resilience: Two 32-year-olds with identical stats may have different durability.

Key Takeaways

1. Injury risk is predictable at the cohort level, with ~78% discrimination between high and low-risk scenarios. Individual prediction remains uncertain, but risk stratification is actionable.

2. Usage and age interact: A young pitcher can absorb high workload safely; an older pitcher cannot. This should inform roster construction—prioritize young depth, manage veteran usage carefully.

3. Simple models work well here. The gap between logistic regression (AUC 0.77) and XGBoost (AUC 0.79) is ~1%. The interpretability win of linear coefficients outweighs marginal accuracy gains.

4. This is a foundation, not an oracle. Real teams combine statistical models with expert evaluation (team medical staff, coaching observations, private biometric data). The model’s value is in surfacing patterns humans might miss and forcing explicit reasoning rather than replacing domain judgment.

5. The architecture scales. Adding more features (biomechanics, training load, genetic markers) is straightforward if data becomes available. The pipeline is reproducible and auditable—key for front-office buy-in.

---

Next steps: Prospective validation through 2025 season; integrate with team scheduling system to flag overload scenarios in advance.