Predicting primary aldosteronism: risk stratification-based and machine learning-based models

Primary aldosteronism (PA) screening is difficult in an unselected population of hypertensive patients, as it can be challenging. We report new algorithms based either on risk stratification or machine learning models to improve the screening of PA.

15,507 patients with confirmed hypertension aged 18-65 years with full workup for PA screening were included in this study. We developed PAstrat, a risk stratification-based algorithm by generating patient groups of interest and assess probability of PA in each group, and two different machine learning models, logistic regression and XGBoost. An external cohort of 768 patients was used for validation. The AUC of the ROC curves for the PAstrat algorithm, logistic regression, and XGboost were, respectively, 0.80, 0.82, and 0.83 in the derivation cohort and 0.79, 0.82, and 0.82 in the validation cohort. Only 0.7% of the population (6 of 768) or 4.2% of patients with PA (6 of 142) had PA despite incorrect prediction by PAstrat (negative predictive value, 0.96). Conversely, among the 128 patients who had an indication for PA screening based on age, kalaemia, or drug resistance but not according to the algorithm, only 3 (0.23%) patients had PA. Logistic regression and XGBoost had, respectively, 49 (6.7%) and 84 (10.9%) false negatives (negative predictive value, respectively, 0.87 and 0.88).

We developed a risk stratification algorithm for PA screening and compared its performances to machine learning models. Our algorithm demonstrated good sensitivity and negative predictive value for the screening of secondary hypertension, with the added value of great interpretability. Machine learning models lacked explainability and demonstrated a lower negative value in the validation cohort which makes them less easily usable.