Head-to-head comparison of AI-based global longitudinal strain measurement - are we ready for this next step?
R Botan, A Winkler, K Sveric, A LinkeAbstract
Background
Global longitudinal strain (GLS) is an established prognostic marker in echocardiography. Harmonisation efforts over the past decade have improved consistency among conventional speckle-tracking software. Recently, artificial intelligence (AI) has enabled fully automated GLS analysis without operator input, demonstrating good agreement with conventional methods and supporting broader clinical implementation.
Purpose
We aimed to evaluate GLS measurements obtained from two independent AI-based software solutions against the conventional manual method as the reference and to determine whether these AI-based measurements are interchangeable (Figure 1A).
Methods
AI-based GLS analysis was performed fully automatically without user interaction, including automated selection of LV apical 2-, 3-, and 4-chamber views, detection of the LV endocardial borders, and longitudinal strain calculation as predefined by the manufacturers (AI#1; AI#2). As the reference standard, conventional manual GLS analysis was performed using independent analysis software by expert readers blinded to the AI results. We enrolled 79 consecutive patients (mean age 63 ± 13 years; 47% male) who underwent clinically indicated transthoracic echocardiography at our tertiary care centre. Image acquisition was performed in accordance with current guideline recommendations. Agreement between manual and AI-based GLS measurements, as well as between the two AI software solutions, was assessed using intraclass correlation coefficients (ICC), mean absolute difference (MAD), and Bland-Altman analysis, including bias and limits of agreement (LoA).
Results
Both AI#1 and AI#2 showed good agreement with manual GLS measurements across the full measurement range (Figure 1B). Agreement was slightly higher for AI#1 (ICC = 0.81, MAD = 1.6%) compared with AI#2 (ICC = 0.78, MAD = 1.9%). Bland-Altman analysis revealed low bias for AI#1 relative to manual GLS (bias 0.7%), with narrow LoA raging from -3.7% to 5.0% (Figure 1C). Similarly, AI#2 showed small bias compared with manual GLS (bias -0.5%), though with slightly wider LoA ranging from -5.4% to 4.3% (Figure 1D). Direct comparison between AI#1 and AI#2 demonstrated good inter-method agreement (ICC = 0.78, MAD = 1.9%), with a mean bias of -1.2% and LoA from -3.2% to 5.5%. Fully automated AI-based GLS analysis was feasible in 95% of patients (n = 75).
Conclusion
AI-based GLS analysis showed good agreement with the manual reference and good inter-method consistency, supporting the reliability of fully automated GLS quantification. However, inter-algorithm differences were observed. Further studies are warranted to determine whether harmonisation of GLS measurements across AI solutions is required to ensure consistent clinical interpretation.Figure 1