Athlete Performance Data: What the Science Is Missing

Athlete performance data looks abundant, because professional teams put sensors on their players and a phone can film a jump. However, the rules that coaches train by rest on far less than that. For example, one widely used workload rule began with 28 cricket bowlers ¹. Likewise, a well-known injury forecasting model was trained on 26 soccer players from one club ². In addition, a review of 204 injury prediction models found that none had been tested on a second group of athletes ³. SAVRN is looking at a project in athletic performance, so I wanted to read the record first. This post is that homework.

Why this matters now

The number of athletes is large and growing. For example, high school sports in the United States counted 8,266,244 participants in 2024-25, which was a record ⁴. College sports set a record too, since the NCAA counted 554,298 athletes in its championship sports that year ⁵. Meanwhile, a federal survey put medically attended sports and recreation injuries at 8.6 million a year in 2011 to 2014. People aged 5 to 24 accounted for 64.9% of them ⁶.

At the professional level, injuries cost money and games. For example, an insurance broker counted 4,456 injuries in Europe’s top five men’s soccer leagues in 2024/25, with €676.14 million paid in wages to injured players ⁷. In addition, a study of the English Premier League estimated that injuries cost a team about £45 million a season in lost performance ⁸. Across 24 European clubs and 11 seasons, a club tended to finish higher in its league in seasons when its injury burden was lower ⁹.

So the need is clear, and the sensors are already on the field. What is thin is the evidence in between, because most published studies are small and most of the athletes in them are men. In addition, teams often treat their data as a competitive advantage, so few findings are ever checked on a second group ¹⁰. As a result, a coach can have a screen full of numbers and still have no tested rule for using them.

What does the record show in numbers?

Table 1 holds the figures this post relies on. I read each one in the paper or report that published it, and the sources are listed at the end.

How small are the studies behind common training rules?

Start with the best-known rule in load management. It compares this week’s workload with the average of recent weeks. In 2016, a review proposed that a ratio of 0.8 to 1.3 was a training “sweet spot,” while 1.5 or more was a “danger zone” ¹¹. The International Olympic Committee then repeated that range in a consensus statement the same year ¹². However, the first studies behind the ratio followed 28 cricket bowlers and 53 rugby league players ^{1, 13}. A later review found 20 studies of the ratio in professional team sport, although every one of the 1,234 athletes was male ¹⁴.

Then the rule was tested. First, a randomized trial divided 34 elite youth soccer teams, both girls and boys, into two groups. Eighteen teams were assigned to manage load by the ratio for 10 months. Nine teams soon withdrew, which left 482 players in the trial. It found no reduction in health problems, since the relative risk was 1.01 ¹⁵. Second, a re-analysis divided each player’s weekly load by one fixed number, the group average, instead of that player’s own recent average. The injury odds came out similar, at 1.95 against 2.45 ¹⁶. Third, a study followed 2,633 adolescent ice hockey players. Injury risk rose steadily as planned hockey time rose relative to the prior four weeks, so there was no sweet spot ¹⁷.

Other familiar advice has the same shape. For example, one finding holds that adolescent athletes who slept under 8 hours were 1.7 times more likely to have had an injury. It came from 112 athletes at one school ¹⁸. For adult athletes, by contrast, a review of 12 prospective studies did not support poor sleep as an independent injury risk ¹⁹. Likewise, training guided by heart rate variability is a common wearable feature, while the pooled trials behind it cover 198 people ²⁰.

The field has measured this pattern in itself. For instance, one review reported a median sample size of 19 in the Journal of Sports Sciences. It also reported that only 12 of 120 studies had worked out in advance how many people they needed ²¹. When researchers repeated 25 published studies from top journals, 7 held up, which is 28% ²².

How many injuries does a study need?

The count that matters is injuries, because a study learns from the athletes who get hurt. In 2003, two researchers worked out the requirement. Specifically, a study needs 20 to 50 injury cases to detect a moderate to strong risk factor. However, it needs about 200 injured athletes to detect a small to moderate one ²³.

