Artificial intelligence (AI) can transform osteoporosis (OP) screening, but its application in high-risk, complex populations like postmenopausal women with chronic kidney disease remains limited. We sought to develop and interpret an explainable AI model specifically for identifying low bone mass (LBM) or OP in this vulnerable group. Using data from the National Health and Nutrition Examination Survey (2005–2018), a nationally representative cross-sectional study, we developed and compared 8 AI models. For the purpose of algorithm development and comparison, the complex survey weights of National Health and Nutrition Examination Survey were not applied. Model performance was assessed by AUC-ROC, recall, precision, F1 score, and Brier score. Interpretability was achieved using SHapley Additive exPlanations, local interpretable model-agnostic explanations (LIME) and generalized additive models to identify key features and their nonlinear interactions. The multilayer perceptron model emerged as the most effective (AUC: 0.72; Precision: 0.84). Interpretation revealed that weight, age, estimated glomerular filtration rate (eGFR), age when heaviest weight, total cholesterol (TC), and age at last menstrual period were top predictors. Critically, we discovered that the relationship between weight and LBM/OP is nonlinear, and that complex interactions exist between weight, age, and age when heaviest weight. This study presents a well-validated and interpretable AI model for LBM/OP screening in postmenopausal women with chronic kidney disease. By moving beyond a “black-box” approach, we provide a tool with potential clinical utility and novel pathophysiological insights, though its implementation requires further external validation in prospective, independent cohorts. This work underscores the potential of AI to enhance understanding of OP risk and highlights the necessity of personalized risk evaluation in this population.
Zhu et al. (Fri,) studied this question.