We finished our first full pilot season in May, and we wanted to share what we learned, plainly, before the next one begins. This is not a marketing post. The numbers below come from the same internal review we share with our partner every quarter.
The pilot ran across the full competitive season with one Ligue 2 professional club. Twenty-three players were monitored through the platform. The data sources included daily training load from GPS vest logs, session rating of perceived exertion (sRPE) collected post-session by the medical staff, subjective wellness questionnaires completed each morning, and historical injury records spanning the two prior seasons. Every signal was ingested into a single database and served through Omen's inference engine, which generates a seven-day injury risk score per player using a gradient-boosted ensemble trained on the club's own historical data.
Over the season the staff ran on Omen, non-contact soft-tissue injuries fell by roughly thirty percent compared with their own prior two-season baseline, same squad, same training context. We hold this number with caution: one season, one staff, and a host of confounding variables that range from fixture congestion to the phase of the moon. We are not yet ready to claim a causative trend, only to report what we read. That said, the direction is consistent across every sub-cohort we analysed: first-team and reserve, early season and late, high-minutes players and rotation players alike.
The predictive model flagged elevated seven-day risk an average of 3.2 days before any injury event during the pilot. The most predictive signals, in order, were a sharp increase in acute training load relative to chronic load (ACWR above 1.5), a decline in self-reported sleep quality over three consecutive days, and a recent history of ipsilateral muscle injury within the preceding sixty days. These three factors alone accounted for roughly seventy percent of the predictive weight in the model's alerts. We publish this because we think the field benefits from knowing which signals actually carry weight, not just that a prediction is possible.
What surprised us less, and reassured us more, was the consistency of the human alignment. In more than nine cases out of ten where Omen flagged a player at elevated seven-day risk, at least one member of the medical staff had independently noted concern about the same player in the same week, recorded in their own notes before seeing the platform's alert. We are not introducing new information. We are surfacing what the staff already half-knows, in time to act on it before it becomes a clinical event. This alignment rate is, in our view, the strongest validation of the approach: the model is not inventing false patterns. It is reading the same signals the staff reads, continuously, without gaps.
The unglamorous finding is that the biggest win is not predictive accuracy. It is time. Before the pilot, the medical staff spent roughly four hours per day gathering data from four separate sources — the GPS vendor's portal, the club's internal injury log, a spreadsheet for wellness questionnaires, and paper session notes — and reconciling them manually into a single view. With Omen, that cross-tool reconciliation dropped to under one hour. The staff reported recovering, on average, three hours per person per day. Three hours that were spent, instead, on the players: more one-on-one treatment time, more preparation for the next session, more sitting with a player who needed to talk.
This time savings also changed how the staff used the data itself. With the manual burden reduced, the rate at which staff entered post-session observations more than doubled across the season. More data in meant better predictions out, creating a virtuous cycle that a purely passive monitoring tool cannot produce. The platform became more useful the more the staff used it, because the usage generated the annotations that refined the model.
We are entering the second season with a larger partner group and a more robust measurement framework. The questions we want to answer next are: does the effect replicate across different squads, different staff structures, and different competitive levels? And can the model sustain its alignment rate as the dataset grows across multiple clubs? We will publish the findings either way.
We take on a limited number of new partners each season. If you would like to talk about it, you know where to find us.