For years, the best biomechanical screening in baseball has lived behind a velvet rope. If you wanted gold-standard motion capture, the kind that measures a pitcher's delivery down to the degree, you needed a stadium-scale, multi-camera system that runs somewhere between $150,000 and $500,000 per venue. That price tag is not a footnote. It is the entire reason this technology has been reserved for a handful of professional organizations and elite facilities while every high school, most colleges, and the overwhelming majority of training programs in the country have been locked out. A 2026 preprint out of the University of Waterloo just asked an audacious question: what if you did not need the lab at all? What if a broadcast feed, the same video you watch on your couch, could do the job?

What the Study Found

The researchers built a pipeline that takes ordinary broadcast footage, the single-camera kind from a televised game, and extracts 18 clinically relevant biomechanical metrics from it. That alone is a serious technical lift, because broadcast video is messy. There is motion blur, compression, a single moving camera, and pitchers contorting into extreme positions. The pipeline had to reconstruct a believable, physically plausible three-dimensional skeleton from all that noise. Then it had to prove the numbers it produced actually meant something.

It did, to a degree that should make the motion-capture industry sit up. Validated against professional tracking on 13 pro pitchers across 156 paired pitches, ranging from 75 to 102 mph, 16 of the 18 metrics matched the gold-standard systems to sub-degree accuracy, with a mean absolute error under one degree. Read that again. Television footage, run through this pipeline, agreed with half-million-dollar stadium hardware to within a single degree on the large majority of measurements. The expensive moat that has protected elite screening suddenly looks a lot shallower than it did a year ago.

But the part that genuinely matters for injury prevention is what happened when they fed those video-derived metrics into a prediction model. Trained across 7,348 pitchers using a machine learning ensemble and scaled to well over 100,000 professional pitching sequences, the model reached an AUC of 0.811 for predicting Tommy John surgery and 0.825 for significant arm injury. Those are not perfect numbers, but for injury prediction in a chaotic, multifactorial system like a throwing arm, they are genuinely strong. And then the model told us something about where the risk actually lives, which is the finding I keep coming back to.

It was not the average mechanics that predicted injury. Variability and extremes carried it, accounting for somewhere between a third and 40% of the model's importance, outranking the average delivery entirely. Specifically, a pitcher's worst 10% of pitches, what their mechanics look like on their most stressed and extreme throws, were stronger predictors than their averages. Hip-shoulder separation at foot plant ranked among the very top features. And sitting above all of it, the single strongest predictor of a future injury was, once again, a prior injury. To be honest, this reminds me of how you judge a bridge. You do not learn much from how it behaves on a calm day with light traffic. You learn everything from how it behaves under the heaviest load, in the worst conditions, at its most stressed moment. The model figured out the same thing about pitchers. The danger is not in the typical pitch. It is in the outliers.

Why This Information Is Important

I want to take the prediction findings seriously before I get to the disruption, because they line up almost eerily with what is already in the research.

Start with the idea that variability and the worst pitches matter more than averages. A 2024 study by Wang and colleagues followed pitchers through a fatiguing bout and found that center-of-mass variability rose and the pitchers leaned on hip and knee compensations, all while their velocity and accuracy stayed unchanged. The breakdown was happening in the variability of the movement before it ever showed up in the output. That is the mechanism behind the model's finding. A pitcher's averages can look stable and even healthy while the spread of their deliveries is quietly widening, and that widening spread is the early warning. If you only ever measured the average, you would see nothing wrong right up until something tore.

The within-pitcher angle is just as well supported. Manzi and colleagues, in 2021, showed that comparing one pitcher to another tells you very little about arm loading, with weak between-pitcher correlations, but comparing a single pitcher to himself tells you almost everything, with correlations above .85. This is why a model keyed to a pitcher's own deviations and extremes works so much better than one benchmarked against the group. The meaningful signal is internal. A pitch that would be unremarkable for one arm might be that pitcher's most dangerous throw of the night, and only his own baseline reveals it.

And the specific top feature, hip-shoulder separation at foot plant, is not an arbitrary output of a black box. A 2025 study by Johnson and colleagues found that under fatigue, hip-to-shoulder separation dropped and sequencing timing shifted after only about 35 pitches. So the exact variable the model flagged as most important is a variable we already know drifts measurably as a pitcher tires. The model is not finding magic. It is finding fatigue, expressed through a kinematic signature, and doing it from a TV feed.

Here is where the screening value becomes obvious, because our current alternatives are not great at this. A 2024 study by Hoshika and colleagues found that roughly 30% of completely asymptomatic players had high-grade UCL damage on MRI, statistically indistinguishable from the symptomatic group. Think about what that means. The most expensive static imaging we have cannot reliably tell you who is about to get hurt, because plenty of pain-free arms already look torn up on a scan. A dynamic, movement-based model that watches how the delivery changes over time is attacking exactly the gap that imaging leaves wide open. It is not redundant with an MRI. It is doing a different and arguably more useful job.

