SUN-367 Artificial Intelligence Systems for Predicting Primary Aldosteronism Subtype

Background: Primary aldosteronism (PA) is a major form of secondary hypertension, and is classified into two subtypes: unilateral and bilateral disease. Subtype classification is important for identifying patients with unilateral PA who can be cured by unilateral adrenalectomy. The gold standard for subtype classification of PA is adrenal vein sampling (AVS), but relatively invasive and costly. The use of artificial intelligence, especially machine learning, having been introduced in some fields of clinical prediction, may be more simple method to identify unilateral PA. Objective: The aim of this study was to develop and compare new prediction systems for diagnosis of PA subtype using machine-learning methods. Methods: We retrospectively analyzed 373 patients with PA (55 ± 11 yrs., female 49%) who underwent AVS at Kanazawa University Hospital between 2005 and 2017. We performed unstimulated bilateral AVS in the morning, and unilateral aldosterone overproduction was confirmed by calculating the ipsilateral adrenal vein aldosterone-to-cortisol concentration ratio divided by the contralateral aldosterone-to-cortisol ratio (i.e., lateralized index ≥ 2). The subjects were diagnosed as having unilateral (n = 192) or bilateral PA (n = 181) based on AVS. After selecting a subset of 31 candidate variables from the literature, we constructed prediction models by five widely-used machine learning classifiers: neural network (multilayer perceptron: MLP), support vector machine (SVM), random forest (RF), k-nearest neighbor (K-NN) and logistic regression (LR). These models were validated in following steps: (i) parameter tuning with randomized ten-fold stratified shuffle split cross-validation (training : validation = 80% : 20%) , repeated 10 times to search candidate parameters, then more 20 times to find best parameters; and (ii) automatic variable selection using recursive feature elimination with RF and LR, also repeated 20 times with randomized ten-fold stratified shuffle split cross-validation to find best feature subset for each machine learning classifier with their best parameters. Finally, the best prediction model was decided based on area under the receiver operating characteristic curve (AUC) mean in the step (ii). Results: The best model was constructed using SVM trained by feature subset based on logistic regression. The AUC mean was 0.660 (max 0.752) and the subset consisted of following eight variables: plasma aldosterone concentration, aldosterone-to-renin ratio, serum potassium, fasting plasma glucose, HDL-Cholesterol, plasma ACTH, adrenal nodule on CT scan, and potassium supplement. Conclusions: We built a new subtype diagnosis prediction system of PA only using a general laboratory test in a single center. This study showed the possibility that machine learning might be useful for the prediction of PA subtype.