Expert knowledge-based weighted K-means clustering for polycystic ovary syndrome phenotype identification
- 1School of Computer Science Engineering and Information Systems, Vellore Institute of Technology , ,
Abstract
Background: Polycystic ovary syndrome (PCOS) is a heterogeneous endocrine disease condition that affects 8%–13% of the female population worldwide. Its clinical heterogeneity makes the binary approach to phenotyping less effective for the biological significance of such phenotypic groups and the personalized management of PCOS. To overcome this issue, we present the expert knowledge-based weighted K -means (EKWKM) approach.
Objective: The proposed EKWKM is a new, weakly supervised phenotyping approach that uses structured clinical knowledge along with mutual information (MI)-based feature weighting, whereby the feature weighting process is weakly supervised, but clustering is unsupervised.
Methods: The EKWKM was benchmarked using a clinical dataset ( n = 541, 41 features) and a large-scale population study ( n = 120,000). A grid search was performed for K ∈ {2, 3, 4} and α = 0.1–0.9. The silhouette score (95% bootstrap confidence interval), the adjusted Rand index (ARI), and the Davies–Bouldin (DB) index were used to quantify cluster quality. Biomarker importance was measured using the Kruskal–Wallis test with Benjamini–Hochberg correction and ϵ ² effect size.
Results: The most effective clinical model ( K = 2, α = 0.9) achieved a silhouette score of 0.3606. The model also achieved ARI = 0.9780 ± 0.0362 and DB = 1.1480 and identified 13 significant biomarkers. The population survey acted as a negative control dataset [random forest area under the curve (AUC) = 0.4974 ± 0.0096] and demonstrated no phenotypic signal. Cross-dataset transfer decreased Cramér’s V from 0.497 to 0.006, which confirmed that EKWKM did not add false patterns. The main part of the clustering was due to data-driven MI, as confirmed by the ablation analysis. Expert knowledge improved clustering robustness.
Conclusion: EKWKM formed clinically meaningful and Rotterdam-based PCOS phenotype clusters. In contrast, the population survey clusters formed symptom-reporting patterns and are not meaningful PCOS phenotypes.
Related
- Assessment of Knowledge and Attitude Regarding Polycystic Ovarian Syndrome Among Nursing Students
- Interpretable Phenotype Discovery Using Temporal Risk-Aware Clustering
- Identification and Validation of Genes Involved in Immune and Ferroptosis in Polycystic Ovarian Syndrome Using Bioinformatics and Experimental Approach
- Knowledge gaps regarding polycystic ovary syndrome/polyendocrine metabolic ovarian syndrome among medical students: a cross-sectional study
- Identification of Mitochondrial Dynamics-Related Genes as Diagnostic Biomarkers for Osteoarthritis Using Weighted Gene Co-expression Network Analysis and Machine Learning
BibTeX
@article{priyanka2026expert,
title = {{Expert knowledge-based weighted K-means clustering for polycystic ovary syndrome phenotype identification}},
author = {Priyanka D. and Sumangali K.},
journal = {Frontiers in endocrinology},
year = {2026},
month = sep,
doi = {10.3389/fendo.2026.1871986},
url = {https://europepmc.org/article/PMC/PMC13662547},
note = {CC BY},
}
RIS
TY - JOUR
TI - Expert knowledge-based weighted K-means clustering for polycystic ovary syndrome phenotype identification
AU - Priyanka D.
AU - Sumangali K.
PY - 2026
DA - 2026/09/30
JO - Frontiers in endocrinology
DO - 10.3389/fendo.2026.1871986
UR - https://europepmc.org/article/PMC/PMC13662547
AB - Background: Polycystic ovary syndrome (PCOS) is a heterogeneous endocrine disease condition that affects 8%–13% of the female population worldwide. Its clinical heterogeneity makes the binary approach to phenotyping less effective for the biological significance of such phenotypic groups and the personalized management of PCOS. To overcome this issue, we present the expert knowledge-based weighted K -means (EKWKM) approach. Objective: The proposed EKWKM is a new, weakly supervised phenotyping approach that uses structured clinical knowledge along with mutual information (MI)-based feature weighting, whereby the feature weighting process is weakly supervised, but clustering is unsupervised. Methods: The EKWKM was benchmarked using a clinical dataset ( n = 541, 41 features) and a large-scale population study ( n = 120,000). A grid search was performed for K ∈ {2, 3, 4} and α = 0.1–0.9. The silhouette score (95% bootstrap confidence interval), the adjusted Rand index (ARI), and the Davies–Bouldin (DB) index were used to quantify cluster quality. Biomarker importance was measured using the Kruskal–Wallis test with Benjamini–Hochberg correction and ϵ ² effect size. Results: The most effective clinical model ( K = 2, α = 0.9) achieved a silhouette score of 0.3606. The model also achieved ARI = 0.9780 ± 0.0362 and DB = 1.1480 and identified 13 significant biomarkers. The population survey acted as a negative control dataset [random forest area under the curve (AUC) = 0.4974 ± 0.0096] and demonstrated no phenotypic signal. Cross-dataset transfer decreased Cramér’s V from 0.497 to 0.006, which confirmed that EKWKM did not add false patterns. The main part of the clustering was due to data-driven MI, as confirmed by the ablation analysis. Expert knowledge improved clustering robustness. Conclusion: EKWKM formed clinically meaningful and Rotterdam-based PCOS phenotype clusters. In contrast, the population survey clusters formed symptom-reporting patterns and are not meaningful PCOS phenotypes.
SN - 1664-2392
ER -