An interpretable machine learning model for preoperative prediction of aldosterone secretion and CYP11B2 status of adrenal gland

BACKGROUND: To investigate whether a machine learning (ML) model integrating CT-based radiomics and clinical features can noninvasively evaluate the aldosterone secretion and CYP11B2 status of the adrenal gland, using Shapley Additive Explanations (SHAP) for model interpretation. METHODS: Patients who underwent adrenalectomy with CYP11B2 immunohistochemistry (IHC) confirmation between January 2022 and February 2025 were analyzed retrospectively. Radiomics models were developed based on different radiomics features selected by ElasticNet. Clinical features were used to construct the clinical model. The radiomics score and clinical features were integrated into the combined model. The SHAP method was applied to calculate feature importance. An external validation cohort from a second center was additionally analyzed to assess generalizability. RESULTS: In total, 140 patients were enrolled in the study. The clinic-radiomics model exhibited superior performance, achieving average AUC values of 0.912, 0.923 and 0.958 in distinguishing classical histopathology, nonclassical histopathology, and non-functional adrenal adenoma (NFA), respectively. The decision curve analysis showed that the combined model performs best. The SHAP method ranked the five features according to their importance. In the external cohort (n = 50), the model achieved AUCs of 0.807, 0.735, and 0.722 for classical histopathology, nonclassical histopathology, and NFA, respectively. CONCLUSION: A machine learning model integrating radiomics and clinical features enables noninvasive prediction of adrenal aldosterone secretion function and histopathological subtypes, with preliminary evidence of generalizability in an independent external cohort.