Artificial intelligence–assisted point-of-care ultrasound training in novice learners: a systematic review and quality appraisal of controlled trials

Point-of-care ultrasound (POCUS) is an essential skill for modern medical practice. Its operator-dependent nature presents two structurally distinct training challenges: the high burden of supervised hands-on practice required to achieve competency, and a critical shortage of credentialed POCUS instructors. Artificial intelligence (AI) has emerged as a potential solution to both barriers, but its impact on the novice learning curve has not been systematically evaluated. A systematic search of PubMed, Embase, and the Cochrane Central Register of Controlled Trials (CENTRAL), supplemented by ClinicalTrials.gov and the WHO International Clinical Trials Registry Platform (ICTRP), was conducted for studies published between 1 January 2020 and 20 May 2026. The review was prospectively registered in PROSPERO (CRD42025267700) prior to data extraction. Risk of bias was assessed using RoB 2 and ROBINS-I; certainty of evidence was graded using the GRADE framework. From 6,871 identified records, 28 full-text articles were assessed for eligibility and six studies (4 randomised controlled trials (RCTs), 2 non-randomised studies; N = 293 participants) met inclusion criteria. In the single low-risk-of-bias trial, a non-inferiority RCT restricted to focused cardiac ultrasound, AI-assisted training met the pre-specified 5-point RACE margin (AI 13.39 [SD 2.79] vs. Traditional 15.77 [SD 1.89]; mean difference [Traditional − AI] 2.38 points; upper limit of the one-sided 95% CI 3.60; p < 0.001), with the traditional group numerically higher at every time point. A second RCT, in which AI was withdrawn before assessment, reported faster apical 4-chamber acquisition (57 s vs. 85 s; p = 0.01) and higher image quality (modified RACE 4.5 vs. 2; p < 0.01). Benefit concentrated in the technically demanding apical views. In the non-inferiority trial, significantly fewer AI-trained participants produced images adequate to permit interpretation of left and right ventricular function, pericardial effusion and volume status (50.0% vs. 81.8%, p = 0.026) despite non-inferior acquisition scores. One trial assessed interpretation, using idealised images rather than participants’ own scans; none assessed clinical decision-making. A high-risk-of-bias RCT (Aronovitz 2024) reported AI superiority, but this finding is substantially confounded by 40% post-allocation exclusions and is presented as exploratory only. Skill retention at three months was comparable between groups. No study reported cost-effectiveness data, and no study adequately assessed clinical decision-making competency. GRADE certainty was low for overall skill acquisition and very low for diagnostic consistency. AI-assisted training matches or improves upon conventional instruction for cardiac image acquisition by novices, with benefit concentrated in technically demanding views. The evidence is confined to cardiac applications assessed on healthy or standardised subjects. Learners’ interpretation of their own images, clinical decision-making and cost-effectiveness remain unevidenced, and trials differ in whether AI remained available at assessment, and therefore estimate different constructs. Until higher-certainty evidence is available, AI-assisted training should be regarded as a complement to, rather than a replacement for, expert-supervised instruction within competency-based POCUS curricula. CRD42025267700 (registered prior to data extraction; no protocol amendments, confirmed by comparison of the final manuscript with the PROSPERO registration record).