Research on the construction and application of a multi-feature auxiliary diagnostic model for uremic pneumonia based on radiomics and deep learning: a retrospective bi-centre study

Abstract Objective This study aimed to develop and validate a diagnostic model that integrates radiomics and deep learning (DL) for the early differentiation of uremic pneumonia (UP) from pulmonary infection (PI) in non-dialysis patients. Methods This study retrospectively analyzed the clinical data and non-contrast CT images of 334 non-dialysis patients with stage 5 chronic kidney disease (CKD). In this study, univariate and multivariate logistic regression analysis methods were employed to screen for clinical risk factors. Radiomics and DL features of lesions were extracted based on non-contrast chest CT images. Clinical model, radiomics model, and DL model were constructed separately, and a combined deep learning-radiomics-clinical (DLRC) model was established using feature fusion methods. The performance of each model was evaluated using the receiver operating characteristic (ROC) curve, the area under the curve (AUC), the calibration curve, and decision curve analysis (DCA). In addition, the generalization ability of each model was validated using an external validation cohort (n = 93). Results The DLRC model demonstrated the highest diagnostic performance, with AUC values of 0.944 (95% CI: 0.911–0.978) in the training set and 0.930 (95% CI: 0.873–0.985) in the test set. The AUC values of each single-feature model in the training set and the test set were as follows: clinical model (0.828 and 0.894), radiomics model (0.862 and 0.813), and DL model (0.881 and 0.855). In the external validation set, the DLRC model continued to show stable diagnostic performance (AUC = 0.885, 95% CI: 0.817–0.952). Conclusion The DLRC model demonstrated outstanding performance in differentiating between UP and PI, which facilitated the early identification of UP and the formulation of individualized diagnosis and treatment plans.