Background <p>The intravertebral shell phenomenon (ISP) is mainly characterized by the formation of a bone defect area resembling an eggshell in the injured vertebra. Traditional prediction methods often fail to capture the complex interactions among the factors of the inner shell phenomenon within the vertebrae. This study aims to establish and validate an interpretable prediction model and web application for ISP based on machine learning (ML), in order to provide suggestions for clinical prevention and treatment.</p> Methods <p>Data on 20 characteristics from 594 patients with type A thoracolumbar fractures&#xa0;without posterior ligamentous complex injury, who underwent posterior pedicle screw fixation in our hospital from January 1, 2015 to December 31, 2023, were collected. The missing data are interpolated using the K-nearest neighbor method or the mode filling method. The synthetic minority oversampling technique (SMOTE) is adopted to address the issue of class imbalance. Optimal feature selection is carried out using the recursive feature elimination (RFE) method. All patients were randomly divided into a training set and a test set in a 7:3 ratio. Subsequently, five ML models, including support vector machine (SVM), logistic regression (LR), random forest (RF), XGBoost, and LightGBM, were used to develop and validate the models based on the selected features. The generalization ability of the model is evaluated using fivefold cross-validation. The performance of the machine learning model was comprehensively evaluated through relevant indicators, such as the area under the receiver operating characteristic curve (AUC) and accuracy. Finally, the best model was selected and the importance of variables was described using SHapley Additive exPlanations (SHAP), as well as the development of a web application.</p> Results <p>Among the 5 ML models, the RF model has the best discriminative ability. In the test set, the AUC value is 0.912. In five-fold cross-validation, the mean AUC value is 0.889. The SHAP values reflecting the importance of the characteristics of ISP patients are as follows: the age score is the highest, followed by the AO spine classification, bone mineral density (BMD), degree of vertebral compression (DVC), body mass index (BMI), gender, degree of vertebral reduction (DVR), and Cobb angle. Ultimately, these eight features were used to generate a web application under the weights of the RF model, in order to facilitate its application in the clinical environment.</p> Conclusions <p>Based on interpretable ML, this predictive model and web application have excellent performance and clinical practicability, which could provide a certain basis for predicting the formation of ISP in patients after thoracolumbar fracture surgery.</p>