AI Model Identifies High-Risk Primary Aldosteronism Patients

How does AI screening for primary aldosteronism work?

An AI model trained on more than 22,000 electronic health records can identify hypertensive patients at risk for primary aldosteronism with an AUROC of 0.71. The retrospective study used data from 1986‑2025 across community clinics and tertiary centers within the Mayo Clinic system, involving 1,833 confirmed cases and 20,431 controls. At a 0.5 risk threshold the algorithm correctly flagged about 70 % of true cases while keeping false positives low. Mayo Clinic researchers note, “The model achieved an AUROC of 0.709, indicating moderate discrimination” (2026).

How machine learning improves hypertension risk assessment

Researchers assembled an ensemble of three tree‑based learners—eXtreme Gradient Boosting, random forest and extremely randomized trees—and fed them variables such as age, sex, diagnosis codes, vital signs, lab values and medication lists. The training set spanned nearly four decades, giving the algorithm exposure to evolving treatment patterns. After training, performance was measured on a held‑out test set of hypertensive patients, reporting an estimated calibration error of 0.037, which suggests the predicted probabilities line up reasonably with observed outcomes. And because the model was evaluated only within the same health system, its behavior elsewhere remains unknown.

Can machine learning reduce false positives in hypertension diagnosis?

An AUROC of 0.71 tells us the model can separate those with primary aldosteronism from those without better than chance, but it is far from perfect. The calibration error of 0.037 indicates that, on average, predicted risk scores are within a few percentage points of the actual rate. The median time from a patient’s first clinic encounter to a formal PA work‑up was five years, meaning many people wait a long time before diagnosis. So, if the tool were embedded in the EHR, clinicians could be nudged to order confirmatory tests years earlier, potentially shortening that delay.

What is the future of automated diagnostic algorithms in clinical settings?

The paper is a preprint and has not yet undergone peer review, so its methods and conclusions have not been independently vetted. It draws on data from a single, integrated health network, which limits generalizability to other regions or health systems. The analysis is observational; the AI’s predictions are associations, not proof that any particular factor causes the disease. And because the model was not tested prospectively, we don’t know how it would perform in real‑time clinical workflows.

How will AI impact everyday clinical practice for hypertension?

If integrated into an EHR, the algorithm could generate alerts for patients whose risk score exceeds a preset threshold, prompting providers to consider aldosterone‑renin testing. This could raise screening rates, which currently sit well below guideline recommendations. However, clinicians would still need to interpret the alert in context and order confirmatory labs, so the tool is an aid, not a replacement for clinical judgment. But the modest false‑positive rate means fewer unnecessary tests compared with blanket screening of all hypertensive patients.

What are the next steps for AI-driven hypertension screening?

The authors plan external validation using data from other institutions to see if performance holds up across diverse populations. A prospective trial that measures whether the alert actually shortens time to diagnosis and improves blood‑pressure control would be the next logical proof point. Regulatory bodies may eventually require evidence of safety and effectiveness before endorsing such AI‑driven decision support. So, keep an eye on follow‑up studies that move this from a promising algorithm to a validated clinical tool.