# AI in Hypertension: Evaluate the Clinical Task

Artificial intelligence may help organize measurements, estimate risk, or support clinical decisions. A useful assessment starts with the intended task, the data used, and whether a tool improves patient care in the setting where it will actually be deployed.

**Learning goal:** Connect assessment, evidence, and a clear next clinical decision. Educational use; individual care requires the treating team’s assessment and applicable protocols.

## Name the task

Distinguish measuring blood pressure, predicting future hypertension, identifying possible secondary causes, and suggesting treatment. These are different problems with different standards. A model that predicts a future event does not establish a current diagnosis or prove that its preferred treatment improves outcomes.

## Examine the reference standard

Ask how the model’s labels and outcomes were obtained and whether measurements were reliable. Inaccurate cuff readings or incomplete records can teach an algorithm the wrong pattern. A cuffless or image-based estimate needs evaluation for its intended measurement use, not just correlation with another value.

## Demand external validation

Performance in development data may not persist at another hospital, time period, or patient population. Review calibration, clinically important errors, missing-data handling, and subgroup performance. A high discrimination score can coexist with poor absolute-risk estimates or unequal errors across patient groups.

## Test the clinical workflow

Identify who receives an output, how they verify it, and what action it changes. Human review is meaningful only when the reviewer has enough information and time to question the result. Prospective evaluation should consider outcomes, workload, false alerts, and unintended treatment changes.

## Protect patients and data

Use authorized data access, clear responsibilities, and appropriate privacy protections. Monitor performance after deployment as patients, devices, and practice change. The NHLBI workshop report identifies substantial implementation and bias challenges; promise in a research paper is not evidence of universal clinical readiness.

## Teach a safe use case

Consider an assistant that flags missing home-pressure readings for staff review. The staff member checks context, contacts the patient, and documents the response. This defined support task is easier to assess than an opaque system that changes medicines without verification or an accountable clinical pathway.

## Apply the framework

Does excellent retrospective accuracy prove an AI tool should change a patient’s medicines?

Show the reasoning

No. The intended treatment decision, external validity, safety, workflow, and clinical outcomes must be evaluated. Retrospective prediction is only one part of the evidence.

## Continue learning

- [Hypertension curriculum](https://urinenephrology.org/2025_UDPA_Lectures_Live/hypertension/index.html)
- [Digital care workflow](https://urinenephrology.org/2025_UDPA_Lectures_Live/hypertension/digital-health-integration.html)

## References and evidence

These sources support the teaching framework. Trial populations, endpoints, and limitations should be checked before applying a result to an individual patient.

1.  Shimbo D, Shah RU, Abdalla M et al.. Transforming Hypertension Diagnosis and Management in The Era of Artificial Intelligence: A 2023 National Heart, Lung, and Blood Institute (NHLBI) Workshop Report. Hypertension. 2025;82(1):36-45. [PubMed 39011653](https://pubmed.ncbi.nlm.nih.gov/39011653/)
2.  Cho JS, Park JH. Application of artificial intelligence in hypertension. Clin Hypertens. 2024;30(1):11. [PubMed 38689376](https://pubmed.ncbi.nlm.nih.gov/38689376/)
3.  Tsoi K, Yiu K, Lee H et al.. Applications of artificial intelligence for hypertension management. J Clin Hypertens (Greenwich). 2021;23(3):568-574. [PubMed 33533536](https://pubmed.ncbi.nlm.nih.gov/33533536/)


---

[Website version](https://urinenephrology.org/2025_UDPA_Lectures_Live/hypertension/artificial-intelligence.html) · Markdown synchronized October 3, 2026.