Literature review and discussion

Search and selection

Targeted searches were run in PubMed and publisher/journal records for running-related injury, running biomechanics, prospective cohorts, systematic reviews/meta-analyses, and running speed/kinematics. We prioritized human studies that inform four design decisions: distinguish current injury status from future incidence; avoid a universal gait marker; describe population and diagnosis; and account for speed. The dataset descriptor is included. Searches were targeted, not exhaustive; no dual-reviewer screening or formal risk-of-bias update was performed.

Evidence and project interpretation

No universal gait marker

Lopes, Mascarinas and Hespanhol synthesized 82 studies involving 5,465 runners. Peak hip adduction was the only variable associated with injury in pooled analysis, and the estimate was strongly influenced by two of five studies. They described the evidence as minimal and highlighted inconsistent definitions and methods. This supports a predeclared feature registry and cautious interpretation.

Population matters

Vannatta, Heinert and Kernozek found 13 eligible prospective studies but only five unique running samples. Pooled findings differed by sex and competition level; moderate evidence for some factors was specific to female recreational runners. Associations should not be transferred automatically to this source cohort.

Prospective incidence differs from current status

Ramskov and colleagues followed 224 injury-free recreational runners for 52 weeks and recorded medically diagnosed injuries. Some baseline factors were associated with subsequent injury. This prospective design answers a different question from a cross-sectional comparison of reported current injury status. The present project cannot establish temporal order or future risk.

Speed can change measured movement

Orendurff and colleagues compared four speeds and found changes in peak lower-limb kinematics and kinetics. Candidate-group speeds in this dataset differ (2.739 m/s healthy; 2.554 m/s injured), so future comparisons should report overlap and predefine speed-aware analyses. These metadata means are descriptive, not evidence of a causal injury effect.

Event parity is not event accuracy

Reprocessing with the official MATLAB workflow, when available, can test computational parity with the source. It cannot independently establish event accuracy; that requires manually adjudicated events or another reference. The 30-person pilot currently has structural checks and source-provided summaries only.

Recent work on the same dataset

Cross-sectional classification has already been studied

Xiao, Li and Dong (2026) applied five machine-learning classifiers to 1,744 runners from the Running Injury Clinic dataset. Their random-forest test ROC-AUC was 0.790 (95% CI 0.747–0.831), sensitivity 0.905 and specificity 0.523. They note a train/test performance gap and call for external validation and subtype-specific work. This makes another generic classifier a weak novelty claim. Our complementary analysis audits repeated-record labels and cohort-time definitions before future modeling; it is descriptive and does not establish clinical transitions.

Xiao Y, Li J, Dong H. Machine learning-based classification of running-related injury status: a cross-sectional study. BMC Sports Science, Medicine and Rehabilitation (2026). DOI: 10.1186/s13102-026-02026-8 ↗

Practical implications

1. Target current reported injury status, not future incidence or causal risk.
2. Do not widen the biomechanical cohort before official event/cycle processing is executed and reviewed.
3. Report source definitions, units, signs, phase windows, missingness, quality and participant-level aggregation.
4. Describe the 30-person balanced sample as a technical pilot, not a representative or clinical sample.
5. Inspect speed/covariate overlap before adjustment; do not extrapolate outside common support.
6. Account for repeated trials and multiple feature testing; treat exploratory subgroup contrasts as hypotheses.
7. Do not derive screening thresholds, diagnosis, treatment recommendations or future-risk claims.

References

1. Ferber R, Brett K, Fukuchi R, Hettinga B, Osis S. A dataset of running biomechanics and clinical measures of runners with and without running-related injuries. Scientific Data. 2024. doi:10.1038/s41597-024-04011-7.
2. Ferber R, Brett K, Fukuchi R, Hettinga B, Osis S. Running Injury Clinic Kinematic Dataset, version 1. Figshare. doi:10.25452/figshare.plus.24255795.v1.
3. Lopes AD, Mascarinas A, Hespanhol L. Are alterations in running biomechanics associated with running injuries? A systematic review with meta-analysis. Brazilian Journal of Physical Therapy. 2023;27(4):100538. doi:10.1016/j.bjpt.2023.100538 · PubMed.
4. Vannatta CN, Heinert BL, Kernozek TW. Biomechanical risk factors for running-related injury differ by sample population: a systematic review and meta-analysis. Clinical Biomechanics. 2020;75:104991. doi:10.1016/j.clinbiomech.2020.104991 · PubMed.
5. Ramskov D, et al. Associations between biomechanical and clinical/anthropometrical factors and running-related injuries among recreational runners: a 52-week prospective cohort study. Injury Epidemiology. 2020. doi:10.1186/s40621-020-00237-2.
6. Orendurff MS, Kobayashi T, Tulchin-Francis K, et al. A little bit faster: lower extremity joint kinematics and kinetics as recreational runners achieve faster speeds. Journal of Biomechanics. 2018;71:167–175. doi:10.1016/j.jbiomech.2018.02.010.
7. Wolski A, et al. Utilising inertial measurement units and force-velocity profiling to explore sprint injury risk: a feasibility study. Sensors. 2025;25(5):1518. doi:10.3390/s25051518.