RT info:eu-repo/semantics/doctoralThesis
T1 Modelo predictivo de fractura osteoporótica con Inteligencia Artificial
A1 Mateo Sotos, Jorge
A2 Universidad de Valladolid. Escuela de Doctorado
K1 Biomedicina
K1 Osteoporotic fracture
K1 Fractura osteoporótica
K1 Artificial Intelligence
K1 Inteligencia artificial
K1 32 Ciencias Médicas
AB Osteoporosis is one of the main causes of fragility fractures in postmenopausal women, with major repercussions on quality of life, mortality, and healthcare costs. Dual-energy X-ray absorptiometry (DXA) remains the gold standard for diagnosis; however, its predictive power is limited, as a significant proportion of fractures occur in women whose bone mineral density (BMD) values fall within the normal or osteopenic range. Therefore, it is crucial to develop more accurate tools capable of identifying high-risk patients before an adverse event occurs.The objective of this thesis was to develop and validate fracture prediction models based on machine learning algorithms, specifically Random Forest (RF) and eXtreme Gradient Boosting (XGB), integrating clinical, demographic, and densitometric variables, as well as parameters derived from 3D-DXA and the Trabecular Bone Score (TBS). A cross-sectional study was conducted including 576 postmenopausal women from two independent cohorts: the HURH cohort (n = 276, patients diagnosed with osteoporosis) and the Camargo cohort (n = 300, general population). The former was used for model training (internal validation using 5-fold cross-validation and a 30% hold-out test), while the latter was reserved for external validation.The discriminative performance of the models was compared with that of other algorithms (K-nearest neighbors, Support Vector Machines, Decision Trees, and Gaussian Naïve Bayes) and with the clinical FRAX index. In the internal test, the RF model achieved an accuracy of 89.24% and an AUC of 0.89; in the external validation, it maintained high performance (accuracy of 87.62% and AUC of 0.87), consistently outperforming all other classifiers and more than doubling the prognostic capacity of FRAX. The XGB model showed slightly superior performance, with improved probability calibration and greater stability across cohorts, confirming its suitability for heterogeneous clinical settings.The variable importance analysis identified previous fracture, parathyroid hormone (PTH) levels, and lumbar T-score as the main predictors, along with other densitometric parameters. Moreover, a reduced set of easily obtainable variables maintained predictive accuracy comparable to the full model, reinforcing its practical applicability in routine clinical contexts.In conclusion, the combination of artificial intelligence and clinical densitometric data optimizes fracture risk stratification in postmenopausal women. The models developed, particularly those based on RF and XGB, provide robust, accurate, and interpretable tools that can be readily implemented in clinical practice, enabling personalized preventive interventions and helping to reduce the morbidity and mortality associated with osteoporotic fractures.
YR 2026
FD 2026
LK https://uvadoc.uva.es/handle/10324/84059
UL https://uvadoc.uva.es/handle/10324/84059
LA spa
NO Escuela de Doctorado
DS UVaDOC
RD 19-abr-2026