EHR-based AI model may enable expanded screening for primary aldosteronism, researchers report
Presenter: Frank G. Lee, MD, Mayo Clinic Rochester, Rochester, MN
AI-based approach to screening primary aldosteronism. Presented June 13, 2026.
A new artificial intelligence (AI) model trained on 30 years of routine electronic health record (EHR) data on 22,264 patients may enable better screening for primary aldosteronism, a common but frequently underrecognized cause of hypertension associated with increased cardiovascular risk, according to Mayo Clinic researchers who presented their findings at ENDO 2026.
The model correctly flagged more than 90% of cases of primary aldosteronism when researchers set the threshold to have high sensitivity. “During testing on patients with high blood pressure who had never been screened previously for primary aldosteronism, our model identified approximately two out of every three patients for further work-up,” said Dr. Frank Lee, who was one of the researchers and who presented the report.
The findings suggest that the AI model could enable broader screening for primary aldosteronism according to the researchers, consistent with recommendations in the 2025 Endocrine Society clinical practice guideline on primary aldosteronism.
Challenges in effective screening
Primary aldosteronism is a leading cause of hypertension, increasing patients’ risk of cardiovascular complications, including stroke, coronary artery disease, atrial fibrillation, heart failure, and renal disease. While its true prevalence is unknown, up to 20% of patients with hypertension are estimated to have primary aldosteronism, according to Dr. Lee. With effective treatments for primary aldosteronism available, early diagnosis could prevent future complications and reduce healthcare costs.
“Clinicians have been challenged to screen primary aldosteronism effectively. The tool developed by our team could offer a solution based on routine information available in a patient's medical records,” Dr. Lee said.
Detecting primary aldosteronism early
Researchers developed the AI screening model using de-identified data from adult patients treated at Mayo Clinic between 1986 and 2025. The main analysis included adults aged 18 years or older with a diagnosis of primary aldosteronism and a control group of patients with negative plasma renin-aldosterone screening results.
Demographic, clinical, laboratory, and medication data were extracted from records collected 30 days or 1 year before the index diagnosis or screening test. Variables included age, sex, hypertension- and hypokalemia-related ICD diagnoses, systolic blood pressure measurements, serum potassium levels, and prescriptions for antihypertensive medications or potassium supplements.
Using five-fold cross-validation, researchers developed an extreme gradient boosting model and assessed its discriminative performance with the area under the receiver operating characteristic curve and calibration with the Brier score. They also assessed the importance of features in improving model interpretability before applying the model to a holdout cohort of 225,887 adults with hypertension who had not undergone screening for primary aldosteronism and did not have a documented diagnosis of it.
A targeted screening approach
Among the 22,264 patients in the Mayo Clinic database, 1,833 had primary aldosteronism and 20,431 were in the control group. The median age was 67 in the primary aldosteronism group and 66 in controls, with 47.0% vs 54.7% female, respectively. The median time from the first model input variable to outcome occurrence was 4.0 years (range 0.3–10.9 years) in the primary aldosteronism group and 5.1 years (range 0.6–12.9 years) in the control group.
Compared with controls, patients in the primary aldosteronism group were more likely to have hypertension-related ICD diagnoses (74.6% vs 71.2%) and a history of hypokalemia (15.5% vs 8.9%). They also had higher median systolic blood pressure (132 vs 126 mm Hg) and lower median serum potassium levels (3.9 vs 4.2 mmol/L), while the median number of prescribed antihypertensive medications was similar between groups (2 vs 2). The model demonstrated moderate discrimination, with an area under the receiver operating characteristic curve of 0.71 for 30-day prediction and 0.67 for 1-year prediction.
Across moderate-risk probabilities (0.3–0.7), the model demonstrated good calibration, with a Brier score of 0.23. For 1-year prediction, a risk threshold of 0.3 yielded a sensitivity of 89% and a specificity of 26%, whereas a threshold of 0.6 reduced sensitivity to 28% but increased specificity to 92%.
When investigators further tested the model using data from 225,887 patients with hypertension and using a risk threshold of 0.3, they identified 153,561 patients (68.0%) for primary aldosteronism screening. Using a higher threshold of 0.6 reduced the number of patients selected for screening to 13,431 (5.9%).
Disclosures
Frank Guo Lee, MD, and Irina Bancos, MD reported no financial relationships to disclose.
References
Guo Lee F, Bancos I, Choudhary A et al. AI-based approach to screening primary aldosteronism. Presented at ENDO 2026, June 13, 2026, Chicago, IL.
Newswise. AI model identifies patients at risk of underdiagnosed cause of high blood pressure. June 13, 2026. https://www.newswise.com/articles/ai-model-identifies-patients-at-risk-of-underdiagnosed-cause-of-high-blood-pressure