Abstract
OBJECTIVE: Current subtype diagnosis of primary aldosteronism relies on adrenal venous sampling and imaging, each with inherent limitations. Lesional adrenal glands with negative CT Imaging is a distinct subtype of primary aldosteronism that has been less frequently studied. The aim of this study was to develop and validate a machine learning and AI model for distinguishing adrenals with transversely negative lesions from normal adrenals Primary Aldosteronism. MATERIALS AND METHODS: We conducted a single-center retrospective study, assessing transverse adrenal scans of 170 PA patients. A specialized iterative method was employed for radiomic feature selection. Subsequently, six conventional machine learning methodologies were utilized to construct the radiomics models. This original data was subsequently applied in the construction of a radiomic model, which was combined with clinical data for the final model construction. RESULTS: 107 radiomic features were extracted from the adrenal scans and 10 features were selected for ML and AI modeling. In the clinical data, values for serum potassium, aldosterone excretion, uric acid, and IVSd were utilized in the model construction. The integration of clinical data further enhanced the model's performance, with an AUC reaching 0.868 in the derived cohort, and an AUC of 0.853 in the temporal validation cohort. CONCLUSION: The study indicates that clinical-radiomic scores can independently serve as diagnostic biomarkers for the specialized PA subtype categorization. We give the proposal for the precise categorization concept in establishing a clinical-radiomic model for PA subtype diagnosis. The model demonstrates substantial potential for both clinical and translational research.