Answers are grounded in this paper and cite their sources.
Real-time artificial intelligence (AI) probe-guidance software may allow operators without formal ultrasound training to acquire echocardiograms, a form of task-shifting for cardiovascular screening. The proportion of such examinations in real patients that reach diagnostic quality has not been quantitatively synthesized. Two reviewers independently searched PubMed/MEDLINE, Google Scholar, Embase, Scopus, and ClinicalTrials.gov (final search 27 September 2026), supplemented by citation-following, for studies reporting the proportion of AI-guided, novice- or nonexpert-acquired echocardiographic examinations in real patients meeting a predefined diagnostic-quality threshold. Proportions were pooled with a binomial–normal generalized linear mixed model (GLMM; principal analysis), and robustness was examined with a Freeman–Tukey/DerSimonian–Laird random-effects model and further sensitivity analyses. Post-hoc subgroup and leave-one-out analyses were performed, and risk of bias was appraised using an adapted JBI prevalence checklist. Eight studies (2365 examination-, scan-, or image-level observations; 74 operators) were pooled. The overall diagnostic-quality yield was 96.3% (95% CI 85.5–99.3%) in the GLMM and 93.0% (95% CI 82.8–99.0%) in the Freeman–Tukey model, and sensitivity analyses gave consistent estimates (91.2–97.8%); between-study heterogeneity was considerable. Six controlled/structured-deployment studies pooled to 98.0% (95% CI 96.7–99.2%), whereas two real-world, decentralized community studies pooled to 63.7% (95% CI 20.6–92.6%), with individual yields of 44.8% and 79.8% (GLMM meta-regression p < 0.001). The two real-world studies also differed from the others in supervision, training, operators, and patient selection. Under controlled conditions, novice and nonexpert operators using AI guidance obtained diagnostic-quality echocardiographic images in nearly all real patients, across five AI systems and multiple operator types. Decentralized community deployments achieved lower and more variable yields, pointing to operator training, supervision, and quality assurance as practical priorities for scale-up. These findings support the feasibility of AI-guided acquisition by novices, particularly under structured deployment conditions, and provide a benchmark for future comparative studies of diagnostic accuracy and clinical impact.
No takes yet. Share an insight, caveat, or question.
Bastola et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: