Individualized high-velocity running exposure and lower-body soft-tissue injury in professional rugby: A within-athlete matched analysis

Sprinting is how most muscle injuries happen in rugby, and nearly every professional club now tracks how fast its players run using GPS units worn between the shoulder blades. That data is expensive to collect, so clubs naturally want to use it to predict who is about to get hurt. This study asked a blunt question: can one team's own data actually answer that?

Researchers took records from 47 male professional rugby union players at a single club over 20 weeks of the 2025 season, covering 1,457 training sessions and matches. Rather than comparing fast players to slow ones, they compared each injured player to himself. For every injury, they looked at the player's sprint exposure in the weeks beforehand, then looked at the same player 28 days earlier, when he was healthy. Sprint exposure was measured relative to each individual's own fastest recorded speed, using thresholds of 85, 90 and 95 percent, and counted two ways: how recently he last hit that speed, and how often he hit it in the previous 7 and 14 days.

Fifteen lower-body soft-tissue injuries turned up, most often in the lower leg (six). Only six of them could be matched to a usable healthy comparison period. Some injuries happened before GPS monitoring started, and one player's healthy comparison date landed on another injury. None of the sprint measures separated injury periods from healthy periods at any threshold. The authors then calculated how big an effect their data could have detected, and the answer was sobering: for the frequency measures, only associations increasing the odds of injury roughly 16 to 25 times over would have shown up. That is not a finding of safety. It is a finding of not enough information.

One incidental discovery deserves attention. In 39 percent of sessions, the day and month in the exported dates had been swapped, an error invisible to the eye and only caught by checking against the club's published fixture list.

The practical lesson is that a single club, however diligent, will rarely have enough injuries to build its own prediction model. Answering this properly needs several teams pooling data.

• No individualized HVR exposure metric distinguished injury periods from matched within-athlete control periods at any threshold.
• At the 85% threshold: recency OR 1.08 (95% CI 0.88–1.32, p = 0.450; 4 pairs, 3 informative), 7-day frequency OR 1.00 (95% CI 0.14–7.10, p = 1.000; 6 pairs, 4 informative), 14-day frequency OR 3.00 (95% CI 0.31–28.84, p = 0.341; 6 pairs, 4 informative).
• At the 95% threshold, recency OR was 1.15 (95% CI 0.61–2.16, p = 0.662; 6 pairs, 2 informative).
• Four of nine unpenalized models showed quasi-complete separation and were not interpreted; the 14-day frequency model at 95% had no informative strata and was not estimable.
• Firth-penalized estimation returned finite estimates for eight of nine models, but no penalized interval excluded the null and informative-strata counts were unchanged.
• Minimum detectable odds ratios ranged from 1.33 (recency, 85%) to 25.41 (14-day frequency, 85%), so only very large associations could have been detected; frequency metrics required ORs of roughly 16–25.