Acute-care AI is narrow and monitored
- The FDA cleared the first real-time cardiogenic-shock monitoring system that continuously reports SCAI-classification status at the bedside. - Other acute-care AI work shows mixed results: high agreement in some trial adjudications but failures in ED chest-pain pilots when tools mishandle live, partial data. - The pattern is clear: ICU-friendly AI succeeds when it is narrow, surveillance-focused, explainable, and tightly integrated into escalation workflows ( ).
In acute care, the AI systems gaining ground are the ones that watch one problem continuously and hand clinicians a bedside status update. (medscape.com) On April 3, 2026, the Food and Drug Administration cleared Etiometry’s platform under 510(k) number K254066, and the company announced the cardiogenic-shock tool on April 8. The software automates hospital-specific shock classification and tracking from physiologic monitoring data. (accessdata.fda.gov, etiometry.com) Cardiogenic shock is a state where the heart cannot pump enough blood to keep organs supplied, and hospitals often stage it the way weather services stage storms. The Society for Cardiovascular Angiography and Interventions system runs from stage A, “at risk,” through stage E, “extremis,” to give teams one shared scale. (scai.org, jscai.org) Etiometry said its tool combines bedside physiologic streams with electronic health record, laboratory, and device data, then shows a continuous view of hemodynamic status and organ function. The company said the clearance covers use from the operating room and intensive care unit to step-down units and telemetry floors. (etiometry.com) That is a narrower job than the emergency-department AI projects aimed at sorting undifferentiated chest pain the moment patients arrive. In the RAPIDxAI trial across 12 South Australian emergency departments and 3,029 patients, AI-assisted care did not reduce the main 6-month outcome, with events in 26.0% of the AI group and 26.4% of usual care. (acc.org, pubmed.ncbi.nlm.nih.gov) RAPIDxAI did change some process measures. Invasive coronary angiography in patients not classified as type 1 myocardial infarction fell to 5.2% from 9.4%, while statin use in patients classified as type 1 myocardial infarction rose to 81.8% from 68.0%. (acc.org) Other chest-pain models still look stronger in retrospective testing than in real-time deployment. A 2025 Open Heart study trained a neural network on 15,048 emergency visits and externally validated it on 14,476 more, with an area under the receiver operating characteristic curve of 0.82 versus 0.79 for a biomarker-only model. (openheart.bmj.com) The cleaner wins have come in back-office adjudication, where the AI reads completed records instead of fragmentary live data. A 2025 Nature Communications paper reported that a domain-specific large language model preadjudicated death, hospitalization, and medication-use events from 1,046 telephone-follow-up vignettes across three centers. (nature.com) A separate 2025 Circulation: Heart Failure study tested natural-language processing against a physician clinical-events committee in the DELIVER trial, which enrolled 6,263 patients, to identify heart-failure hospitalizations from trial records. The authors wrote that committee review remains the gold standard, but described the manual process as labor-intensive and poorly reproducible. (pubmed.ncbi.nlm.nih.gov) The common thread is not “AI doctor” software making open-ended calls. It is monitored software handling a bounded task, using a standard like SCAI staging, and feeding an escalation workflow that clinicians already use at the bedside. (scai.org, etiometry.com)