Very few studies reach that bar. For example, the landmark knee screening study followed 205 female athletes and recorded 9 torn anterior cruciate ligaments, or ACLs. It reported that one landing measure predicted injury with 78% sensitivity and 73% specificity ²⁴. Eleven years later, a study of 710 elite female players recorded 42 new noncontact ACL injuries. However, none of its five landing measures predicted injury in players who had not been hurt before ²⁵. Figure 1 sets seven well-known studies against the bar, including three team sport models that recorded 18, 32 and 51 injuries ^{26, 27, 28}.

The arithmetic shows why one team cannot settle the question alone. Across 58 studies, female athletes tore an ACL at a rate of 1.5 per 10,000 athlete exposures, where an exposure is one practice or game ²⁹. Therefore a study needs about 1.33 million exposures to record 200 injuries at that rate. Suppose a squad of 25 logs 150 practices and games each per season, which is 3,750 exposures. That squad would need about 356 seasons. As a result, the only route to the number is many teams keeping the same record in the same way.

How well do injury prediction models work?

In a 2026 review of 97 sports medicine AI studies, prediction and estimation models made up 57.7% of applications, so this part of the record matters ³⁰. Models are usually scored on a scale where 0.5 is a coin flip, while 1.0 is perfect.

The best-known model used GPS data from 26 professional soccer players at one Italian club, who had 23 injuries. Its precision was 0.50, so half of its warnings were false alarms ². A review of machine learning injury models found 11 studies, with scores from 0.52 to 0.87 ³¹. However, a broader review of 30 studies and 204 models rated 98% of the models at high or unclear risk of bias. In addition, no study had tested its model on outside athletes ³.

When models face a harder test, the scores fall. For example, hamstring models built on 2013 data from elite Australian footballers had a median score of 0.52 on 2015 data ³². The largest ACL cohort in a team ball sport analyzed 283 variables for 791 elite female players, and 60 of them tore an ACL. Although the data was rich, its best model scored 0.63 ³³. Similarly, a model for 314 youth basketball and floorball players scored 0.65 ³⁴. In 2026, researchers took an ACL model to an independent group of 320 female Australian football and soccer players. The score was 0.67, while the model overstated the risk ³⁵.

A more complex method has not been the answer. Across 125 head-to-head comparisons at low risk of bias in orthopedic sports medicine prediction models, machine learning showed no advantage over ordinary logistic regression ³⁶. Meanwhile, a review of 97 sports medicine AI studies found only one that tested a prevention strategy built on a model ³⁰.

Who is missing from the data?

Women are the largest gap. An audit of 5,261 publications in six sport science journals found that 31% studied men only, while 6% studied women only. In addition, women were 34% of all participants ³⁷. The gap is wider in AI research, since a review of 58 team sport studies found a pooled sample that was 97% male ³⁸. In elite soccer research, likewise, women were 7% of all participants ³⁹.

The gap matters because outcomes differ. For example, female athletes tear an ACL at 1.7 times the rate of male athletes across sports ²⁹. Among adolescents the ratio is 1.40 overall, although it reaches 4.14 in basketball ⁴⁰. In high school sports played by both sexes, girls had 2.22 times the concussion rate of boys ⁴¹. However, the research into why is thin. For instance, a 2026 analysis of muscle injury across the menstrual cycle found 3 eligible studies with 318 athletes, and it could make no recommendation ⁴².

Other groups are close to absent. For example, an audit of 842 sports medicine studies found that 3.8% reported race, while 2.0% reported ethnicity ⁴³. Para athletes made up 0.4% of participants in 767 studies of the athlete’s heart ⁴⁴. However, injuries at the Paralympic Games run at 14.3 per 1,000 athlete days, against 6.5 at the Olympic Games ⁴⁵. In addition, only 19% of 186 validation studies for wearable optical sensors reported skin tone. The median share of participants with the darkest skin tones was 0% ⁴⁶.

