arxiv: 2604.14534 · v1 · submitted 2026-04-16 · 💻 cs.LG · stat.AP

Recognition: unknown

An unsupervised decision-support framework for multivariate biomarker analysis in athlete monitoring

Fernando Barcelos Rosito , Sebasti\~ao De Jesus Menezes , Simone Ferreira Sturza , Adriana Seixas , Muriel Figueredo Franco

Authors on Pith no claims yet

Pith reviewed 2026-05-10 12:18 UTC · model grok-4.3

classification 💻 cs.LG stat.AP

keywords unsupervised clusteringathlete monitoringmultivariate biomarkersphysiological statesmechanical damagemetabolic stressdecision support

0 comments

The pith

An unsupervised framework distinguishes mechanical damage from metabolic stress in athletes using multivariate biomarker profiles without injury labels.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a framework that applies unsupervised clustering to multiple biomarkers collected from amateur soccer players during training. This addresses the common problems of small sample sizes, mixed data scales, and lack of confirmed injury records that limit standard risk models. By working in the combined biomarker space, the method finds groups that align with mechanical damage, metabolic stress, and normal balance. It shows through added synthetic data that the groups stay stable and can catch subtle risks that single-biomarker checks overlook. Such an approach could support better individualized decisions in sports medicine where labeled data is scarce.

Core claim

The framework identifies coherent profiles that distinguish mechanical damage from metabolic stress while preserving homeostatic states by using Ward hierarchical clustering for monitoring and Gaussian Mixture Models for stability analysis in the joint biomarker space, learned from real data of amateur soccer players and validated with synthetic augmentation.

What carries the argument

the modular computational framework operating in the joint biomarker space with preprocessing, safety screening, unsupervised clustering, and centroid-based interpretation

Load-bearing premise

The unsupervised clusters correspond to clinically meaningful physiological states that can be interpreted as mechanical damage versus metabolic stress without any ground truth labels or independent clinical validation.

What would settle it

Independent clinical validation data showing injury types for the same athletes would falsify the claim if athletes clustered as mechanical damage do not exhibit higher mechanical injury rates than those in other clusters.

read the original abstract

Purpose. Athlete monitoring is constrained by small cohorts, heterogeneous biomarker scales, limited feasibility of repeated sampling, and the lack of reliable injury ground truth. These limitations reduce the interpretability and utility of traditional univariate and binary risk models. This study addresses these challenges by proposing an unsupervised multivariate framework to identify latent physiological states in athletes using real data. Methods. We propose a modular computational framework that operates in the joint biomarker space, integrating preprocessing, clinical safety screening, unsupervised clustering, and centroid-based physiological interpretation. Profiles are learned exclusively from amateur soccer players during a competitive microcycle. Synthetic data augmentation evaluates robustness and scalability. Ward hierarchical clustering supports monitoring and etiological differentiation, while Gaussian Mixture Models (GMM) enable structural stability analysis in high-dimensional settings. Results. The framework identifies coherent profiles that distinguish mechanical damage from metabolic stress while preserving homeostatic states. Synthetic data augmentation demonstrates feasibility and detection of latent silent risk phenotypes typically missed by univariate monitoring. Structural analyses indicate robustness under augmentation and higher-dimensional settings. Conclusion. The framework enables interpretable identification of latent physiological states from multivariate biomarker data without injury labels. By distinguishing mechanisms and revealing silent risk patterns not captured by conventional monitoring, it provides actionable insights for individualized athlete monitoring and decision making.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

A modular unsupervised pipeline for athlete biomarkers that applies standard clustering but leaves the mechanical-vs-metabolic state labels without independent checks.

