An artificial intelligence model trained on routinely collected longitudinal electronic health record data identified patients at risk for primary aldosteronism up to 1 year before diagnosis, according to a research abstract presented at ENDO 2026, the Endocrine Society's annual meeting, in Chicago. 

Current Endocrine Society guidelines recommend universal screening for primary aldosteronism (PA) in patients with hypertension. Because universal screening can be difficult to implement in routine practice, researchers from Mayo Clinic in Rochester, Minnesota, sought to develop an AI tool to prioritize patients for testing.  

“The scope of impact is substantial as nearly half of US adults have hypertension, and an estimated 5% to 15% of them have PA,” first author Frank Lee, MD, told AACE Endocrine AI. “Early screening is critical. Identifying PA sooner can prevent years of unrecognized disease exposure that contributes to atherosclerotic disease, stroke, and heart attacks.” 

The Mayo Clinic team developed and validated an extreme gradient boosting (XGBoost) machine learning model using de-identified electronic health record data collected between 1986 and 2025 through the Mayo Clinic Platform, a federated infrastructure that contains multimodal clinical data.  

The analysis included 22,264 adults, comprising 1,833 patients with PA and 20,431 control patients who had negative plasma renin and aldosterone screening results. Predictor variables were extracted either 30 days or 1 year before diagnosis or screening and included age, gender, hypertension- and hypokalemia-related diagnosis codes, systolic blood pressure, serum potassium levels, and prescribed antihypertensive medications and potassium supplements. 

Compared with controls, patients with PA more frequently had hypertension-related diagnoses (75% vs 71%) and hypokalemia-related diagnoses (16% vs 9%), as well as higher median systolic blood pressure (132 vs 126 mm Hg), and lower median potassium concentrations (3.9 vs 4.2 mmol/L). Median age was similar between groups (67 for PA vs 66 for control), while women comprised 47% of the PA cohort and 55% of controls. 

Using fivefold cross-validation, the model achieved an area under the receiver operating characteristic curve (AUROC) of 0.71 for predicting PA 30 days before diagnosis and 0.67 for predicting risk 1 year before diagnosis. The model demonstrated moderate discrimination and was best calibrated across moderate risk probabilities (0.3 to 0.7), with a Brier score of 0.23. 

Model performance varied according to the selected risk threshold. For 1-year prediction, a threshold of 0.3 yielded 89% sensitivity and 26% specificity, whereas a threshold of 0.6 yielded 28% sensitivity and 92% specificity. 

When applied to a separate cohort of 225,887 patients with hypertension, the model identified 153,561 patients (68%) for PA screening at the lower risk threshold of 0.3. Raising the threshold to 0.6 reduced the number recommended for screening to 13,431 patients (6%). 

The researchers noted that the tool is not intended to replace the current standard of care for PA, and lab-based assessments remain essential to the diagnostic pathway. “What this model offers is a prescreening strategy for clinicians that continuously updates a patient's risk score with each clinical encounter and presents an opportunity for a more dynamic approach than a single, static lab test,” said Dr. Lee. “We believe it represents an important first step toward more informed, data-driven clinical decision-making for PA.” 

Dr. Lee and Irina Bancos, MD reported having no financial relationships to disclose. The other researchers did not share conflicts of interest.