In a statement released last Friday, the attorney for injured ace Jose Fernandez—who underwent Tommy John surgery to repair his torn ulnar collateral ligament the same day—accused the Marlins coaching staff of not picking up on an “unanticipated change” in his delivery caused by discomfort in his second-to-last start preceding the surgery.

The attorney’s claim implies that the Marlins should have recognized the writing on the wall, which makes this issue a touchy one. The club’s response was somewhat predictable: manager Mike Redmond questioned whether there had been such a change and asserted that Fernandez should have notified the team if he had been experiencing discomfort.

Fernandez’s silence is not unprecedented. Most pitchers are competitive people. Last year, Jaime Garcia admitted to battling shoulder discomfort in the four starts that preceded a surgery to repair his torn rotator cuff and labrum. The Cardinals were racing toward the playoffs, and Garcia wanted to be a part of that effort. Garcia didn’t talk about his left shoulder to anyone on the Cardinals staff until the middle of that fourth start. By then, surgery was unavoidable.

What’s the Injury Zone?
Years ago, former online baseball writer (and current Tampa Bay Rays analyst) Josh Kalk developed a machine learning algorithm that was designed to predict pitcher arm injuries just before they occurred. This type of a model makes use of relevant and publicly available pitch velocity, movement, and release point PITCHf/x data. It overlooks all qualitative indicators, but that doesn’t mean that those aren’t important. I’d recommend taking a look at this Doug Thorburn piece for a detailed mechanical breakdown of Fernandez’s warning signs.

The term “Injury Zone” may be misleading, because (in at least some cases) pitchers who test positive by this model’s standards tend to be injured already. When I reevaluated Kalk’s model last February, I suggested that this kind of tool isn’t necessarily an injury preventer, but instead a mechanism to prevent an existing arm injury from worsening.

An important distinction between my model and Kalk’s is that mine omits all non-fastball data. I’m hesitant to do that when looking at Fernandez, but I’d take it over any alternative. When asked about Fernandez’s injury on the Effectively Wild podcast last week, Thorburn argued that the motion behind a breaking ball is inherently dangerous to a pitcher’s arm. In fact, Fernandez’s exceptional breaking ball might be the missing piece to the injury puzzle—we don’t know. Regardless, I want to make it clear that I am omitting it from this neural network.

The model uses historical data to generate a number between zero and one that represents the likelihood of injury on the next pitch. In the full-length study, I drew an arbitrary cutoff of .80 (or an 80 percent chance of injury) for all Injury Zone candidates. Why did I choose this number? Well, after repeated trials, I kept noticing that injured pitchers received all sorts of predicted values (ranging from 0.0-1.0), but I never found a single case of a healthy pitcher receiving a mark above .80.

So, if a manager removes a pitcher who scored below that critical .80 threshold, he might not actually be injured. Something to keep in mind, though: a false negative is a lot more concerning than a false positive. I wouldn’t mind a few incorrect early hooks in exchange for one missed ligament tear, and I’m sure the Marlins wouldn’t either.

Enter Fernandez
Fernandez’s silence about this pre-existing arm fatigue makes him the perfect Injury Zone candidate. His attorney argues that he was not himself in his pre-injury start (and the one before that). The PITCHf/x cameras in each park may or may not support that claim.

Before I get to Fernandez’s injury zone numbers, I’d like to highlight something I found in his 2013 numbers (which were used as a baseline for this study) that surprised me quite a bit. In my revision of the model, I introduced a measure that captures the consistency of a pitcher’s release point.

Release point variance turns out to be a pretty strong predictor of injury risk, and of a pitcher’s ability to throw strikes. For reference, the game’s most consistent pitchers (Cliff Lee) post 10-pitch release point standard deviations that stick around .09-.10 feet. The game’s wildest and most volatile arms (Carlos Marmol, anyone?) approach the .20 feet mark.

Fernandez, who is lauded for his repeatable delivery, was actually closer to the latter group than he was to the former in 2013. He posted values ranging from .15-.17, with those readings increasing as he fatigued. We may have assumed that Fernandez has never exhibited any causes for concern, but that’s an incorrect assumption. As Thorburn noted recently, even the game’s cleanest deliveries don’t always prevent a trip to the operating table.

I ran the model five times for Fernandez and consistently arrived at the same conclusion. There’s a conclusion that I thought I’d have gotten to by this point: Fernandez was visibly fatigued toward the end of his final start, and I assumed (before doing this analysis) that I’d be calculating the millions of dollars worth of destruction caused by a huge managerial mishap in this article.

