• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Sunday, November 22, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Health

Machine Learning Shows Modest Accuracy Predicting Tendinopathy Risk in Athletes

by
November 21, 2026
in Health
Reading Time: 5 mins read
0
Machine Learning Shows Modest Accuracy Predicting Tendinopathy Risk in Athletes

Machine Learning Shows Modest Accuracy Predicting Tendinopathy Risk in Athletes

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

A prospective cohort study tracking 56 student-athletes over a competitive season found that machine-learning models could achieve modest internal discrimination for predicting incident tendinopathy, though substantial statistical uncertainty and small sample sizes limit their clinical deployment. Researchers evaluated 58 distinct model and class-imbalance configurations using repeated stratified nested cross-validation, finding that tree-based ensembles demonstrated the most consistent internal reliability. Specifically, Random Forest algorithms combined with synthetic oversampling strategies achieved the strongest diagnostic stability, yet their overall predictive precision underscored that algorithmic screening in low-data sports medicine settings remains strictly exploratory rather than immediately applicable to routine clinical practice.

Tendinopathy represents a common, load-related musculoskeletal disorder characterized by localized tendon pain, swelling, and impaired mechanical function, presenting a substantial burden across elite and recreational sporting domains. Clinical manifestations arise through complex interactions between repetitive mechanical loading, tendon matrix disorganization, impaired physiological adaptation, biomechanical dynamics, and individual biological susceptibility. The clinical challenge is particularly evident in high-load jumping sports; previous cross-sectional investigations have identified jumper’s knee in 14.2% of elite athletes overall, with prevalence reaching 44.6% in volleyball players and 31.9% in basketball players, while non-elite cohorts demonstrate an overall prevalence of 8.5%. Despite this prevalence, early diagnosis and risk stratification remain difficult because tendinopathy lacks a singular structural or biochemical biomarker, and imaging abnormalities such as tendon thickening or neovascularity can frequently appear in asymptomatic individuals without correlating directly with symptoms.

To investigate whether multi-algorithmic frameworks could identify athletes at heightened risk prior to symptom onset, investigators recruited participants across two distinct settings governed by an identical overarching protocol. The study enrolled student-athletes registered across 16 university sports clubs at the University of East London in the United Kingdom, alongside male and female athletes registered across four under-21 professional sports teams in Riyadh, Saudi Arabia. All participants were aged 18 to 21 years and actively competing in organized athletics at baseline. To isolate incident cases and limit measurement confounding, participants were required to be completely pain-free at enrolment with no documented history of musculoskeletal pathology within the preceding three months, while athletes presenting with chronic medical conditions or non-sport-related disorders were excluded.

At the baseline assessment visit, investigators documented a structured battery of candidate predictors grouped into four prespecified clinical domains: individual and sport-related characteristics, anthropometric parameters, neuromuscular performance markers, and psychological or recovery constructs. The individual domain captured chronological age, biological sex, ethnicity, employment status, nicotine consumption, primary sport type, player position, leg dominance, and prior seasonal tendon complaints. Sleep parameters were catalogued using the self-administered Athlete Sleep Screening Questionnaire, which produced a composite Sleep Difficulty Score alongside individual indices of total sleep duration, sleep-quality satisfaction, sleep latency, maintenance difficulties, and medication use, employing the validated Arabic translation for Saudi participants. Athlete burnout was evaluated across 15 Likert-scaled items from the Athlete Burnout Questionnaire, covering emotional and physical exhaustion, reduced athletic accomplishment, and sport devaluation.

Physical biomechanics were quantified using standardized clinical tools administered under controlled laboratory conditions prior to the competitive season. Dynamic postural stability was captured via the Y-Balance Test during single-leg stance across anterior, posteromedial, and posterolateral directions, performing three trials per limb to normalize maximal reach distance against limb length and generate composite stability scores. Lower-limb joint range of motion was recorded in degrees using calibrated goniometric assessments, measuring hip flexion with the knee both extended and flexed, hip abduction, hip internal and external rotation, knee flexion, and ankle dorsiflexion under extended and flexed knee positions. Over the subsequent competitive season, prospective injury surveillance tracked incident tendinopathy defined strictly by palpable tendon pain confirmed during clinical examinations and corroborated by clinical imaging reports when routine care records were accessible.

