Early Identification of Low Bone Density in Primary Ovarian Insufficiency: A Machine Learning Approach

77INTRODUCTIONPrimary Ovarian Insufficiency (POI) is a condition characterizedby the premature loss of ovarian function before theage of 40, leading to estrogen deficiency and several systemiccomplications, including reduced bone mineral density(BMD). Although dual-energy X-ray absorptiometry is consideredthe gold standard for assessing BMD, its use may belimited by costs and waiting times. In this context, machinelearning models may help identify patients at higher risk oflow BMD at an earlier stage.OBJECTIVESThe aim of this study was to develop a machine learningmodel for the early prediction of reduced bone mineral densityin women with POI, using medical history, clinical, laboratory,genetic, and instrumental data collected in clinicalpractice. A secondary objective was to identify the variablesmost strongly associated with the outcome through model interpretabilitytechniques.METHODSThe analysis was conducted on a clinical dataset of 250patients with POI from the Pediatric and Adolescent GynecologyUnit, Careggi University Hospital, including demographic,clinical, hormonal, genetic, and autoimmune characteristics.After a priori variable selection, based on literatureand redundancy criteria, and multiple imputation of missingdata, the dataset was split into a development (80%) and atest set (20%). Subsequently, the development set was splitinto a training (80%) and a validation (20%) sets; several supervisedclassification algorithms including linear penalizedmodels, Support Vector Machines, Random Forest, GradientBoosting Machine, AdaBoost, XGBoost, and Gaussian Process,were trained on the training set. For each algorithm, hyperparametertuning was performed using 5-fold cross-validationwithin the training set, and the optimal hyperparameterconfiguration was selected based on the area under thereceiver operating characteristic curve (AUC). The resultingmodels, fitted with their optimal hyperparameters, were thenevaluated on the validation set for model selection. Finally,the best-performing model was assessed on the test set to obtainan unbiased estimate of its predictive performance. Interpretabilitywas assessed using Shapley values.RESULTSThe analyzed population is relatively young, with a meanage of 27 years, and it’s characterized by marked hormonalvariability, reflecting the underlying heterogeneity of the clinicalcondition. In this context, ensemble and boosting-basedmachine learning models demonstrated the best overall discriminativeperformance compared to other approaches. Inparticular, Random Forest achieved the highest predictive accuracyon the independent test set, with an AUC of 0.803, followedclosely by Gradient Boosting Machine (AUC = 0.793),XGBoost (AUC = 0.769), and AdaBoost (AUC = 0.746),confirming the ability of tree-based ensemble methods in thissetting.Calibration analysis further indicated a good agreementbetween predicted probabilities and observed outcomes. Finally,SHAP analysis highlighted the most relevant predictorsof the outcome, including key clinical and laboratory variablessuch as FSH levels, body mass index, and age at menarche,in line with established clinical evidence. CONCLUSIONSMachine learning models based on clinical and laboratoryvariables represent a promising tool for the early identificationof POI patients at risk of reduced bone mineral density,improving risk stratification and clinical decision-makingsupport.