S-47-2: DIAGNOSIS OF PRIMARY ALDOSTERONISM USING ARTIFICIAL INTELLIGENCE

Abstract

Testing for primary aldosteronism (PA) involves an arduous multistep process culminating in either resection of an adrenal in patients identified with unilateral disease after adrenal venous sampling (AVS) or mineralocorticoid receptor blockade in patients without AVS-based evidence of lateralized aldosterone secretion. Multiple difficulties plague each step, including the confirmatory step where immunoassay interferences lead to many false-positive results (1). Mass spectrometry-based steroidomics provides an alternative approach that combined with artificial intelligence might be used for both screening and stratification of patients for therapeutic intervention. The approach may be particularly useful for identification of patients with unilateral disease due to somatic KCNJ5 mutations, who after imaging studies may proceed directly to adrenalectomy without need for confirmatory studies or AVS. To establish this approach, we developed machine learning (ML) models using a retrospective cohort of 462 patients tested for PA, including 139 with unilateral disease of whom 58 had KCNJ5 mutations (2) ML models are now undergoing external validation in a prospective multicenter study (PROSALDO), which has to date enrolled 542 eligible patients. Among these 186 have received follow-up, which has enabled exclusion of disease in 73 and confirmation of unilateral disease in 38 others according to post-surgical biochemical cure. Among the latter, genotyping in 19 patients identified KCNJ5 mutations in nine, all of whom showed post-operative cure and were correctly predicted by ML to have PA with KCNJ5 mutations. Among these nine and four others who were not part of the PROSALDO trial (3) 10 had their adrenals removed without AVS evidence of lateralization. In an interim analysis (with recognition that ML models were likely compromised by incorrect classifications due to immunoassay inaccuracy), among the 542 PROSALDO patients, 138 and 404 were respectively classified with and without PA. According to areas under receiving-operating characteristic curves (AUC), our ML model for screening performed equally poorly (AUC = 0.927) compared to aldosterone:renin ratios (ARR) determined by mass spectrometry (AUC = 0.909) and immunoassay (AUC = 0.895) measurements of aldosterone. Significantly improved performance was achieved by incorporation of plasma potassium and renin with steroidomics-based models (AUC = 0.9511). Results to date indicate promise for ML approaches to both more accurately screen for PA than possible with the ARR and also identify patients with unilateral disease due to KCNJ5 mutations who may directly proceed to surgery without need for confirmation studies or AVS.