| Objective Osteoporosis, a skeletal disease with a high incidence, insidious onset, and long course, which can cause a series of severe symptoms including fractures, is one of the leading causes of disability and mortality among the elderly in our country. There are many physiological test indexes related to osteoporosis. However, there is no mature and unified method to screen and to diagnosis at present. Methods Artificial intelligence technology was used to select features from indicators of clinical osteoporosis patients using a variety of feature-related algorithms. Based on this, a multi-level ensemble learning framework was proposed, including SAB-SVMKNN algorithm, which combined internal homogeneous learning with external heterogeneous learning. Boosting and Bagging algorithms were integrated in learning with Stacking to build a diagnostic prediction model with better performance and adaptability. Results Eight of the most important features for osteoporosis were selected from 31 clinical indicators in the original data by feature selection algorithm, which improved the accuracy of each model by an average of 9.2%, and the corresponding model accuracy in this study increased by 18.6%, reaching a final accuracy of 94.8%. Conclusion Feature selection is of great significance for clinical diagnosis and the study of osteoporotic diseases. The constructed prediction model can improve the diagnosis accuracy for physicians. |