“Does it actually work?” — answered with numbers.
Our analysis and our product rest on a pilot in a top-level university environment and on academic research. Every number here comes with its conditions and its limits.
Results
We do not inflate the numbers. The conditions and the limits come with them.
How this compares with ACWR
ACWR (acute:chronic workload ratio) has been the conventional way to flag injury risk. It judges risk from a single ratio — how sharply recent load has risen against the longer-term average.
| Metric | What it means | ACWR (conventional) | Odonata |
|---|---|---|---|
| Precision | Of the players flagged as high risk, the share who went on to be injured | ~5% | ~70% |
| Recall | Of the injuries that occurred, the share flagged in advance | ~20% | ~20% |
Recall stays at the same level as the conventional metric, while precision rises from ~5% to ~70%.
In practical terms: of ten players flagged for attention, fewer than one used to go on to be injured. Now seven do.
How these numbers are calculated
Breakdown of the 100 players validated
- Flagged high risk and injured (true positive): 7
- Flagged high risk but not injured (false positive): 3
- Flagged low risk but injured (missed): 28
- Flagged low risk and not injured (true negative): 62
Precision = 7 ÷ (7 + 3) = ~70%
Recall = 7 ÷ (7 + 28) = ~20%
- Validated on University of Tsukuba Football Club data from 2022–2025.
- Limited to muscular, non-contact ligament and overuse injuries (including stress fractures).
- Within a single model, precision and recall trade off against each other. Raising one tends to lower the other.
- Results will differ with the volume and quality of data, and with the club. We validate and tune on each club’s own data.
How we did it
Combine
Bring wellness, GPS, schedule and injury history together into a single dataset, and generate features across more than 200 views of the data.
Model per club
Build the model on that squad’s own data, training it to recognise states resembling those that preceded past injuries.
Implement
Put it in front of staff, in the app. It returns a daily risk score for each player, together with the factors driving it.
Use it
As one input into the decision, not the decision itself. Staff decide; the model is a second pair of eyes.
The data was there. It just was not connected.
University of Tsukuba Football Club, a top-level university environment, already had the raw material: GPS, wellness reports and medical records. But each sat in a separate place, with no way to read them across the same player. That is where the pilot started, in June 2026.
Research base
The team includes doctoral researchers at the University of Tsukuba, certified clinical psychologists and a physiotherapist.
Psychological network approach
Research that treats a psychological state as a network of interacting symptoms and factors, applied directly to wellness analysis.
Neuroscience, cognitive science and sport science
A background in perception, decision-making and motor control shapes how the product is designed: to support the decision, not to make it.
A team that implements its own research
Nothing is outsourced. The researchers design and build it themselves, which keeps the distance from paper to pitch very short.
This page keeps being updated.
Results from clubs using the platform will be published here, with their conditions and limits.
Tell us what data your club already has.
An initial conversation and a quote cost nothing. “We have data but we are not using it” and “we are struggling with paper and spreadsheets” are both good places to start.