A Metabolic-Inflammatory Phenotype of Pelvic Floor Dysfunction: A Machine Learning-Based Discovery in a Nationally Representative U.S. Cohort

Pelvic floor dysfunction (PFD) is a highly prevalent and heterogeneous condition among women. The traditional view of PFD as a single clinical entity limits the understanding of its complex pathophysiology and hinders the development of personalized therapies. We aimed to deconstruct the heterogeneity of PFD by using unsupervised machine learning to identify distinct clinical subtypes in a nationally representative sample of U.S. women. This cross-sectional study included 7,291 female participants aged ≥20 from the National Health and Nutrition Examination Survey (NHANES) 2005-2012. We employed a two-stage analytical approach. First, K-means clustering was performed on PFD-positive women using 12 physiological features to identify subtypes. Second, we developed and validated seven multiclass machine learning models to predict subtype membership (Healthy Control, Phenotype 1, Phenotype 2). The best model was interpreted using SHAP (SHapley Additive exPlanations). Two distinct clinical subtypes were identified among PFD-positive women. Phenotype 1 (Metabolic-Inflammatory) was characterized by severe metabolic disturbances, including central obesity (mean BMI 36.1 kg/m²), and high prevalence of diabetes (21%) and hypertension (50%). In contrast, Phenotype 2 (Metabolically-Healthy) exhibited PFD symptoms but had a metabolic profile nearly identical to healthy controls (mean BMI 25.4 kg/m²; diabetes prevalence 3.7%). Our prediction models accurately distinguished the three groups, with a Neural Network achieving the highest macro-AUC of 0.848. SHAP analysis revealed that waist circumference was overwhelmingly the most important predictor for differentiating the subtypes. PFD is not a monolithic disorder but comprises at least two distinct clinical subtypes. One subtype is intrinsically linked to systemic metabolic and inflammatory disease, while the other appears to be driven by more traditional risk factors. This novel, data-driven classification provides a new framework for understanding PFD and may pave the way for precision diagnostics and targeted therapeutic strategies.