DOI: 10.4103/bbrj.bbrj_83_26 ISSN: 2588-9834

Predicting Inflammatory Adaptation to Exercise via Circulating miR-146a and Machine Learning

Fatma Hassan Abd Elbasset Mourgan
Show PDF Cite

Abstract

Background:

Circulating microRNAs are an emerging noninvasive tool for monitoring physiological adaptations to exercise. miR-146a is of special interest as it plays an important role in inflammatory regulation and immune-response feedback at the cellular level.

Methods:

Ten exercise-trained and 10 nonexercise controls were included for this study. The fold change of miR-146a expression was determined. High miR-146a expression (the sample median split [median = 1.029]) was considered the primary outcome. To prevent outcome leakage, the following factors were included as model predictors: age, body mass index, and exercise-training status.

Results:

miR-146a fold-change values ranged from 0.656 to 1.854, with an overall mean of 1.080 ± 0.288. The exercise-trained group showed a numerically higher mean expression than nonexercise controls (1.127 ± 0.211 vs. 1.033 ± 0.354), although this difference was not statistically significant (Welch P = 0.483; Mann–Whitney P = 0.224). High miR-146a expression was observed in 7/10 exercise-trained participants and 3/10 controls (Fisher exact P = 0.179), suggesting a directional but underpowered group trend. In leave-one-out cross-validation, k-nearest neighbors achieved 75% accuracy (area under the receiver operating characteristic curve [AUC] = 0.710), quadratic discriminant analysis with quadratic feature expansion achieved 75% accuracy with the highest AUC (0.790), and polynomial linear discriminant analysis achieved the highest balanced classification performance with 80% accuracy (AUC = 0.770, Brier = 0.200, sensitivity = 0.80, and specificity = 0.80).

Conclusions:

In this small exploratory cohort, optimized classification improved the miR-146a signal to a defensible moderate level. The 80% best-model accuracy should be retained r because it was obtained using nonleakage predictors and clean internal validation.

More from our Archive