read the letter

The paper puts together a modular pipeline that preprocesses multivariate biomarkers, applies clinical safety screening, runs Ward clustering for monitoring and GMM for stability checks, and then interprets the resulting profiles as mechanical damage, metabolic stress, or homeostatic states. It works with real microcycle data from amateur soccer players and adds synthetic augmentation to probe scalability. This framing directly tackles the practical headaches of small cohorts, mixed scales, and absent injury labels, which is the main thing it gets right.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes a modular unsupervised framework for multivariate biomarker analysis in athlete monitoring. It combines preprocessing, clinical safety screening, Ward hierarchical clustering, and Gaussian Mixture Models (GMM) applied to real data from amateur soccer players over a competitive microcycle, supplemented by synthetic data augmentation. The central claim is that the resulting clusters identify coherent latent physiological profiles that distinguish mechanical damage from metabolic stress while preserving homeostatic states, enabling detection of silent risk phenotypes without requiring injury ground-truth labels.

Significance. If the unsupervised clusters can be shown to map reliably to distinguishable physiological regimes, the framework would address key limitations in athlete monitoring (small cohorts, heterogeneous scales, no labels) by providing an interpretable, multivariate alternative to univariate risk models. The modular design, use of both real microcycle data and synthetic augmentation for robustness testing, and structural stability analysis in higher dimensions are constructive elements. However, the current lack of quantitative validation metrics and independent confirmation of the post-hoc physiological interpretations substantially limits the immediate impact and generalizability.

major comments (3)

[Abstract, Results] Abstract and Results: The reported outcomes are described only qualitatively (e.g., 'coherent profiles that distinguish mechanical damage from metabolic stress') with no quantitative metrics such as silhouette scores, Davies-Bouldin indices, cluster stability measures, or statistical comparisons of centroid differences. This absence makes it impossible to evaluate the strength or reproducibility of the claimed distinctions.
[Methods, Results] Methods and Results: The mapping of Ward/GMM clusters to specific physiological states (mechanical damage vs. metabolic stress vs. homeostatic) is performed post-hoc via inspection of biomarker centroid elevations. No cross-validation against external criteria (expert annotation, longitudinal injury records, or controlled experiments) is described, so the etiological differentiation rests on domain assumptions rather than independent evidence.
[Abstract] Abstract: Synthetic data augmentation is invoked to demonstrate robustness and detection of 'latent silent risk phenotypes,' yet no details are supplied on augmentation parameters, generation process, or how the augmented data were used to validate (or potentially bias) the observed profiles. This leaves open the possibility of circularity in the robustness claims.

minor comments (2)

[Methods] The manuscript would benefit from explicit statements of the number of clusters/components selected, the criteria used for selection, and any sensitivity analyses around this choice.
[Methods] Notation for the biomarker variables and the exact preprocessing steps (normalization, safety screening thresholds) could be presented more formally, perhaps in a dedicated table or pseudocode block, to improve reproducibility.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for their constructive comments, which highlight opportunities to strengthen the quantitative aspects and transparency of our work. We address each major comment below and indicate planned revisions.

read point-by-point responses

Referee: [Abstract, Results] Abstract and Results: The reported outcomes are described only qualitatively (e.g., 'coherent profiles that distinguish mechanical damage from metabolic stress') with no quantitative metrics such as silhouette scores, Davies-Bouldin indices, cluster stability measures, or statistical comparisons of centroid differences. This absence makes it impossible to evaluate the strength or reproducibility of the claimed distinctions.

Authors: We agree that quantitative metrics are needed for rigorous evaluation. In the revised manuscript, we will add silhouette scores, Davies-Bouldin indices, cluster stability measures via bootstrap resampling, and statistical comparisons (e.g., Kruskal-Wallis tests with post-hoc analyses) of biomarker centroid differences to objectively support the reported distinctions. revision: yes
Referee: [Methods, Results] Methods and Results: The mapping of Ward/GMM clusters to specific physiological states (mechanical damage vs. metabolic stress vs. homeostatic) is performed post-hoc via inspection of biomarker centroid elevations. No cross-validation against external criteria (expert annotation, longitudinal injury records, or controlled experiments) is described, so the etiological differentiation rests on domain assumptions rather than independent evidence.

