Clustering of non-invasive biomarkers to distinguish phenotype-related response differences in women with PCOS after the 16-week Nordic interval training intervention: A prospective cohort study.

BACKGROUND

Polycystic ovary syndrome (PCOS) is associated with metabolic and reproductive abnormalities. Response variability due to exercise may reflect underlying biological heterogeneity in PCOS phenotypes. We explored changes in non-invasive PCOS biomarkers following a 16-week Interval Nordic Walking (INW) to identify phenotype-related response differences.

METHODS

A total of 7500 women's data were obtained from the Yazd Reproductive Sciences and Amir al Mómenin Hospital (2019-2024). PCOS was diagnosed according to the Rotterdam criteria. XGBoost feature selection was implemented to choose 17 key predictors. Multiple ML classifiers were evaluated, and model interpretability was achieved using SHAP values. 80 women with different PCOS phenotypes completed a 16-week INW intervention. PCA was applied to reduce the features to 75% variance. Then clustering algorithms (K-means, GMM, Hierarchical, Spectral, DBSCAN) were evaluated using Silhouette scores, Calinski-Harabasz indices, and bootstrap resampling (Jaccard index).

RESULTS

XGBoost and SHAP analysis confirmed AMH, FSH, and Anti-TPO as the most influential predictors of PCOS, while metabolic indices like HbA1c, BMI, and lipid profiles exhibited moderate but clinically relevant effects. The spectral model provided better separation among evaluated algorithms; however, overall cluster separation remained modest. It suggested two exploratory data-driven response patterns, termed high-response (more frequently observed phenotypes one and two), which demonstrated greater observed reductions in testosterone and DHEA-s along with improvements in HDL and total cholesterol following the intervention. Low-response (more frequently observed phenotypes three and four) showed relatively minimal response to INW. Based on mean Jaccard index = 0.45, cluster stability was moderate.

CONCLUSIONS

Our study suggests the heterogeneous phenotypic response pattern of PCOS following a 16-week INW intervention and supports further investigation of nonpharmacological strategies for PCOS.