Machine learning prediction of metabolic dysregulation in women with polycystic ovary syndrome: development and validation of hemato-inflammatory predictive models

Insulin resistance and dyslipidemia are prevalent in 50–70% of women diagnosed with polycystic ovary syndrome (PCOS), which is distinguished by metabolic heterogeneity. Non-invasive biomarkers that are accessible under restricted circumstances are necessary for the early detection of metabolic dysfunction. Hematological inflammatory indices can function as markers that may suggest metabolic dysregulation, as they are derived from routine complete blood count parameters. Our objective is to create and verify machine learning models that anticipate metabolic dysregulation in PCOS by utilizing hematological inflammatory indices. This cross-sectional study enrolled 200 women stratified into four groups (n = 50 each): PCOS-Obese, PCOS-Lean, Control-Obese, and Control-Lean. Seven input features were selected based on clinical relevance and multicollinearity management: four hematological inflammatory indices (Neutrophil-to-Lymphocyte Ratio NLR, Platelet-to-Lymphocyte Ratio PLR, Systemic Immune-Inflammation Index SII, Monocyte-to-Lymphocyte Ratio MLR), hemoglobin, age, and Body Mass Index (BMI). Machine learning algorithms (Random Forest, Gradient Boosting, Support Vector Regression, Linear Regression) were trained using 70% of data (n = 140) and validated on 30% independent test set (n = 60) with 5-fold nested cross-validation. All preprocessing standardization was performed only on training data to prevent data leakage. Target metabolic outcomes included C-reactive protein (CRP), defined as a systemic inflammatory biomarker, triglycerides, glucose, HDL-cholesterol, and Homeostasis Model Assessment for Insulin Resistance (HOMA-IR). PCOS-Obese women demonstrated significantly elevated hematological inflammatory indices compared to controls (NLR: 2.58 ± 1.76 vs. 1.61 ± 0.73, p < 0.001; PLR: 150.99 ± 68.14 vs. 91.83 ± 25.65, p < 0.001; SII: 739.46 ± 519.98 vs. 429.57 ± 231.97, p < 0.001; MLR: 0.25 ± 0.13 vs. 0.15 ± 0.06, p < 0.001). Age-adjusted ANCOVA confirmed group differences were independent of age (all p < 0.001). Random Forest achieved modest test set R² values for triglycerides (R²=0.571) and CRP (R²=0.248). However, models for HDL-cholesterol demonstrated negative R² values (range: 0.110 to 5.276 across algorithms), indicating hemato-inflammatory indices alone do not predict this metabolic parameter. This discrepancy between high cross-validation R² (e.g., triglycerides CV: 0.641 ± 0.068) and lower test set R² (0.015–0.571) reflects model instability at the individual prediction level, despite group-level discrimination. While hematological inflammatory indices effectively distinguish PCOS-obese phenotypes at the group level (indicating distinct inflammatory signatures), their predictive utility for individual metabolic outcome forecasting remains limited. These preliminary exploratory findings suggest hemato-inflammatory profiling may serve as a screening signal for metabolic risk stratification in resource-constrained settings; however, external validation in larger, prospective, multicenter cohorts (n ≥ 500–1000) is required. Enhanced models incorporating additional metabolic, hormonal, and genetic parameters are necessary for clinical implementation. This work contributes to the emerging application of machine learning to hemato-inflammatory biomarkers in PCOS metabolic phenotyping, though future research must address current predictive limitations. Not applicable.