Authors: The post-hoc mapping follows standard practice for unsupervised methods in the absence of ground truth, which is a core motivation of the study. Interpretations draw on established biomarker-physiology links from the literature. We will expand the Methods and Discussion with additional references and sensitivity analyses for robustness. However, cross-validation against injury records or expert annotations is not feasible, as these data were not collected, reflecting the real-world constraints the framework targets. revision: partial
Referee: [Abstract] Abstract: Synthetic data augmentation is invoked to demonstrate robustness and detection of 'latent silent risk phenotypes,' yet no details are supplied on augmentation parameters, generation process, or how the augmented data were used to validate (or potentially bias) the observed profiles. This leaves open the possibility of circularity in the robustness claims.

Authors: The Methods section already details the augmentation parameters and process. To enhance clarity, we will briefly summarize these in the Abstract and add a Results subsection explicitly describing the generation method, parameters, application to robustness testing, and validation metrics to demonstrate independence from the original profiles. revision: yes

standing simulated objections not resolved

Independent cross-validation of physiological interpretations against external criteria such as injury records or expert annotations, which are unavailable in the study dataset and inherent to the unsupervised, label-free setting.

Circularity Check

0 steps flagged

No circularity detected in derivation chain

full rationale

The paper's framework applies standard unsupervised clustering (Ward, GMM) to preprocessed biomarker data from real athlete microcycles, followed by post-hoc centroid inspection to assign physiological labels. This interpretation step relies on external domain knowledge of biomarker meanings rather than deriving the mechanical-vs-metabolic distinction mathematically from the clustering equations themselves. No equations reduce a claimed prediction to a fitted input by construction, no uniqueness theorems are imported via self-citation, and synthetic augmentation is used only for robustness checks, not to generate the primary profiles. The central claim therefore remains an empirical partitioning plus interpretive overlay, not a self-referential loop.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The framework rests on standard unsupervised clustering assumptions and domain knowledge for state interpretation, with synthetic data used for robustness checks but no new entities postulated.

free parameters (1)

Number of clusters or GMM components
Critical choice for profile identification but unspecified in the abstract.

axioms (1)

domain assumption Unsupervised clusters represent distinct, interpretable physiological states
Invoked to assign meaning to mechanical damage, metabolic stress, and homeostatic profiles without ground truth.

pith-pipeline@v0.9.0 · 5534 in / 1071 out tokens · 48762 ms · 2026-05-10T12:18:46.133034+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

20 extracted references · 13 canonical work pages

[1]

The role of biomarkers in monitoring chronic fatigue among male pro- fessional team athletes: A systematic review,

G. Contreras-D´ ıaz, A. Galiano, J. M. Garc´ ıa- Mansoet al., “The role of biomarkers in monitoring chronic fatigue among male pro- fessional team athletes: A systematic review,” Sensors, Vol. 24, p. 6862, 2024

2024
[2]

Blood-based biomarkers for man- aging workload in athletes: Considerations and recommendations for evidence-based use of established biomarkers,

N. Haller, M. Behringer, T. Reichel et al., “Blood-based biomarkers for man- aging workload in athletes: Considerations and recommendations for evidence-based use of established biomarkers,”Sports Medicine, Vol. 53, No. 7, pp. 1315–1333, 2023. [Online]: https://doi.org/10.1007/s40279-023-01836-x

work page doi:10.1007/s40279-023-01836-x 2023
[3]

Biochemical and perceptual markers of physiological stress during acute exercise overload in u20 elite basketball players,

J. M. L´ opez-Cuervo, A. Rojas-Jaramillo, A. Garc´ ıa-Caro, J. Gonz´ alez-Santamaria, G. Humeres, J. R. Stout, A. Odriozola- Mart´ ınez, and D. A. Bonilla, “Biochemical and perceptual markers of physiological stress during acute exercise overload in u20 elite basketball players,”Stresses, Vol. 5, No. 3, p. 52, 2025. [Online]: https://doi.org/ 10.3390/stre...

work page doi:10.3390/stresses5030052 2025
[4]

