Enhancing clinical care in obstructive hypertrophic cardiomyopathy: the use of artificial intelligence for echocardiographic follow-up during treatment with mavacamten
A Van Hoogdalem, P P Zwetsloot, A Schinkel, R van der Boon, D Bowen, R de Boer, A Hirsch, M MichelsAbstract
Background
Since the introduction of mavacamten for treatment of obstructive hypertrophic cardiomyopathy(oHCM), workload on cardiomyopathy clinics has increased due to the necessity of frequent follow-up, including echocardiograms. Artificial intelligence (AI-)enabled echocardiography analysis software could relieve part of this burden by automated echo assessments, of which first versions are approved for clinical use.
Purpose
Assess whether AI can accurately be used in patients with oHCM during treatment with mavacamten.
Methods
Symptomatic oHCM patients treated with mavacamten for at least 6 months were included between January 2024 and October 2025. All patients provided written informed consent. Echocardiograms at baseline, 12 weeks and if available at 1 year of follow-up were analyzed by US2.ai (US2.v2(.3.0)) and results were compared to assessment by an experienced cardiologist. Parameters of interest were left ventricular ejection fraction (LVEF), LV end diastolic volume (LVEDV), LV end systolic volume (LVESV), intraventricular septum (IVS), posterior wall thickness (PWT), LV end diastolic dimension (LVEDD) and global longitudinal strain (GLS). LV mass was calculated using the Devereux formula, based on IVS, LVEDD and PWT. LV outflow tract (LVOT) gradient was automatically measured by AI with the maximal gradient algorithm for the aortic valve, an undesignated function of this software. LAVi and right ventricular (RV) parameters were only measured with AI. Analyses include Wilcoxon rank tests, linear mixed models and type III ANOVA.
Results
A total of 70 oHCM patients treated with mavacamten were included of whom 49 reached 1 year of follow-up (Table 1 for baseline characteristics). AI measurements of LVEF, PWT and LVEDD were comparable to human measurements (p>0.05). However, there were significant differences between human and AI in LVEDV, LVESV and IVS. For these measurements, the AI-derived values were lower. For GLS, AI measurement was significantly higher. The LVOT gradient measurement with AI was significantly lower than the values measured by humans (median=29mmHg [IQR 13,43] vs61mmHg [IQR 20,96] respectively, p<0.001) (Figure 1). During follow-up, both methods showed a significant decrease in IVS, LV mass and LVOT gradient. AI measurements were significantly lower for all three parameters (p<0.001). In addition, US2.ai measured LAVi (41ml/m2 at baseline vs 31.8ml/m2 at 1 year, p=0.002) and RV parameters (RVEDV:40.7ml vs 43.9ml, p=0.001; RVESV:14ml vs 20.8ml, p<0.001) that were not reported in focused echocardiography reports.
Conclusions
US2.ai measured LVEF accurately compared to humans, supporting its potential applicability in clinical mavacamten practice providing unbiased automated measurements and adjunct parameters (e.g. LAVi, RV function). However, the current approved AI software is not suitable for the assessment of the LVOT gradient and dedicated oHCM packages are needed.Figure 1 - Paired analyses per parameterTable 1 - Baseline characteristics