Diagnostic Accuracy And Clinical Safety Of Artificial Intelligence For Detecting And Flagging Critical Laboratory Results: A Systematic Review

Abstract

Background: Delayed or missed critical laboratory results remain a preventable source of patient harm. The practice of flagging them dates to George Lundberg’s original 1972 definition of a critical value as a result “life-threatening unless something is done promptly” (Lundberg, 1972), and the manual call-and-document workflow built around that definition has changed remarkably little in the half-century since. Artificial intelligence (AI) and machine learning (ML) tools are now increasingly proposed to detect or flag these results faster and more reliably than existing threshold-based alerting.

Objective: To evaluate the diagnostic accuracy and clinical safety of AI/ML systems for detecting or flagging critical laboratory and physiological results, and to examine how far the regulatory and implementation evidence has kept pace with the technical claims.

Methods: We searched PubMed/MEDLINE, Embase, CINAHL, Scopus, Web of Science, IEEE Xplore, and the Cochrane Library from [start date] to [search date], following PRISMA 2020 and its diagnostic test accuracy extension, PRISMA-DTA. Two reviewers screened records independently. [Risk of bias was assessed with QUADAS-2, supplemented by the QUADAS-AI framework, and certainty with GRADE.] Because study designs and reference standards varied too widely to pool, findings were synthesised narratively by critical-value category.

Results: Of 89 records, 11 studies were included. AI-enabled electrocardiography for detecting severe hyperkalemia was the most mature application, with two independent studies. Hypokalemia detection, anemia detection, and glucose-abnormality prediction were each represented by one to three studies, with wider and less consistent accuracy ranges. No included study reported a clinical safety outcome such as notification time, missed-result rate, or clinician override behaviour, and none tested for automation bias.

Conclusions: AI and machine learning can detect several types of critical result with accuracy approaching or matching laboratory confirmation, most convincingly for ECG-based hyperkalemia detection. Whether these tools improve, or compromise, the safety of the notification pathway itself is currently unknown, because almost no study in this literature has measured it.