Hematology, hormones, inflammation, and muscle damage in elite and professional soccer players: a systematic review with implications for exercise,

K. Saidi, A. B. Abderrahman, A. C. Hackney, B. Bideau, S. Zouita, U. Granacher, and H. Zouhal, “Hematology, hormones, inflammation, and muscle damage in elite and professional soccer players: a systematic review with implications for exercise,”Sports medicine, Vol. 51, No. 12, pp. 2607– 2627, 2021. [Online]: https://doi.org/10. 3390/ijerph21111394

2021
[5]

Overtraining syndrome: A complex systems phenomenon,

L. E. Armstrong, J. L. VanHeest, H. O’Connor, W. J. Kraemer, and R. Meeusen, “Overtraining syndrome: A complex systems phenomenon,”Frontiers in Network Physiology, Vol. 1, p. 794392, 2022. [Online]: https://doi.org/10.3389/fnetp.2022. 794392

work page doi:10.3389/fnetp.2022 2022
[6]

Digital medicine and the curse of dimensionality,

V. Berisha, C. Krantsevich, P. R. Hahn, S. Hahn, G. Dasarathy, P. Turaga, and J. Liss, “Digital medicine and the curse of dimensionality,”npj Digital Medicine, Vol. 4, No. 1, p. 153, 2021. [Online]: https://doi.org/10.1038/s41746-021-00521-5

work page doi:10.1038/s41746-021-00521-5 2021
[7]

Machine learning methods in sport injury prediction and prevention: A systematic review,

H. Van Eetvelde, L. D. Mendon¸ ca, C. Ley, R. Seil, and T. Tischer, “Machine learning methods in sport injury prediction and prevention: A systematic review,”Journal of Experimental Orthopaedics, Vol. 8, No. 1, p. 27, 2021. [Online]: https://doi.org/10. 1186/s40634-021-00346-x

2021
[8]

Blood sample profile helps to injury forecasting in elite soccer players,

A. Rossi, L. Pappalardo, C. Filetti, and P. Cintia, “Blood sample profile helps to injury forecasting in elite soccer players,” Sport Sciences for Health, Vol. 19, No. 1, pp. 285–296, 2023. [Online]: https://doi.org/10. 1007/s11332-022-00932-1

2023
[9]

Lim- itations of separating athletes into high or low-risk groups based on a cut-off: A clinical commentary,

J. M. Losciale, L. K. Truong, P. Ward, G. S. Collins, and G. S. Bullock, “Lim- itations of separating athletes into high or low-risk groups based on a cut-off: A clinical commentary,”International Jour- nal of Sports Physical Therapy, Vol. 19, No. 9, pp. 1151–1164, 2024. [Online]: https://doi.org/10.26603/001c.122644

work page doi:10.26603/001c.122644 2024
[10]

Machine learning approaches to injury risk pre- diction in sport: A scoping review with evidence synthesis,

C. Leckey, N. van Dyk, C. Doherty, A. Lawlor, and E. Delahunt, “Machine learning approaches to injury risk pre- diction in sport: A scoping review with evidence synthesis,”British Journal of Sports Medicine, Vol. 59, No. 7, p. e108576, 2024. [Online]: https://doi.org/10. 1136/bjsports-2024-108576

2024
[11]

A hierarchical clustering approach for examining potential risk factors for bone stress injury in runners,

J. A. Martin, M. R. Stiffler-Joachim, C. M. Wille, and B. C. Heiderscheit, “A hierarchical clustering approach for examining potential risk factors for bone stress injury in runners,”Journal of Biomechanics, Vol. 141, p. 111136, 2022. [Online]: https://doi.org/ 10.1016/j.jbiomech.2022.111136

work page doi:10.1016/j.jbiomech.2022.111136 2022
[12]

Using unsupervised machine learning to characterize recovery patterns in elite canoe- kayak athletes across the Olympic training year,