The final analytical cohort consisted of 56 athletes, among whom 16 individuals (28.6%) developed incident tendinopathy during seasonal follow-up while 40 participants (71.4%) remained uninjured. The cohort had a median age of 20.19 years, with 52 participants (92.9%) identifying as male. Football was the predominant sport, representing 58.9% of the cohort, followed by basketball at 14.3%, volleyball at 3.6%, handball at 1.8%, and other sporting activities accounting for 21.4%. Prior to model training, a structured missingness audit revealed that all 12 non-questionnaire demographic, anthropometric, range-of-motion, and balance predictors possessed complete data across all 56 athletes. Conversely, six questionnaire-derived recovery metrics exhibited 41.07% missingness, which was addressed using multiple imputation by chained equations with five imputed datasets and 20 iterations, ensuring all imputation operations were strictly isolated within training partitions.

To evaluate predictive capacity without introducing optimistic data leakage, the engineering pipeline implemented repeated nested cross-validation consisting of five repetitions of stratified four-fold outer validation, yielding 20 distinct outer testing folds. Within each training partition, an inner three-fold cross-validation loop conducted randomized hyperparameter optimization, probability calibration via sigmoid mapping, and guarded decision-threshold selection. Rather than adopting an uncalibrated default threshold of 0.5, candidate thresholds were learned using inner out-of-fold predictions constrained to target a positive predictive value of at least 0.60 and a sensitivity guard of at least 0.40, with balanced accuracy maximized among qualifying points. Synthetic oversampling methods, including standard SMOTE and adaptive synthetic sampling, were applied exclusively to training splits, generating roughly 18 synthetic minority event observations per fold to balance training distributions.

Among 58 evaluated configurations spanning linear models, kernel support vector machines, gradient-boosted trees, and tabular deep neural networks, tree-based ensembles demonstrated the highest internal consistency. Random Forest paired with ADASYN achieved the highest mean ranking discrimination, recording an area under the receiver operating characteristic curve of 0.679 (corrected 95% CI, 0.485 to 0.872) and a precision-recall area of 0.604 (95% CI, 0.403 to 0.804). Random Forest with SMOTE attained the highest tuned balanced accuracy at 0.630 (95% CI, 0.509 to 0.751), with a mean sensitivity of 0.525, specificity of 0.735, and an area under the precision-recall curve of 0.625 (95% CI, 0.430 to 0.819). Unresampled Random Forest yielded comparable stability, generating an ROC-AUC of 0.658 and balanced accuracy of 0.624, with both leading ensemble variations maintaining Brier reliability scores of 0.20 and avoiding complete zero-recall classification failures across all 20 outer folds.

Other algorithmic families exhibited lower or more volatile diagnostic trajectories across test partitions. Support vector machines with radial basis kernels utilizing SMOTE produced an ROC-AUC of 0.657 (95% CI, 0.469 to 0.846) and a balanced accuracy of 0.609. Gradient-boosting architectures showed intermediate performance, with LightGBM and CatBoost using ADASYN recording ROC-AUC values of 0.654 and 0.650, respectively. In contrast, complex tabular neural networks suffered notable instability; TabNet without resampling yielded zero true-positive recall in 25% of testing folds, while TabTransformer achieved an ROC-AUC of 0.50. Furthermore, a simple post-hoc clinical benchmark combining age and prior tendinopathy history within identical nested folds achieved a balanced accuracy of 0.646 and an ROC-AUC of 0.599, demonstrating that heavily parameterized machine-learning algorithms did not decisively outperform parsimonious baseline heuristics under sparse-sample constraints.

Model interpretability assessed via SHapley Additive exPlanations revealed that predictive signals were distributed across biomechanical, psychological, and recovery domains rather than concentrated in a single physical metric. Anterior reach distance on the Y-Balance Test for both non-dominant and dominant limbs emerged as the most consistent contributor to model decisions, alongside burnout dimensions including emotional exhaustion and reduced athletic accomplishment. Additional influential features included ankle dorsiflexion with knee extension, hip internal and external rotation ranges, body weight, height, age, and self-reported sleep quality satisfaction. Permutation analyses utilizing 200 label shuffles indicated that classical tree-based models maintained discrimination above conditional chance expectations, whereas deep tabular architectures heavily overlapped with null distributions.

