Abstract
PURPOSE: This study aims to employ machine learning for automated age stratification in patients with polycystic ovary morphology (PCOM) and to clarify the age-specific contributions of key clinical characteristics, thereby improving diagnostic accuracy. PATIENTS AND METHODS: A total of 192 ovaries with corresponding clinical and ultrasound data were analyzed. Automated age stratification was performed using the K-means unsupervised clustering model along with the elbow method. XGBoost, Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Network (ANN) algorithms were applied to compare full and selected feature sets. Model performance was evaluated using five-fold cross-validation with metrics including accuracy, sensitivity, F1-score, and area under the curve (AUC), validating the stratification robustness and feature selection strategy. RESULTS: Automated age stratification categorized PCOM patients into three distinct age groups. A stable positive correlation was observed between ovarian volume and follicle count across all strata; however, key diagnostic drivers varied significantly by age. Beyond ovarian volume, follicle count, and BMI, the 18-22 years group was primarily influenced by menstrual cycle phase. The 23-29 years group was characterized by the number of pregnancies and the ovarian stromal artery blood flow RI. In the 30-40 years group, the ovarian stromal artery blood flow S/D ratio and the number of live births showed increasing importance. CONCLUSION: This study highlights the differential role of age-sensitive indicators in ultrasound-based diagnosis of PCOM and provides a framework for more personalized diagnostic assessment.