Machine learning and penalized logistic regression-based models for the identification of potential characteristic genes in primary hyperparathyroidism.

PURPOSE: Primary hyperparathyroidism (PHPT) is an endocrine disorder in which one or more parathyroid glands oversecrete parathyroid hormone, leading to hypercalcemia and potential complications such as kidney stones, bone fractures, and other systemic issues. Machine learning (ML) is extensively applied to diagnose medical diseases due to its high accuracy and ease of implementation.

METHODS

The microarray expression dataset (GSE83421) was selected from the Gene Expression Omnibus Database (GEO) for analysis. Differential expression analysis was performed with false discovery rate (FDR) correction. Weighted gene co-expression network analysis (WGCNA) identified correlated gene modules. Consensus feature selection using six algorithms (LASSO, SVM-RFE, random forest, Boruta, XGBoost, and decision tree) was applied to identify robust biomarkers and mitigate overfitting. A penalized logistic regression model was developed and evaluated using leave-one-out cross-validation (LOOCV).

RESULTS

Ten candidate genes were identified by intersecting 24 DEGs with the WGCNA red module. Consensus ML selected five hub genes (LOC51092, KIAA1479, PRO2214, TBX4, and ZNF361). T. The LOOCV-validated penalized logistic regression model effectively distinguished PHPT from normal controls.

CONCLUSION

Our analysis identified a 5-gene signature (LOC51092, KIAA1479, PRO2214, TBX4, and ZNF361) with diagnostic potential for PHPT. The model was internally validated using LOOCV. However, without nested feature selection or external validation, the reported performance likely represents an optimistic estimate. Given the limited sample size, these findings represent preliminary hypotheses for PHPT diagnosis, requiring validation in independent cohorts.