The study authors emphasized several substantial methodological limitations that prevent the direct translation of these predictive models into athletic screening programs. The analytical cohort was constrained to 56 individuals and 16 incident events, producing wide cross-validation confidence intervals that reflect severe statistical uncertainty and sparse-data limitations. Because the cohort comprised 92.9% male participants, sex-specific risk inference was unachievable. Furthermore, the dataset pooled university student-athletes from London with professional under-21 athletes from Riyadh without retaining site indicators for leave-one-site-out domain shift validation, while exposure factors such as match minutes and training load were not recorded. Consequently, the authors concluded that these findings provide exploratory, hypothesis-generating architectural benchmarks, affirming that extensive prospective validation across larger, diverse cohorts is mandatory before automated predictive systems can guide clinical injury prevention.

Subject of Research: Machine-learning frameworks for incident tendinopathy risk prediction in athletes under small-sample and class-imbalanced conditions

Article Title: Exploratory development and internal validation of leakage-resistant, imbalance-aware machine-learning models for incident tendinopathy in athletes

Article References: Elgebaly, A., Muir, D., Hammad, R., Imam, M., Alsulaiman, M. S., & Abdalla, H. S. (2026). Exploratory development and internal validation of leakage-resistant, imbalance-aware machine-learning models for incident tendinopathy in athletes. Frontiers in Medicine, 13, Article 1918715. https://doi.org/10.3389/fmed.2026.1918715

Image Credits: AI Generated

DOI: 10.3389/fmed.2026.1918715

Keywords: tendinopathy, machine learning, random forest, injury prediction, sports medicine, nested cross-validation, class imbalance, SMOTE, ADASYN, Y-Balance Test, athlete burnout, sleep quality, receiver operating characteristic, precision-recall curve, Brier score, SHAP analysis, prospective cohort, student-athletes

News Source: Teresa Odom. (November 21, 2026). Machine Learning Shows Modest Accuracy Predicting Tendinopathy Risk in Athletes. Scienmag.

Tags: ADASYNAthlete BurnoutBrier scoreclass imbalanceInjury predictionMachine Learningnested cross-validationprecision-recall curveprospective cohortRandom Forestreceiver operating characteristicSHAP analysissleep qualitySMOTESports MedicineStudent AthletestendinopathyY Balance Test
Share12Tweet7Share2ShareShareShare1

Related Posts

Loneliness Strongly Associated With Depression Among Older Adults in Golestan Iran

Loneliness Strongly Associated With Depression Among Older Adults in Golestan Iran

November 22, 2026
Self-Powered Neuroprosthesis Harvests Walking Energy to Improve Mobility After Stroke

Self-Powered Neuroprosthesis Harvests Walking Energy to Improve Mobility After Stroke

November 22, 2026

Trifluoroacetate Counterion Reduces Atherosclerosis and Plasma Lipids in Mice via PPAR-α

November 22, 2026

RAGE, Not TLR4, Drives Inflammatory Pain in Human Skin

November 22, 2026

POPULAR NEWS

  • Loneliness Strongly Associated With Depression Among Older Adults in Golestan Iran

    Loneliness Strongly Associated With Depression Among Older Adults in Golestan Iran

    29 shares
    Share 12 Tweet 7
  • Self-Powered Neuroprosthesis Harvests Walking Energy to Improve Mobility After Stroke

    29 shares
    Share 12 Tweet 7
  • Comparative Genomics Unveils Evolutionary Lineages and Stress Genes Across Four Vigna Legumes

    29 shares
    Share 12 Tweet 7
  • Trifluoroacetate Counterion Reduces Atherosclerosis and Plasma Lipids in Mice via PPAR-α

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Tunnel Lining Study Shows Evaporation Accelerates Steel Fibre Corrosion in Cracks

Recycled Geopolymer Powder Can Enhance Slag Geopolymer Strength at Controlled Replacement Levels

Combustion-Synthesized Magnesium Ferrite Catalysts Boost Efficiency Across Biodiesel Blends

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 89 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.