And the previous-injury finding deserves its own moment, because it never stops being true. Erickson and colleagues, in 2025, found that after a revision UCL reconstruction, only 59% of professional pitchers returned to their prior level, and that pitchers needing a second surgery had often returned too early from the first. Injury begets injury. Every model, including this one, keeps landing on prior injury history as the dominant predictor, which tells us that protecting a pitcher's first injury, and respecting the long shadow it casts, is among the most important things we can do.

Now, the disruption. Stack all of that on top of a pipeline that runs off broadcast video, and the implications for access are enormous. If sub-degree screening no longer requires a half-million-dollar installation, then the same caliber of analysis that has been reserved for a few professional venues could in principle reach high schools and colleges that have never been able to afford it. Even more compelling is the possibility of doing this on the fly. Imagine recognizing that a pitcher's variability is climbing and his hip-shoulder separation is decaying, in real time, during the game, in the inning where it is happening, long before it becomes an injury. That is the future this kind of model points toward, and it is a future worth being excited about.

How Can This Information Be Applied

The first shift, and it does not require any new technology at all, is to stop screening the average delivery and start watching variability and the worst pitches. Whatever tools you have, your eyes, video, a radar gun, the question is not what a pitcher's typical mechanics look like, it is how much they are spreading out and what the ugliest throws look like under fatigue. The model is telling us the danger lives in the outliers, so that is where our attention belongs.

The second shift is to monitor each pitcher against himself, not against a benchmark. A widening gap from a pitcher's own baseline is the signal, and a number that looks fine compared to the population can still be a red flag compared to where that specific athlete was a month ago. Individualized monitoring is not a nice-to-have here, it is the whole point.

The third is to treat prior injury as the dominant risk factor it keeps proving to be. A pitcher returning from an arm injury is not the same risk profile as one who has never been hurt, no matter how good he looks, and the ramp back deserves more patience and more scrutiny than we usually give it.

I do want to keep my excitement honest, because this is a preprint and the validation has real limits. The metric-extraction step was checked against tracking on only 13 pitchers, which is a small sample for a claim this big, and an AUC around 0.81 is strong for injury prediction but is nowhere near a crystal ball. This is a promising proof of concept, not a finished product, and the individual variability that runs through all of pitching means no model will ever hand us a clean yes or no for a given arm. But the direction is the thing. The barrier was never that we did not know what to measure. It was that measuring it well cost more than most programs would ever have. If that barrier is falling, the whole conversation about who gets access to good screening changes.

Conclusion

Two stories live inside this study, and both are worth telling. The first is that, yet again, the research points at the same handful of truths: previous injury is the strongest predictor we have, and the risk hides in a pitcher's variability and worst pitches rather than in their tidy average delivery. The second story is the one that could actually change the game, which is that the gold-standard screening we have walled off behind a half-million-dollar price tag may be reproducible from a broadcast feed. If that holds up beyond this preprint, the lab does not stay in the stadium anymore. It walks out into every gym, every high school field, and eventually into the game itself, watching for the moment a pitcher's mechanics start to drift before anyone, including the pitcher, can feel it. We are not there yet. But for the first time, it does not look like science fiction. It looks like a budget line that just got a lot smaller.

References

Scalable Injury-Risk Screening in Baseball Pitching From Broadcast Video. Preprint, 2026.

Wang SM, Huang TS, Chen SH, et al. Effect of Repetitive Pitching on the Control of Lower Extremity Joints and Center of Mass in Collegiate Baseball Pitchers. Sports Health. 2024.

Manzi JE, Estrada JA, Dowling B, et al. Intra- versus Inter-pitcher Comparisons: Associations of Ball Velocity With Throwing-Arm Kinetics in Professional Baseball Pitchers. Journal of Shoulder and Elbow Surgery. 2021.

Johnson AL, Kokott W, Dziuk C, Cross JA. Assessment of Muscular Fatigue on Hip and Torso Biomechanics in Adolescent Baseball Pitchers. 2025.

Hoshika S, Matsuki K, Takeuchi Y, et al. Microscopic Magnetic Resonance Imaging Comparing Asymptomatic and Symptomatic Ulnar Collateral Ligament Injuries. The American Journal of Sports Medicine. 2024.

Erickson BJ, Camp CL, et al. Outcomes of Revision Elbow Medial Ulnar Collateral Ligament Reconstruction in Professional Baseball Players. The American Journal of Sports Medicine. 2025.