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

Abstract

77
INTRODUCTION
Primary Ovarian Insufficiency (POI) is a condition characterized
by the premature loss of ovarian function before the
age of 40, leading to estrogen deficiency and several systemic
complications, including reduced bone mineral density
(BMD). Although dual-energy X-ray absorptiometry is considered
the gold standard for assessing BMD, its use may be
limited by costs and waiting times. In this context, machine
learning models may help identify patients at higher risk of
low BMD at an earlier stage.
OBJECTIVES
The aim of this study was to develop a machine learning
model for the early prediction of reduced bone mineral density
in women with POI, using medical history, clinical, laboratory,
genetic, and instrumental data collected in clinical
practice. A secondary objective was to identify the variables
most strongly associated with the outcome through model interpretability
techniques.
METHODS
The analysis was conducted on a clinical dataset of 250
patients with POI from the Pediatric and Adolescent Gynecology
Unit, Careggi University Hospital, including demographic,
clinical, hormonal, genetic, and autoimmune characteristics.
After a priori variable selection, based on literature
and redundancy criteria, and multiple imputation of missing
data, the dataset was split into a development (80%) and a
test set (20%). Subsequently, the development set was split
into a training (80%) and a validation (20%) sets; several supervised
classification algorithms including linear penalized
models, Support Vector Machines, Random Forest, Gradient
Boosting Machine, AdaBoost, XGBoost, and Gaussian Process,
were trained on the training set. For each algorithm, hyperparameter
tuning was performed using 5-fold cross-validation
within the training set, and the optimal hyperparameter
configuration was selected based on the area under the
receiver operating characteristic curve (AUC). The resulting
models, fitted with their optimal hyperparameters, were then
evaluated on the validation set for model selection. Finally,
the best-performing model was assessed on the test set to obtain
an unbiased estimate of its predictive performance. Interpretability
was assessed using Shapley values.
RESULTS
The analyzed population is relatively young, with a mean
age of 27 years, and it’s characterized by marked hormonal
variability, reflecting the underlying heterogeneity of the clinical
condition. In this context, ensemble and boosting-based
machine learning models demonstrated the best overall discriminative
performance compared to other approaches. In
particular, Random Forest achieved the highest predictive accuracy
on the independent test set, with an AUC of 0.803, followed
closely 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 this
setting.
Calibration analysis further indicated a good agreement
between predicted probabilities and observed outcomes. Finally,
SHAP analysis highlighted the most relevant predictors
of the outcome, including key clinical and laboratory variables
such as FSH levels, body mass index, and age at menarche,
in line with established clinical evidence.

CONCLUSIONS
Machine learning models based on clinical and laboratory
variables represent a promising tool for the early identification
of POI patients at risk of reduced bone mineral density,
improving risk stratification and clinical decision-making
support.