Research on Assisting X-ray Diagnosis of Osteoporotic Vertebral Compression Fractures Using Interpretable Machine Learning Models and Radiomics Features

Objective: To improve early diagnosis rates, this study applies a combination of radiomics and machine learning algorithms to aid in the X-ray diagnosis of osteoporotic vertebral compression fractures (OVCF). Methods: Data were collected from 852 patients from January 2016 to December 2023, including lateral X-rays of the L1 vertebra and demographic information. The cohort included 589 patients with lumbar back pain but normal MRI results, and 263 patients diagnosed with various degrees of OVCF by MRI. Patients were randomly divided into training (70%) and validation (30%) groups. X-ray images were annotated to extract radiomics features, which were then selected to finalize the radiomics score, along with meaningful clinical factors. Five machine learning algorithms were utilized to model and compare the diagnostic efficacy of clinical prediction models, radiomics models, and combined models, identifying the optimal model group and machine learning algorithm. The SHAP method was employed for further explanatory analysis. Results: Variables showing significant differences between groups included gender, smoking history, trauma history, history of lumbar surgery, residential area, history of glucocorticoid treatment, age, and VAS score. Through t-tests, intraclass correlation coefficients (ICCs), and LASSO regression analysis (Least Absolute Shrinkage and Selection Operator), eight radiomics features were identified to establish a Radscore. Multifactorial logistic regression analysis identified gender, smoking history, trauma history, lumbar surgery history, residential area, and Radscore as independent risk factors for OVCF. The combined model outperformed the other two. Due to overfitting in the Random Forest algorithm, KNN was determined to be the best machine learning algorithm. SHAP bar graphs displayed the influence factors in descending order of impact: residential area, Radscore, trauma history, gender, smoking, and lumbar surgery history. SHAP swarm plots revealed a broad distribution of Radscore, underscoring its significant predictive influence. Conclusion: The diagnostic model developed through radiomics and machine learning algorithms reached an ideal level of effectiveness, with KNN in the combined model group demonstrating the highest diagnostic efficacy for assisting in the early X-ray diagnosis of OVCF.