SUN-238 A Machine Learning Model for Risk Stratification of Primary Aldosteronism

Abstract Disclosure: K. Zekarias: None. J. Kohlenberg: None. M. Xu: None. S. Ma: None. Background: Primary aldosteronism (PA) is a leading, often undiagnosed cause of secondary hypertension. Objective: This study aimed to develop an electronic-health-record (EHR) based machine learning model to predict PA in people with hypertension. Methods: We conducted a retrospective cohort study utilizing EHR data from 1,649 people screened for PA between April 13, 2018, and May 2, 2023 at an academic health system in the upper Midwest, USA. PA screening was conducted using standardized guideline recommended risk factor assessment. PA diagnosis was defined as 1) urine aldosterone > 12 µg/24h after oral salt loading or 2) PAC > 10 ng/dL 4 hours after saline infusion or 3) plasma aldosterone concentration (PAC) > 20 ng/dL with renin < 1 ng/mL/h or 4) Aldosterone to Renin Ratio (ARR) > 50 with PAC ≥ 10 ng/dL. We extracted 74 potential PA predictor variables from EHR data and applied various feature selection and classification methods to develop predictive models. Results: Among 1,649 people screened for primary aldosteronism (PA), 166 (10%) were diagnosed with PA. Compared to people without PA, those with PA were older (mean age 61 ± 13 years vs. 56 ± 17 years, p=0.001). There was no difference in gender (48% male, 52% female, p=0.7). The prevalence of PA among those screened was higher in Black patients (20%) compared to White patients (7%) and other racial groups (5%), with a statistically significant difference (p < 0.001). Those with PA exhibited increased end organ damage (48%) compared to those people without PA (37%, p=0.02).The model selection procedure resulted in a logistic regression model with 13 predictors, with cross validated area under the receiver operating characteristic curve (AUC) 0.7+/-0.021. Key risk factors predicting PA diagnosis were age, race, body mass index, hypokalemia, sodium and bicarbonate concentrations, high-density lipoprotein levels, obstructive sleep apnea, end stage renal disease, uncontrolled blood pressure (B/P >130/80), and prescription of > 4 blood pressure medications, diastolic blood pressure > 80, and Glomerular Filtration Rate (GFR).Discussion/Conclusion: We have developed a predictive model with moderate performance designed to optimize PA screening through risk stratification. Notably, the cohort in this study was screened for PA following current clinical guideline recommendations. The fact that our model achieved AUC=0.7 indicates that additional clinical information not currently included in the screening guideline can enhance screening performance. Following rigorous validation with prospective data, we aim to integrate this prediction tool directly into the clinical workflow as a decision support tool for PA screening. The clinical significance of this approach lies in its potential to enhance targeted screening for PA and to improve screening rates by leveraging a clinical workflow decision support tool. Presentation: Sunday, July 13, 2025