SAT-830 - AI-Based Approach to Screening Primary Aldosteronism

Introduction
The 2025 Endocrine Society Primary Aldosteronism (PA) guidelines recommend universal screening for all patients with hypertension but targeted approaches are needed to practically guide referrals and testing. We aimed to develop and validate an artificial intelligence (AI)-based tool to identify at-risk patients, with 30-day and 1-year lead-time, using longitudinal electronic health records (EHR) data.

Methods
Multi-site de-identified EHR data from Mayo Clinic patients from 1986-2025 was utilized for development and validation. Adults (≥18 years old) with a PA diagnosis (disease cohort) or a negative plasma renin/aldosterone screening test (control cohort) were included. Variables were abstracted 30 days or 1 year prior to diagnosis or screening test, and included age, gender, hypertension- and hypokalemia-related ICD diagnoses, systolic blood pressure (BP) measurements, potassium blood levels, and prescribed antihypertensive or potassium supplement medications. An XGBoost-based model was developed using 5-fold cross-validation. The model was assessed for discriminative performance using AUROC and calibration with Brier score. Feature importance was analyzed for model explainability. The model was then applied to a holdout validation dataset of hypertension patients without PA diagnosis or screening.

Results
Overall, 22,264 patients (1,833 in the disease cohort; 20,431 in the control cohort) were identified. Median age of disease cohort was 67 vs. 66 for controls; with 47.0% vs. 54.7% female respectively. Median timeframe from first model input variable to outcome was 4.0 y (range: 0.3-10.9) for disease group and 5.1 (0.6-12.9) for controls. A summary of the predictor variables in the disease vs. control groups were: 74.6% vs. 71.2% with hypertension-related ICD diagnoses, 15.5% vs. 8.9% with hypokalemia ICD diagnosis, median SBP 132 vs. 126, median potassium level 3.9 vs. 4.2, and median number of prescribed antihypertensive medications of 2 vs. 2. The model had an AUROC of 0.71 for 30-day prediction and AUROC 0.67 at 1-year. The model was well-calibrated for moderate risk probabilities (0.3-0.7) with Brier score 0.23. A risk threshold of 0.3 had 89% and 26% specificity for 1-year prediction, whereas a risk threshold of 0.6 had 28% sensitivity and 92% specificity. When applied to a holdout test cohort of 225,887 patients with hypertension, the model identified 153,561 (68.0%) for screening using a risk threshold setting of 0.3. A risk threshold setting of 0.6 identified 13,431 (5.9%) for screening.

Conclusion
An AI model using data automatically abstracted from the longitudinal health record was able to identify patients at risk of PA with a 1-year lead-time. Adjusting the risk threshold can provide a granular/targeted approach to screening and offer cost-effective advantages to current universal screening guidelines.

*Unless otherwise noted, all abstracts presented at ENDO must not be released to the press or the public until the date and time of presentation. For oral presentations, the abstracts are embargoed until the session begins. The Endocrine Society reserves the right to lift the embargo on specific abstracts that are selected for promotion prior to or during ENDO.*