Competitive Runners · Training History & Injury Risk
A longitudinal, athlete-aware reanalysis of public competitive-running training histories. I audited 74 masked athletes and 42,766 selected event windows, then evaluated training and recovery signals using temporal, athlete-disjoint and combined validation. An interactive bilingual site exposes the trajectories, modest predictive signal and review burden.
Context
Longitudinal training data introduce time, overlapping histories, individual baselines, rare labels and participant dependence. These challenges matter before choosing an algorithm.
Problem
What changed in an athlete’s training before a retained injury flag, and how much ranking signal survives strict temporal and athlete separation?
My role
Independent data audit, corrected lag interpretation, feature design, statistical description, validation, calibration, reproducible code, literature discussion and interactive communication. The original researchers collected the measurements.
What I built
Eleven notebooks, checksummed acquisition, past-only personal-history audit, regularized logistic baselines, independent calibration and a matched 14/21-day extension with athlete weighting, fixed forest and partition stability. The website explores saved held-out outputs; it does not run a clinical model.
Data or evidence used
Primary unseen-athlete/future evaluation: AP 0.0200 versus test label fraction 0.0161. At a 10% calibration review budget, the held-out sample requires about 60 reviews per captured positive label. All comparisons and a cited narrative literature discussion are public.
Limitations
One competitive Dutch team, creator-selected windows, future-conditioned healthy controls, relative dates and incomplete clinical ascertainment. No causal intervention effects, clinical thresholds, true athlete-day alert rate or validated event lead time can be established. SoccerMon and Wu data are not pooled with these runners.
Next validation step
Prospective continuous surveillance with complete clinical outcomes, recording-availability timestamps and independent external cohorts before considering operational decisions.
Source data and published analysis
- Lövdal, den Hartigh & Azzopardi. Replication Data, Dataverse v2.0. DOI 10.34894/UWU9PV. ↗
Original measurement source, CC0. Credited separately from this analysis.
- Lövdal et al. (2021). Injury Prediction in Competitive Runners With Machine Learning. IJSPP. DOI 10.1123/ijspp.2020-0518. ↗
Original study, with a separate 2024 erratum. Article rights are distinct from data rights.