Cytokeratin 7 (CK7) expression in clear cell renal cell carcinoma (ccRCC) may reflect tumor phenotype and biological heterogeneity, but its clinical role remains exploratory. This study aimed to investigate associations between preoperatively available clinical and CT imaging features and CK7 expression and to develop an exploratory machine learning model for CK7 characterization. This multicenter retrospective study included 230 patients with pathologically confirmed ccRCC from three institutions, including 139 patients in the training cohort and 91 patients in the validation cohort. Clinical variables and qualitative and quantitative CT features were assessed by two radiologists. CK7 positivity was defined as cytoplasmic staining in at least 10% of tumor cells. After least absolute shrinkage and selection operator regression, four machine learning models using prespecified configurations were developed in the training cohort and evaluated in the validation cohort. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive predictive value, negative predictive value, accuracy, F1 score, calibration analysis, and decision curve analysis. Classification thresholds were determined in the training cohort using the maximum Youden index and then applied to the validation cohort. Model stability was assessed using 1000 bootstrap resampling iterations. Among 230 patients, 67 had CK7-positive tumors. The final predictors included age, cystic component, corticomedullary attenuation, corticomedullary enhancement ratio, nephrographic attenuation, and tumor size. The logistic regression model was selected as the final model based on its discrimination, calibration performance, interpretability, and overall generalizability. In the validation cohort, the logistic regression model achieved an AUC of 0.714 (95% CI: 0.589, 0.828), with a sensitivity of 0.625 and a specificity of 0.627. Bootstrap resampling supported the stability of performance estimates. This exploratory machine learning model-development study suggests that preoperatively available clinical and CT imaging features may be associated with CK7 expression in ccRCC. Further prospective validation using larger cohorts is required before clinical translation. CK7 status may provide complementary information on tumor phenotype in ccRCC. The LR model showed balanced discrimination, calibration, and interpretability. Preoperative clinical and CT features may help characterize CK7-related tumor heterogeneity.
Lin et al. (Mon,) studied this question.