Validating Machine Learning Models Against the Saline Test Gold Standard for Primary Aldosteronism Diagnosis

Figure 1. Feature Importance of Random Forest During Feature Selection (A) The bar chart depicting feature importance of the first feature selection subset for predicting essential hypertension and unilateral primary aldosteronism. (B) The box plot of permutation feature importance of the subset. (C) The dendrogram of feature correlations of the subset. (D) The heatmap of feature correlations of the subset. COPD = chronic obstructive pulmonary disease; dBP = diastolic blood pressure; HTN = hypertension; Hyper = hyperthyroidism; myocardiac = myocardial infarction; PRA = plasma renin activity; SBP = systolic blood pressure.

Figure 2. The AUROC of EH-uPA-biPA of Test Set (A) Receiver-operating characteristic (ROC) curve of Random Forest and the AUC value is 0.968. (B) ROC curve of deep neural networks (DNN) and the AUC value is 0.943. (C) ROC curve of XGBoost and the AUC is 0.974.

Figure 3. Waterfall Plots of SHAP Values Exploring the Important Expectation (A) The first extracted feature for essential hypertension vs unilateral primary aldosteronism of 1 sample. (B) The first extracted feature for essential hypertension vs unilateral primary aldosteronism of all samples. (C) All extracted features of 1 sample. (D) All extracted features of all samples. Blue symbols depict a negative effect and red depict a positive effect. The number represented the strength of the influence. The x-axis was the impact value. SHAP = Shapley additive explanations; other abbreviations as in Figure 1.

Figure 4. Feature Selection Disparities Across Sites and Permutation Feature Importance Disparities This figure shows feature selection disparities across sites and permutation feature importance disparities for the models trained with source data from the TAIPAI data set and the features permuted in the 3 classifiers (ie, EH vs uPA, EH vs biPA, and uPA vs biPA) at each validation site (ie, South Korea and Mainland China) using all the features (please see Supplemental Methods). Aldost = aldosterone; hypert = hypertension duration (years); other abbreviations as in Figure 1.

Central Illustration. The Proposed Features Extracted for Confirming Primary Aldosteronism This figure demonstrates the validation of machine learning models (Random Forest, XGBoost, and Deep Neural Networks (DNN)) against the saline infusion test (SIT) for diagnosing primary aldosteronism (PA). Data from 3,833 patients (1,978 PA and 1,855 essential hypertension cases) across 3 regions were used, with feature selection and extraction applied to improve model performance. XGBoost achieved the highest accuracy (0.889) and area under the receiver-operating characteristic curve (AUROC) (0.974), followed closely by Random Forest (Accuracy: 0.883, AUROC: 0.968) and Deep Neural Networks (Accuracy: 0.887, AUROC: 0.943). The results indicate that feature selection and extraction methods enhance diagnostic accuracy, making the machine learning models a comparable alternative to traditional SIT diagnosis for PA.