When Minute-Resolution Monitoring Meets Session-Level Injury Labels: Landmark-Based Discrimination in Elite Women's Football
Evangelos Chatzidimitriou, Konstantinos Tserpes
cs.LG
Quick verdict
A landmark framework for session-level injury labels with minute-level monitoring; honest and robust, but five positive athletes cap the conclusions.
Summary
The paper tackles a unit-of-analysis problem in sports-injury ML: monitoring data arrive at minute resolution while injury labels mark only an athlete-session (day). Instead of copying session labels onto minutes, the authors build one representation per athlete-session at six fixed landmarks (10-60 min) using only information observed up to each point, keeping the label as a same-day injury-associated session indicator, and validate with athlete-disjoint folds. Using 2020 SoccerMon data (2,259 sessions, 27 athletes; 22 positives from five athletes, all in Team A), they compare representations (PRE/CUM/DYN), four models (logistic regression, Random Forest, XGBoost, TabPFN), and training-only SMOTE/CTGAN augmentation. Primary CUM+DYN logistic discrimination is non-monotonic (ROC-AUC 0.499-0.607 across landmarks), a shape that survives a fixed 2,104-session common cohort and 100 alternative fold allocations; equal-athlete weighting shifts estimates substantially. TabPFN performs best at 30 min (0.692), beats logistic regression at later landmarks but not consistently Random Forest. SMOTE/CTGAN gains are condition-specific rather than universal. The decisive limitation is that all positives stem from five athletes, capping precision and external validity. Code is public; raw data require third-party access.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A unit-aligned landmark framework - one representation per athlete-session at fixed 10-60 min landmarks using only information observed up to each point - avoids fabricating minute-level supervision when injury labels exist only at session level.
- Primary CUM+DYN logistic-regression discrimination is landmark-dependent and non-monotonic (ROC-AUC 0.499, 0.557, 0.607, 0.428, 0.402, 0.367 across 10-60 min), a shape that persists on a fixed 2,104-session common cohort and across 100 alternative fold allocations.
- TabPFN improves later-landmark discrimination relative to logistic regression (paired ROC-AUC contrasts exclude zero at 40-60 min; 0.692 at 30 min) but does not consistently outperform Random Forest.
- Training-only SMOTE/CTGAN augmentation provides condition-specific rather than universal benefit (e.g., Random-Forest+SMOTE at 30 min positive; logistic-regression+CTGAN negative at 30-40 min), and the CTGAN audit shows substantial distributional drift despite no near-duplicate memorization.
- Within this cohort, inferential precision is limited by only five injury-positive athletes, and the paper explicitly claims no minute-specific prediction, onset localization, or operational alerting capability.
Strengths
The unit-of-analysis problem is framed correctly and attacked at its root
Section 1/3.1 refuse to replicate session labels onto minutes; Sec. 3.1 explicitly restricts interpretation to a same-day discrimination score, and the design is positioned against event-history landmarking and multiple-instance learning without conflating them.
Robustness battery unusual for this literature
Fixed 2,104-session common cohort, 100 alternative negative-athlete fold allocations, leave-one-positive-athlete-out, and equal-athlete weighting (Sec. 3.5, 4.1); the non-monotonic ROC-AUC shape (0.499/0.557/0.607/0.428/0.402/0.367) persists in all of them.
Calibrated claims with paired uncertainty
Genuine synthetic-data audit
Reproducibility discipline
Weaknesses
Effective sample is five athletes, not 380k rows
No external or out-of-cohort validation of any kind
Landmark-trajectory uncertainty is wide and the central peak is not formally tested
Best-performing feature family is demoted on data-quality grounds
Augmentation grid is exploratory with direction instability
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This arXiv preprint (cs.LG) addresses a unit-of-analysis problem in sports-injury ML: minute-resolution monitoring predictors coexist with injury labels that only identify an injury-associated athlete-session (day), with unknown within-session onset. The authors' answer is a landmark-style framework: at six fixed elapsed-time landmarks (10, 20, 30, 40, 50, 60 min), each athlete-session is summarized once, using only information observed up to that session-clock point; the label stays a same-day session indicator (Sec. 3.1). The paper is explicit about what it is not: no minute-specific prediction, no onset localization, no operational alerting.
Data are the 2020 season of SoccerMon (elite Norwegian women's football; Sec. 3.2, Table 1): 380,193 minute rows, 3,743 sessions, 48 athletes; the modelling cohort is 2,259 Team-A sessions from 27 athletes, because all 22 positives come from five athletes in Team A (Team B: 1,484 sessions, zero positives). Comparisons: representations PRE (14 contextual), CUM (30 cumulative), DYN (33 dynamic; primary CUM+DYN = 63), plus combined families up to 77 features; learners logistic regression (primary), Random Forest, XGBoost, TabPFN; training-only SMOTE/CTGAN augmentation (3 ratios, 5 seeds; Sec. 3.7).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What problem does this paper solve in sports injury prediction?
It addresses the label-resolution mismatch: wearable data are recorded every minute, but injury labels often only say that a given athlete on a given day was injured, without the within-session onset time. Copying the day-level label onto every minute would fabricate supervision. The paper instead summarizes each athlete-session once at six fixed elapsed-time landmarks (10-60 minutes), using only information observed up to each landmark, and keeps the label at session level.
What are the main results and how strong are they?
On 2020 SoccerMon data from elite Norwegian women's football (2,259 sessions, 27 athletes), discrimination was non-monotonic across landmarks: primary logistic-regression ROC-AUC rose to about 0.607 at 30 minutes and fell to 0.367 at 60 minutes. TabPFN reached the single best value (0.692 at 30 min) and beat logistic regression at later landmarks but not consistently Random Forest. SMOTE/CTGAN augmentation helped only in specific model-landmark combinations. A key caveat: all 22 positive sessions came from just five athletes, so intervals are wide and the results are cohort-specific, not deployment-ready performance.
Why do only five athletes matter if there are 380,193 data rows?
Minute rows are highly dependent within an athlete; the number of independent injury-positive individuals is what drives statistical power. All 22 injury-associated sessions came from five athletes on one team (the second team had none), so precision, stability and external validity are limited no matter how dense the sensor data are - a point the paper itself emphasizes.
Can I reproduce this study?
Partially. The full analysis pipeline is public on GitHub (minute-level-injury-risk-monitoring), and the authors report a frozen rerun reproducing TabPFN metrics to differences below 4e-7. However, the raw SoccerMon data are not redistributed; you must request access from the dataset curators, so the headline numbers cannot be regenerated from the paper and code alone.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.