# Artificial Intelligence Improves Apical Four-Chamber Window Quality in Experienced but Not Novice Users

## TL;DR

AI helps experienced users take better heart ultrasound images but slows them down and doesn't help novices much.

## Contribution

Shows AI improves image quality for experienced users but not novices, affecting training strategies.

## Key findings

- Experienced users had better image quality with AI assistance, especially for imaging plane and structure visibility.
- AI increased acquisition time for novices and showed no improvement in image quality for them.
- Experienced users outperformed novices in all image quality criteria regardless of AI use.

## Abstract

Introduction: The most current point-of-care ultrasound (POCUS) machines incorporate artificial intelligence (AI) features to assist users in obtaining cardiac windows. Prior studies have focused on evaluating the helpfulness of AI acquisition assistance for novice users. This study sought to determine the immediate effect of AI assistance on acquisition time and image quality of apical four-chamber (A4C) cardiac ultrasound windows obtained by both novice and experienced users.

Methods: Fourteen novices with limited POCUS training during medical school and 10 experienced second- and third-year emergency medicine residents recorded A4C windows with and without AI assistance on three standardized patients in randomized order. Acquisition time in seconds was compared with the Mann-Whitney U test. Chi-square analysis was used to compare the proportion of recordings demonstrating each of the three image quality criteria: all essential structures visible, correct image plane, and proper probe placement.

Results: The median (interquartile range) acquisition time was longer with AI assistance in both user groups, though the difference was only significant for the novice users (136 (109) seconds; 75 (66); p < .01), not for the experienced users (98 (132); 66 (47); p = .18). All A4C quality criteria were significantly more likely by experienced than novice users. For experienced users, the visibility of all essential structures trended more likely with AI assistance (0.87 (0.70, 0.95); 0.67 (0.49, 0.81); p = .06) and the correct imaging plane was significantly more likely with AI assistance (0.60 (0.42, 0.75); 0.37 (0.22, 0.55); p = .03). There was no difference in proper probe placement with AI assistance for experienced users. No significant differences in image quality criteria were observed in the novice user subgroup.

Conclusion: The results provide strong evidence that the immediate effect of AI assistance was associated with longer acquisition times in novice users obtaining A4C windows, with a similar trend among experienced users. AI assistance was also associated with a higher proportion of quality criteria amongst the experienced users, but not the novice users. Therefore, medical educators should consider the experience level of POCUS learners when incorporating AI acquisition features into education and clinical practice.