Clinical service evaluation of the feasibility and reproducibility of novel artificial intelligence based-echocardiographic quantification of global longitudinal strain and left ventricular ejection fraction in trastuzumab-treated patients

Introduction: Cardiotoxicity is a potential prognostically important complication of certain chemotherapeutic agents that may result in preclinical or overt clinical heart failure. In some cases, chemotherapy must be withheld when left ventricular (LV) systolic function becomes significantly impaired, to protect cardiac function at the expense of a change in the oncological treatment plan, leading to associated changes in oncological prognosis. Accordingly, patients receiving potentially cardiotoxic chemotherapy undergo routine surveillance before, during and following completion of therapy, usually with transthoracic echocardiography (TTE). Recent advancements in AI-based cardiac imaging reveal areas of promise but key challenges remain. There are ongoing questions as to whether the ability of AI to detect subtle changes in individual patients is at a level equivalent to manual analysis. This raises the question as to whether AI-based left ventricular strain analysis could provide a potential solution to left ventricular systolic function analysis in a manner equivocal to or superior to conventional assessment, in a real-world clinical service. AI based automated analyses may represent a potential solution for addressing the pressure of increasing echocardiographic demands within limited service-capacity healthcare systems, in addition to facilitating more accurate diagnoses.

Methods: This clinical service evaluation aims to establish whether AI-automated analysis compared to conventional methods (1) is a feasible method for assessing LV-GLS and LVEF, (2) yields moderate to good correlation between the two approaches, and (3) would lead to different clinical recommendations with serial surveillance in a real-world clinical population.

Results and Discussion: We observed a moderate correlation (r = 0.541) in GLS between AI automated assessment compared to conventional methods. The LVEF quantification between methods demonstrated a strong correlation (r = 0.895). AI-generated GLS and LVEF values compared reasonably well with conventional methods, demonstrating a similar temporal pattern throughout echocardiographic surveillance. The apical-three chamber view demonstrated the lowest correlation (r = 0.423) and revealed to be least successful for acquisition of GLS and LVEF. Compared to conventional methodology, AI-automated analysis has a significantly lower feasibility rate, demonstrating a success rate of 14% (GLS) and 51% (LVEF).