Young athletes are a special case, because a growing body changes the risk. In one academy study, non-contact and overuse injuries ran at 24.5 per 1,000 match hours around the growth spurt, against 11.5 before it ⁴⁷. However, the three growth studies found for this post followed 26, 76 and 63 players. All of them were boys in soccer academies ^{47, 48, 49}. In fact, one of those studies took 21 years of records at a single club to assemble ⁴⁹.

Small samples also produce numbers that travel too far. For example, one study of 36 players at a single club found that hamstring injury odds rose 1.78 times with each year of age ⁵⁰. By contrast, a study of 1,401 professional players found no link between age and hamstring injury ⁵¹.

Where has large-scale data already paid off?

Where someone has kept a large and consistent record, the results have been usable. For example, UEFA’s injury study followed elite men’s soccer clubs for 18 seasons, covering 3,302 players and 11,820 injuries. Over that period, injury rates fell 3% a season in both training and matches ⁵².

American school sports have a similar record. Specifically, one surveillance study has logged 134,167 injuries over 67,255,905 athlete exposures in 20 years of high school sports ⁵³. That record is where the concussion comparison between girls and boys came from. Likewise, a concussion study run by the NCAA and the Department of Defense reports more than 53,000 enrolled athletes and cadets ⁵⁴. Its analysis of 1,071 concussions found no overall sex difference in median recovery, at 13.5 days for women against 11.8 for men. However, women took longer in contact sports and at Division II and III schools, while men took longer in limited-contact sports ⁵⁵.

Large records also change rules. In Ivy League football in 2015, kickoffs were 6% of plays, although they produced 21% of concussions. After a 2016 rule change, the concussion rate on kickoffs fell from 10.93 to 2.04 per 1,000 plays ⁵⁶. Similarly, the NFL says it simulated 10,000 seasons before it adopted a new kickoff format ⁵⁷. In 2024, the league reported that returns rose 57%, while the concussion rate on the play fell 43% against the 2021 to 2023 average ⁵⁸. However, the next season showed why the record has to stay open. With 1,157 more returned kickoffs, concussions on kickoffs rose from 8 to 35, although the league says the rate stayed below the old format ⁵⁹.

Registration data can work the same way. For instance, Portugal’s registration records covered 126,285 boys and showed a birth date effect in every age group. The effect was far stronger at selection. Boys born in the first quarter of the year had 12.36 times the odds of being on the under-17 national team, against boys born in the last quarter ⁶⁰.

What will more data not fix?

Three findings complicate the case for data, and they belong in this post.

First, some answers already exist and go unused. Programs that include the Nordic hamstring exercise cut hamstring injuries by about half across 8,459 athletes ⁶¹. However, a survey of 50 elite men’s soccer clubs found that 83.3% of club seasons did not follow the program ⁶². Meanwhile, in elite European men’s soccer, hamstring injuries rose from 12% to 24% of all injuries between the 2001/02 and 2021/22 seasons ⁶³.

Second, screening has a limit that volume does not remove. A 2016 review found no injury screening test with adequate accuracy. The reason is overlap, because scores for athletes who get hurt look much like scores for those who do not ⁶⁴.

Third, more measurements are different from more athletes. For example, a review of 45 workload studies found that a single study tested between 1 and 336 variables ⁶⁵. In addition, a review of 56 studies found that asking athletes how they felt tracked training load better than objective measures did ⁶⁶. Professional baseball also keeps a league-wide injury tracking system. Its records show that 2,281 pitchers had elbow ligament surgery from 2010 to 2023, and the yearly number rose ⁶⁷.

Can a phone camera replace the motion lab?

Smartphone video can collect movement data at a fraction of a motion lab’s cost, since it needs no markers and no lab space. For example, the authors of one open system put a conventional lab at more than $150,000 in equipment. Their setup, by contrast, runs on two or more smartphones and costs under $700. In a field test, they collected data on 100 people in 8 hours ⁶⁸.

