The landscape of artificial intelligence-enabled medical devices in the EU and the US intended for intensive care units

Comprehensive Summary

This article looked at how many AI-enabled medical devices are actually on the market for ICU use in the US and EU, and whether they are ready for real-world clinical care. Using a multimethod search, the authors identified 36 ICU-specific AI devices in total. Of these, 21 were available in the US only, 9 were available in the EU only, and 6 were available in both regions. The first ICU device, Visensia, was approved in the US in 2008, and there was then a 10-year gap before authorizations resumed in the US in 2018 and in the EU in 2021. Most US devices (27) were classified as risk class II, while EU devices ranged from class I to IIb. In the US, most devices (24) reached the market through the 510(k) pathway, and only 3 used De Novo authorization. Functionally, the majority of devices (32 of 36) were designed for assessment rather than intervention. The most common tasks were prediction (13 devices) and quantification or feature localization (11 devices). The main input data were physiological signals such as vital signs or ECGs (15 devices), followed by EHR data (9) and imaging data (9). Compared with the broader FDA AI device landscape, ICU devices relied less on imaging and more on signals and EHRs, which fits the ICU setting where continuous monitoring and early warning matter most. The main practical meaning is that AI in critical care is already moving from research into real products, but these tools are still mostly narrow, task-specific systems. The study argues that approval alone does not prove patient benefit, so future work must focus on prospective outcome studies, better workflow integration, and stronger evaluation of real-world safety and usefulness.

Outcomes and Implications

At the bedside, this research means ICU AI should be used for continuous monitoring, early warning, prognosis, complication prediction, and the optimization of supportive therapy, not as a replacement for clinician judgment. The main change is that predictive scores and alerts must be embedded in escalation pathways, so they can support intensified monitoring, early diagnostics, or therapy initiation when patients deteriorate. Because ICU environments are time-critical and have high alarm burden, these tools need strong specificity, contextual prioritization, usability, and fit to clinical workflows. This also means that regulatory clearance or certification alone is not enough; prospective evaluation in real-world settings and post-market surveillance are needed to show clinically meaningful ICU outcomes and safety endpoints.