Written by Chris Thom
Spoon Feed
In this prospective trial of 209 patients, artificial intelligence software was able to accurately measure systolic and diastolic cardiac function.
Is the future of POCUS in AI?
AI is making inroads in nearly all fields of medicine, with POCUS being no exception. AI can help democratize POCUS skills by making image acquisition and interpretation easier for end users. In addition, measurements that previously would require manual tracing and measurement can now be rapidly acquired through AI software. However, the accuracy of these AI software packages on real-world patients is often limited. This study enrolled a cohort of 209 adult ED patients with risk factors for heart failure, but no previously known heart failure or valvular dysfunction. Four ultrasound-fellowship trained ED physicians acquired cardiac images using the Mindray TE X machine. Visual assessment of ejection fraction (EF) and manual measurements of diastolic function were performed. This was followed by AI software to auto-measure EF and diastolic doppler variables. Images were then reviewed by two ultrasound-fellowship trained experts for final determination of manual assessments. In assessing systolic dysfunction, AI was 86% (95%CI 57%-98%) sensitive and 95% specific (95%CI 91%-98%). For diastolic dysfunction, AI was 92% sensitive (95%CI 86%-96%) and 94% (95%CI 87%-98%).
How does this change my practice?
It is time to be optimistic about many of the AI software offerings on modern POCUS machines. Some are targeted at the novice user for image acquisition and labeling purposes, while others can efficiently perform quantitative measurements and save substantial time for any user. The current study adds evidence that this particular Mindray AI software performs well, though the sensitivity value for EF and its wide confidence interval give some pause. The auto-EF features also likely hold the most benefit for novice users who are not adept at visual estimation. A future study that uses novice users as the “test pilots” for this AI software could be very additive, as that is the population that theoretically stands to benefit the most.
Pro clips and tips
I’ve personally gotten used to the auto-EF and auto-VTI features on cardiac assessments given the time saving potential (auto-VTI) and as a backup to visual estimation (auto-EF). While these seem to function well, I will admit that I do carefully watch the 2D images for clarity and gestalt while the AI software does its work. The AI software is much more accurate when being fed high-quality images where it can clearly detect the anatomical structures (e.g. the endocardial border for EF assessment). Use of these features on high-quality imaging data sets will likely result in accurate values and determinations. However, we should remain wary of AI results that seem discordant with the eye of an experienced physician sonographer.


Source
Diagnostic accuracy of artificial intelligence for identifying systolic and diastolic cardiac dysfunction in the emergency department. Am J Emerg Med. 2024 Dec;86:115-119. doi: 10.1016/j.ajem.2024.10.019. Epub 2024 Oct 15. PMID: 39426020
