Non-invasive Prediction of CYP11B2-Defined Subtypes in Primary Aldosteronism Using <sup>18</sup>F-Pentixafor PET/CT and Machine Learning.
Authors
Affiliations (3)
Affiliations (3)
- PET/CT Department, The Second Affiliated Hospital of Harbin Medical University, No. 246, Xuefu Road, Nangang District, Harbin, Heilongjiang Province, China.
- Department of Pathology, The Second Affiliated Hospital of Harbin Medical University, No. 246, Xuefu Road, Nangang District, Harbin, Heilongjiang Province, China.
- PET/CT Department, The Second Affiliated Hospital of Harbin Medical University, No. 246, Xuefu Road, Nangang District, Harbin, Heilongjiang Province, China. Electronic address: [email protected].
Abstract
This study aims to develop and validate an interpretable machine learning model that integrates clinical data, radiomics, and deep learning (DL) features extracted from <sup>18</sup>F-AlF-NOTA-Pentixafor positron emission tomography/computed tomography (PET/CT) images for the non-invasive prediction of pathological subtypes in primary aldosteronism (PA). In this single-center retrospective study, we included 89 patients diagnosed with PA or non-functioning adrenal adenomas who underwent <sup>18</sup>F-Pentixafor PET/CT between February 2024 and May 2025. Predictive models were built by integrating clinical data, PET/CT radiomics, and DL features. A two-stage feature selection strategy was employed, which utilized the minimum redundancy maximum relevance method followed by stepwise regression based on the Akaike information criterion. Four distinct models were constructed using the support vector machine algorithm, and their hyperparameters were optimized via stratified five-fold cross-validation. Model performance was rigorously evaluated by the area under the receiver operating characteristic curve (AUC), calibration analysis, and decision curve analysis. Furthermore, model interpretability was achieved using Shapley Additive Explanations (SHAP) to elucidate feature contributions. The combined model demonstrated superior diagnostic accuracy in the test set, with an AUC of 0.907, perfect sensitivity (1.000), and an F1-score of 0.923. It significantly outperformed the clinical, radiomics, DL models individually (p<0.01). SHAP analysis identified lesion-to-adrenal ratio, maximum standardized uptake value, and selected PET/CT radiomics and DL features as key contributors, revealing biological alignment with CXCR4 and CYP11B2 expression. An interpretable machine learning model can non-invasively predict surgically confirmed PA subtypes, defined by immunohistochemistry for CYP11B2. This approach may reduce the reliance on invasive adrenal vein sampling and facilitate personalized surgical decision-making.