In this study, we aimed to develop 2-year fracture prediction models using machine learning approaches based on routine blood test markers in patients newly diagnosed with osteoporosis. We retrospectively analyzed data from 497 elderly patients (121 men and 376 women) newly diagnosed with osteoporosis without prior fractures at Jinjiang Municipal Hospital between 2011 and 2023. The dataset was randomly divided into training and validation cohorts. Various readily available medical characteristics at the initial diagnosis of osteoporosis were used to develop the prediction models employing the linear, nonlinear, and SuperLearner approaches. Model performance was assessed using the area under the curve (AUC) and the SHapley Additive exPlanation was used to interpret the best-performing models. In total, 229 (46.1%) of the 497 participants experienced fractures (at any site) within 2 years of osteoporosis diagnosis. The SuperLearner model demonstrated the highest AUC (0.677, 95% confidence interval: 0.591, 0.762) for predicting the outcome of interest in the internal validation set. Age at diagnosis, serum sodium levels, and other factors emerged as the top-ranked features. Older patients with hyponatremia, low aspartate transferase levels, and hypocalcemia have an increased imminent risk of fracture. We successfully developed and validated a fracture risk prediction model for patients newly diagnosed with osteoporosis. The combination of routine blood test markers and advanced machine learning algorithms has demonstrated a predictive value for forecasting fractures in individuals with osteoporosis. This study developed a model to predict fractures in patients with newly diagnosed osteoporosis. Leveraging routine blood test data and machine learning, the SuperLearner model demonstrated high accuracy in identifying key predictors. This innovation offers promise for improving fracture prevention and management and enhancing overall patient care.
Chen et al. (Tue,) studied this question.