A radiomics-based interpretable model to predict the pathological grade of peritoneal pseudomyxoma

Objective To construct a machine learning model that integrates delayed-phase enhanced CT radiomics with clinical features and employs Shapley Additive Explanations (SHAP) for interpretable analysis, aiming to provide accurate prediction and analysis for the pathological grading of appendiceal pseudomyxoma peritonei (PMP). Materials and methods The clinical and imaging data of 158 patients with appendiceal PMP confirmed by pathology in Aerospace Center Hospital from January 2015 to April 2024 were retrospectively analyzed, including 85 cases of low-grade appendiceal PMP and 73 cases of high-grade appendiceal PMP.The Clinical data included age, sex, white blood cell(WBC),carcinoembryonic antigen (CEA), cancer antigen 199 (CA199), CA125, D-dimer, CA-724, CA-242 and preoperative Computed tomography imaging for peritoneal cancer index(CT-PCI).Clinical indicators with statistical significance were screened by univariate and multivariate analysis.The radiomics features of the region of interest (ROI) of appendiceal lesions were extracted from the preoperative enhanced thin-section CT images, and the combined screening methods were used to screen out the optimal radiomics features.The five-fold cross validation method was employed to partition the dataset into training and testing sets.Logistic Regression classifier was used to establish three prediction models based on clinical indicators, radiomics and clinical-imaging combination.The interpretable SHAP value was used to analyze the model to clarify the weight of each group in the model.The receiver operating characteristic(ROC) curve was used to distinguish the performance of each model, and the prediction results of each model were quantified by the area under the curve (AUC).DeLong test was used to compare the AUC differences between three models.The net benefits of each model under different threshold probabilities were quantified by Decision Curve Analysis (DCA) to evaluate the clinical applicability of each model. Results The AUC values for the radiomics model, clinical model, and their combined model in the training set were 0.84 (95%CI 0.78-0.90), 0.82 (95%CI 0.75-0.88),and 0.91 (95%CI 0.86-0.95), respectively.The AUC values in the testing set were 0.80 (95%CI 0.73-0.86), 0.78 (95%CI 0.71-0.86), and 0.88 (95%CI 0.82-0.93), respectively.The Delong test revealed that the combined model's predictive performance was significantly superior to that of the radiomics model and the clinical model(P< 0.05). The DCA of the testing set showed that the net clinical benefit of the radiomics model and the combined model was higher than that of the clinical model.The net clinical benefit of the combined model was higher than that of the radiomics model in most threshold probability intervals.SHAP analysis showed that radiomics features were more important than clinical indicators, and the top five features were as follows:wavelet-LHH_glcm_InverseVariance,original_shape_Elongation,square_glszm_SizeZoneNonUniformityNormalized,wavelet-LLL_firstorder_Median and CA-199. Conclusion The combined prediction model combining radiomics and clinical features, and using SHAP interpretable analysis method, has a good predictive value for the pathological grade of appendiceal PMP, which provides a non-invasive and efficient method for clinical decision-making, helps to provide individualized treatment plans for patients, and improves the prognosis of patients.