Comparison of artificial intelligence and physician assessment of global longitudinal strain before and after surgical aortic valve replacement
Z Abramikas, G Jagminaite, G Tumalaviciute, G Samoskaite, D Matuliauskas, G Balciunaite, D Palionis, S Glaveckaite, N ValevicieneAbstract
Background
Global longitudinal strain (GLS) as assessed by echocardiography is a sensitive marker of left ventricular systolic function with established prognostic value in severe aortic stenosis. Artificial intelligence (AI)–based strain analysis may improve standardization and reproducibility.
Purpose
To evaluate agreement between physician- and AI-derived GLS measurements before and after surgical aortic valve replacement (AVR), and to assess their ability to detect functional recovery.
Methods
Sixty patients (66.3 ± 8.1 years; 58,3% female) with severe aortic stenosis undergoing surgical AVR were included. Pre- and 12-month post-operative transthoracic echocardiograms were assessed by physicians and by a CE-marked, commercially available AI-based strain analysis software. Agreement was evaluated using Pearson correlation, Bland–Altman analysis, and intraclass correlation coefficient (ICC(3,1); two-way mixed-effects model, single measurement, absolute agreement). Proportional bias was assessed by regression of inter-method differences on measurement means. Longitudinal changes were analysed using paired testing. Statistical analyses were performed in Python using SciPy and pingouin.
Results
Pre-operatively, AI-derived GLS showed strong correlation with physician measurements (r = 0.72–0.79). For averaged GLS, agreement was good (ICC(3,1) = 0.77; 95% CI 0.65–0.86), with a mean physician–AI difference of 2.71% and Bland–Altman limits of agreement of approximately −6.4% to 11.8% (Figure 1). No significant proportional bias was observed. Post-operatively, correlation for averaged GLS remained moderate (r = 0.67), and inter-method bias increased to 3.65%, with a significant increase in the 3-chamber view (p = 0.003, Figure 2). The magnitude of inter-method differences was comparable to ranges reported for inter-observer variability in GLS assessment. Both methods detected significant GLS improvement at 12 months (physician +2.63%, p < 0.001; AI +1.69%, p = 0.001).
Conclusions
AI-derived GLS demonstrates good agreement with physician assessment and reliably detects post-operative functional recovery after surgical AVR. Although a modest systematic offset was observed, the magnitude of inter-method differences was within ranges commonly reported for inter-observer variability in GLS assessment. These findings support the potential integration of AI-based strain analysis into clinical workflows.Figure 1Figure 2