Researchers presented an AI-based screening approach for primary aldosteronism at ENDO 2026 that leveraged three decades of de-identified electronic health record (EHR) data. The development set comprised more than 22,000 patients drawn from the Mayo Clinic Platform, a federated, privacy-preserving infrastructure containing multimodal clinical data collected between 1986 and 2025. After model training, the team tested the algorithm on a substantially larger set of 225,887 adults with hypertension.
The study team included Frank Lee, M.D., of Mayo Clinic, who served as the lead researcher, and Irina Bancos, M.D., MS, who participated in the ENDO 2026 discussion. The source article reports the retrospective analysis and model testing as presented at the meeting; additional publication details or peer-reviewed manuscript identifiers were not reported in the source.
The AI screening model used routinely recorded clinical features available in EHRs. Input variables reported in the source include:
The team implemented an XGBoost architecture, a gradient-boosted decision-tree machine learning library, to predict which patients were at risk of primary aldosteronism. The model was designed to identify risk up to 12 months before a formal clinical diagnosis in the test cohort.
When evaluated on the test cohort, the model demonstrated a high case-detection rate at the threshold set to identify people at low risk. Specifically, at that operating point the algorithm correctly flagged more than 90% of primary aldosteronism cases while missing fewer than 10%. Using the same threshold, approximately two-thirds (about 66%) of participants with high blood pressure who had never been screened previously for primary aldosteronism were identified by the model as candidates for additional work-up.
In practical terms, the investigators reported that the tool could identify roughly two out of every three unscreened patients with hypertension for further diagnostic evaluation. The source frames these results as evidence that an AI model based on routinely available medical-record information may be a feasible approach to selective screening for primary aldosteronism.
Primary aldosteronism occurs when the adrenal glands produce excess aldosterone, leading to sodium retention, potassium depletion, and elevated blood pressure. Patients with primary aldosteronism face greater risk of cardiovascular disease than people with primary (essential) hypertension.
The true population prevalence is uncertain, but the study lead reported an estimate that up to 20% of patients with hypertension may have primary aldosteronism. Because effective targeted treatments exist, earlier diagnosis can alter prognosis by reducing cardiovascular complications and may also lower healthcare costs through more appropriate management.
The source article highlights the Endocrine Society’s clinical practice guideline released in 2025, “Primary Aldosteronism: An Endocrine Society Clinical Practice Guideline,” which calls for expanded screening for this condition. The investigators positioned their AI tool as potentially supporting the guideline recommendation by facilitating more efficient identification of patients who warrant biochemical testing and specialist referral.
The source provides key methodological and performance findings but does not report several details important for clinical implementation. Specifically, the article does not include full external-validation results beyond the described test cohort, prospective implementation data, calibration statistics, or information on how the model performs across demographic or clinical subgroups. Potential sources of bias inherent to EHR-derived models and the operational workflow required to integrate such a model into practice were not described in the source.
Clinicians and health systems considering adoption would need additional information not reported here: prospective validation in other populations, decision thresholds aligned with local screening capacity, clinical pathways for confirmatory testing, and assessment of equity and unintended consequences. The study team reported feasibility and promising retrospective performance at ENDO 2026, but the source does not detail next steps for prospective trials or regulatory review.
In summary, an XGBoost-based model trained on more than 30 years of Mayo Clinic Platform EHR data and tested on 225,887 adults with hypertension identified a high proportion of primary aldosteronism cases when set to a low-risk threshold. The tool flagged most cases up to 12 months before diagnosis and recommended selective biochemical screening for roughly two-thirds of unscreened patients with high blood pressure. Investigators emphasized that earlier detection is clinically meaningful because targeted therapies can lower cardiovascular risk. The source did not report prospective implementation outcomes, broader external validation, or detailed subgroup performance, and those areas would be important next steps before routine clinical deployment.