Machine Learning Classification of Serum miR-210 as a Biomarker of Exercise Training Response

MicroRNAs (miRNAs) in circulating blood are increasingly recognized as biomarkers of exercise-induced physiological adaptation. Of these, miR-210, a hypoxiainducible miRNA associated with angiogenesis, mitochondrial control, and cellular stress response, has shown mixed exercise-associated modifications. Its classifier status for individuallevel training response has yet to be established.

Ten young adults (n=10) completed standardized exercise training. Demographic, anthropometric, and training characteristics were recorded, and serum miR-210 expression was quantified using qRT-PCR. Upregulation of miR-210 was the primary outcome, defined as a median divide of expression levels (≥ median = high; < median = low). Predictors were training frequency (sessions/week) and age. Three Machine learning (ML) models-ridge-penalized logistic regression (LR), random forest classifier (RFC), and Support Vector Machine (SVM) were evaluated with Leave-One-Out Cross-Validation (LOOCV). Performance metrics were accuracy, area under the curve (AUC), Brier score, sensitivity, specificity, precision, and calibration.

The variability for serum miR-210 was 0.677 to 1.220 (median 0.782) in five subjects per outcome group. Ordinary LR was 80% accurate, while the ML classifiers performed significantly better: ridge-penalized logistic regression, 90% accuracy (AUC 0.88, sensitivity 1.00, specificity 0.80); random forest, 90% accuracy (AUC 0.92, specificity 1.00, sensitivity 0.80); and support vector machine, 90% accuracy (AUC 0.88, sensitivity 1.00, specificity 0.80). Calibration analysis verified the accuracy of predicted probabilities, and RFC had the most consistent agreement, making it the best-performing model. There were negative, non-significant correlations between miR-210 and frequency of training (r = -0.256) and age (r = -0.117).

ML models identify upregulation of serum miR-210 with ≥90% accuracy, and its utility as a discriminative biomarker of inter-individual exercise responses is hence suggested Conclusion: In this small exploratory cohort (n = 10), machine learning methods detected a distinctive sign in the expression level of miR-210 following exercise training. These findings are preliminary and hypothesis-generating, and larger, adequately powered investigations are needed to ascertain whether miR-210 has clinically evocative predictive value as an exercise biomarker.