Artificial intelligence‐assisted echocardiography: contextualising the promise

We read with interest the multicentre study by Borde et al. that showed good-to-excellent agreement between US2.AI software and expert clinicians across most echocardiographic parameters in the peri-operative setting [1]. However, we wish to raise two interrelated considerations that may affect the generalisability of these findings. First, of the 202 patients enrolled, 102 (51%) were undergoing cardiac surgery [1]. The study's conclusion that anaesthetists can obtain actionable cardiac assessments using artificial intelligence-assisted transthoracic echocardiography may not be directly transferable to the non-cardiac surgical setting, where the need for point-of-care echocardiography is arguably greatest. In cardiac surgery, patients are managed by cardiac anaesthesia teams with advanced echocardiography expertise and immediate transoesophageal echocardiography backup. By contrast, the ‘real-world’ challenge lies in the non-cardiac operating theatre: a general anaesthetist, often without on-site support, facing a high-risk patient under time pressure. The promise of artificial intelligence – compensating for limited operator experience – is most relevant in this non-cardiac context, yet nearly half the validation cohort came from cardiac surgical cases. We wonder whether the authors have considered a subgroup analysis restricted to non-cardiac surgery patients undergoing non-cardiac surgery. Second, the study reported only moderate agreement for two haemodynamic parameters: inferior vena cava collapsibility (r = 0.641) and cardiac output (r = 0.675) [1]. These values are lower than the strong correlations observed for left ventricular ejection fraction (r = 0.845) and right ventricular size (r = 0.860). We believe this disparity is not incidental. Inferior vena cava collapsibility depends on accurate identification of respiratory phase; this is a dynamic parameter that artificial intelligence algorithms may struggle to capture reliably as they typically analyse fixed time windows [2]. Furthermore, the clinical utility of inferior vena cava metrics in patients who are breathing spontaneously is limited by substantial inter-individual variability [3]. Cardiac output calculation, derived from left ventricular outflow tract diameter squared multiplied by velocity-time integral, is inherently sensitive to measurement error: a small error in diameter is amplified by squaring [4]. In peri-operative practice, where acoustic windows are often suboptimal, this sensitivity may be magnified further. These two considerations are linked: if artificial intelligence performance on dynamic haemodynamic parameters is only moderate even in the cardiac surgical cohort – where operators are highly experienced and imaging conditions optimal – then performance in the non-cardiac setting, where operators may be less experienced and conditions more challenging, could be even less reliable. This suggests that clinical deployment of artificial intelligence-assisted echocardiography should be parameter-specific, with human interpretation reserved where artificial intelligence performance is weaker. We agree that further studies are needed to assess the impact of artificial intelligence-assisted echocardiography on clinical outcomes [1]. We would encourage validation in non-cardiac surgical populations; operator-stratified analyses (experts vs. novices) to assess whether artificial intelligence compensates for inexperience; and algorithm refinement for dynamic parameters (inferior vena cava collapsibility, cardiac output), perhaps by integrating respiratory phase detection.