However, accuracy depends on the movement. That system’s joint angle error averaged 4.5 degrees in 10 healthy adults ⁶⁸. In jumping, a pooled analysis of 20 studies found errors of 5.3 degrees at the hip and 4.4 at the knee ⁶⁹. In addition, a review of 53 studies found that camera systems differed from marker systems by 0.2 to 28.6 degrees, depending on the measure ⁷⁰. As a result, records from different systems cannot simply be mixed.

Training data for sports is scarce as well. One athlete pose dataset holds about 1.3 million frames, although all of them come from eight athletes ⁷¹. Another covers 24 people ⁷². An open soccer event dataset, the largest when it was released in 2019, covers 1,941 matches and 4,299 players, all from men’s competitions. In addition, it records events on the ball instead of movement or injury ⁷³.

The volume at the top of the sport is of a different order. For example, the NFL says its system uses 38 cameras in a stadium and processes about 6.8 million video frames in each week of games. Meanwhile, practices add more than 500 million tracking data points a week ⁷⁴.

Who owns an athlete’s data?

The rules depend on where the athlete plays. In the NFL, for example, the 2020 labor agreement states that each player owns the personal data collected by his sensors ⁷⁵. In the NBA, the 2023 agreement makes team wearables voluntary, while it gives a player full access to his data. It also allows a fine of up to $250,000 for a team that uses the data in contract talks ⁷⁶.

College athletes sit under a different law. Specifically, federal guidance states that at most colleges, student health records kept by a campus clinic fall under the student privacy law known as FERPA. Therefore they are excluded from the health privacy law known as HIPAA, even when the school is covered by it ⁷⁷. In soccer, meanwhile, the players’ union and FIFA published a charter in 2022 that lists eight data rights. In the union’s survey of 119 players, 80% wanted access to their own data ⁷⁸.

These rules differ by league, by country and by level of play. Therefore a shared dataset has to settle consent, access and ownership before the first record is collected.

What would a useful athlete dataset contain?

The studies above point to a short list, while Figure 4 shows the evidence behind each item.

1. Enough injuries, because detecting a small to moderate risk factor takes about 200 injured athletes. That means many organizations pooling their records ²³.
2. One definition of injury. The International Olympic Committee published a recording standard in 2020, since 11 separate statements for single sports or settings had already appeared ⁷⁹.
3. Exposure counts. A rate needs a denominator. Therefore every practice and game is logged, including the ones without an injury.
4. Demographics on every record. Specifically, that means sex, age, growth stage, level of play, race, ethnicity and impairment class. Of 13 studies that analyzed race, 8 found differences in outcomes ⁴³.
5. A held-out group, because a model has to be tested on athletes from a different organization before anyone relies on it ³.
6. Consent and ownership in writing, settled before collection starts.

How SAVRN is approaching this

We build AI factories, so our part of this problem is the computing and the control of the data. Both are large. For example, the smartphone system above needed 31 hours of computing for 100 people, which is about 19 minutes each ⁶⁸. Therefore one session for each of the 554,298 NCAA athletes would take about 172,000 hours on the same machine. That is more than 19 years, so the work has to run in parallel on far more capacity.

Control matters as much as capacity. In September we wrote about why we train a model on our own data instead of handing the data over. The same choice applies to athlete records, since they are health records about named people. Likewise, our university AI infrastructure work starts with the institution deciding where data can go and who can use it. In addition, our dataset catalog listed 1,578 open datasets with their licenses on September 27, 2026, because a model can only be traced if its data can be.

We set out the principle behind that work, and a design case for athlete development, in Data to Intelligence. The project we are looking at is at an early stage, and nothing is announced. However, the record above sets the terms for it. First, the dataset would pool many organizations, because no single team records 200 athletes with one specific injury, such as an ACL tear, on its own. Second, it would cover women, para athletes and growing young athletes from the first day, since the current record is thin on all three. Third, it would test every model on outside athletes before use. Finally, it would be measured by injuries prevented, because prediction alone has not been shown to protect athletes.