Early PCOS Diagnosis with Minimal Features via Explainable Machine Learning

Polycystic Ovary Syndrome (PCOS) is a common and complex endocrine disorder affecting women. This study aims to develop explainable machine learning models for the diagnosis of PCOS using minimal clinical data. An anonymized dataset comprising 1,000 records obtained from the Kaggle platform was used to compare eight different algorithms. Ensemble and Decision Tree models achieved the highest accuracy (99.9%), with the Decision Tree selected for its superior explainability. The model’s decision process was based on menstrual irregularity, $\mathrm{BMI}\gt25 \mathrm{~kg} / \mathrm{m}^{2}$, testosterone $\geq 41$ $\mathrm{ng} / \mathrm{dL}$, and antral follicle count $\geq 10.5$. The findings align with known risk factors reported in the literature. Results suggest that highly accurate diagnostic support systems can be developed using a small number of accessible variables, supporting the future development of home-based screening methods.