A. Foucaud, F. Durand, and H. Meric, “Using unsupervised machine learning to characterize recovery patterns in elite canoe- kayak athletes across the Olympic training year,”Frontiers in Sports and Active Living, Vol. 7, p. 1629924, 2025. [Online]: https: //doi.org/10.3389/fspor.2025.1629924

work page doi:10.3389/fspor.2025.1629924 2025
[13]

Blood biomarker profiling and monitoring for high-performance physiology and nutri- tion: current perspectives, limitations and recommendations,

C. R. Pedlar, J. Newell, and N. A. Lewis, “Blood biomarker profiling and monitoring for high-performance physiology and nutri- tion: current perspectives, limitations and recommendations,”Sports Medicine, Vol. 49, No. Suppl 2, pp. 185–198, 2019. [Online]: https://doi.org/10.1007/s40279-019-01158-x 14

work page doi:10.1007/s40279-019-01158-x 2019
[14]

Position-specific biomarker responses to match vs. VAMEVAL test modalities in elite female soccer players,

A. G. Souglis, G. C. Bogdanis, C. Chryssan- thopoulos, N. Apostolidis, and N. D. Geladas, “Position-specific biomarker responses to match vs. VAMEVAL test modalities in elite female soccer players,”Cogent Engi- neering, Vol. 11, No. 1, p. 2331188, 2024. [Online]: https://doi.org/10.1080/23311886. 2024.2447399

work page doi:10.1080/23311886 2024
[15]

Unsupervised clustering of biochemical markers reveals health profiles associ- ated with function and survival in active aging,

R. Gonz´ alez-Martos, J. Galeano, C. Ramirez- Castillejo, N. Gusi, E. Gesteiro, G. Vicente- Rodriguez, I. Ara, and A. Guadalupe-Grau, “Unsupervised clustering of biochemical markers reveals health profiles associ- ated with function and survival in active aging,”Scientific Reports, Vol. 15, No. 1, p. 30546, 2025. [Online]: https: //doi.org/10.1038/s41598-...

work page doi:10.1038/s41598-025-14580-1 2025
[16]

Classifying soccer players based on physical capacities and match- specific running performance using machine learning,

D. Popczyk, “Classifying soccer players based on physical capacities and match- specific running performance using machine learning,”Journal of Sports Science and Medicine, Vol. 24, pp. 1–12, 2025. [Online]: https://doi.org/10.52082/jssm.2025.764

work page doi:10.52082/jssm.2025.764 2025
[17]

J. F. Hair, W. C. Black, B. J. Babin, and R. E. Anderson,Multivariate Data Analy- sis, 8th Edt. Hampshire, Cengage Learning, 2019

2019
[18]

Synthetic data in biomedicine via generative artificial intel- ligence,

B. van Breugel, T. Liu, D. Oglic, and M. van der Schaar, “Synthetic data in biomedicine via generative artificial intel- ligence,”Nature Reviews Bioengineering, Vol. 2, No. 12, pp. 991–1004, 2024. [Online]: https://doi.org/10.1038/s44222-024-00245-7

work page doi:10.1038/s44222-024-00245-7 2024
[19]

Review of generative ai for synthetic data generation: a health- care perspective,

H. M. Waseem, S. U. Islam, N. Matragkas, G. Epiphaniou, T. N. Arvanitis, and C. Maple, “Review of generative ai for synthetic data generation: a health- care perspective,”Artificial Intelligence Review, 2025. [Online]: https://doi.org/10. 1007/s10462-025-11440-2

2025
[20]

Indi- vidualised dietary strategies for Olympic combat sports: Acute weight loss, recov- ery and competition nutrition,

R. Reale, G. Slater, and L. M. Burke, “Indi- vidualised dietary strategies for Olympic combat sports: Acute weight loss, recov- ery and competition nutrition,”European Journal of Sport Science, Vol. 17, No. 6, pp. 727–740, 2024. [Online]: https: //doi.org/10.1080/17461391.2017.1297489 15

work page doi:10.1080/17461391.2017.1297489 2024