However, the numbers alone suggest that Redmond might not have made the worst managerial decision of the 2014 season. Fernandez’s last 15 fastballs on May 9 earned him Injury Zone ratings ranging from .71-.74—very troubling, but not a slam-dunk case. If you remove Fernandez here, you also remove a few healthy pitchers. Obviously, we have the benefit of the postmortem, but Redmond would have drawn quite a bit of flak for pulling an uncomplaining Fernandez before pitch 80.

The image below plots injuries in red, healthy starts in blue, and the intersection of the two in purple. In different instances of the same model, the results aren’t always distributed exactly like they are in the one shown below, but the peaks generally follow the same trend.

Fernandez did exhibit a decline in horizontal and vertical spin deflection, but the velocity drop is the biggest and most visible cry for help. His fastball velocity fell by almost two and half standard deviations from its normal level at that pitch count, suggesting that this was the type of in-game decline that occurs only about once every two seasons.

Fernandez didn’t lose his level of release point consistency, as some injured starters tend to do just before departing. His last 15-pitch variance (.170 feet) was almost exactly in line with his 2013 average for the 65-80-pitch range. His release point (which isn’t adjusted for calibrations differences between parks) was definitely elevated, but not dramatically.

Could a case have been made for pulling Fernandez even earlier? The early velocity drop might imply that one could argue for an even earlier hook, but the neural network disagrees. I went as far back as 40 pitches before Fernandez’s exit and found results in the .31-.33 range for that point in time. The model’s predicted probability does increase as we approach the injury, though, as we’d expect.

After going back to the beginning of the May 9 start, I thought it might make sense to look at Fernandez on May 4 (when he allegedly began to experience discomfort) as an injury zone candidate. His velocity did fall off a bit toward the end of that outing, and he also began to rely heavily on his breaking stuff. His desire to stay away from the fastball might be an indicator in and of itself, but it also means that I can’t conduct a meaningful analysis using the same model I’ve employed here.

As is always the case with pitching injuries, the red flags preceding Fernandez’s exit look a lot clearer after the fact. It’s difficult to gauge how these PITCHf/x warning signs would work in practice, but the value gained in trading a Tommy John surgery for a DL stint could not be clearer in Fernandez’s case. The Marlins (and the other 29 teams) will have plenty of time to figure out how to keep tabs on the next Jose Fernandez while we wait for him to return to the hill in 2015.

You need to be logged in to comment. Login or Subscribe
a wonderful article and exactly the reason i subscribe. why don't teams pay more attention to this stat?
Are managers able to see Pitch f/x data in-game? Is there someone following along in the clubhouse that could relay the information to the manager? I know there is a tradition of the next game starter doing pitch charting, is this still necessary if a manager could get real-time Pitch f/x?
If managers wanted, they could certainly get real-time PITCHf/x data. Fans thousands of miles away can do that just by using the Gameday app online. From what I know, however, it would be very unusual for a front office-style tool to influence in-game management. That's just how things are now. The problem with using a model like this, which Noah has said, is that it is really good at post-hoc explanations for injuries -- but not necessarily predicting confidently enough in time for someone to run out to the mound and say "stop the fight!" Fernandez should have been taken out earlier by Redmond, IMHO, given a 7-10 mph dropoff in fastball velocity that happened very suddenly. But by that point, the torn UCL was already there.
Great article. The gist is clear. As someone not experienced with neural networks, I do not understand what is being referenced on the X-axis of your neural network. At .8, are we talking about a pitcher who has an 80% chance of injury on the next pitch or a pitcher who has an 80% greater chance of injury?
That's exactly right. The bar above .8 counts the number of pitchers who received this rating, and it implies that each of these guys have an 80% of injury on the next pitch.
Thanks Noah for running this, and great finding!
Do teams have immediate access to in-game pitch/fx data?
Yes, reverseman, they do. Read the above comments.
I know the point has been brought up regarding this research, but this is not an early enough indicator to help prevent injuries. Talking to anyone in the sports med community, you hear these are all injuries built up over years of abuse. Maybe Fernandez comes out two starts earlier with an even more sensitive model, but it's just as likely he gets an MRI that shows a partial tear, with surgery recommended. Ligaments and tendons don't go from 100% to 0% function. I'm definitely interested in the work, especially because I think it will be useful in identifying pitchers who are trying to pitch through injury. We just need to keep reasonable expectations with it.
I was going to ask, would pulling someone at 0.7 or whatever actually decrease their likelihood of injury, or just postpone it until the next start? It's not like a damaged ligament or muscle will return to 100% function in the < 1 week between starts.