A novel AI-derived digital biomarker for monitoring disease progression in ATTR-CM: first-in-trial use of a computer vision AI-ECG algorithm within a phase 3 pivotal randomized controlled trial
R Khera, P M Croon, S S Vasisht, T Bergemann, G Murtagh, J Whang, L DhingraAbstract
Background
Transthyretin amyloid cardiomyopathy (ATTR-CM) is a progressive infiltrative cardiomyopathy. Artificial intelligence electrocardiogram (AI-ECG) algorithms have emerged as powerful tools to screen for cardiomyopathies such as ATTR-CM. However, the potential of AI-ECG prediction scores to function as a quantitative digital biomarker for longitudinal disease monitoring has not been established. In the phase 3 ATTRibute-CM randomized controlled trial (RCT; NCT03860935), acoramidis, a near-complete (≥90%) transthyretin stabilizer, reduced cardiovascular mortality and cardiovascular-related hospitalizations. The trial obtained serial ECGs of enrolled participants at baseline and follow up.
Purpose
We report the first deployment of a computer vision AI-ECG algorithm, operating directly on ECG images without digital signal data, as a digital biomarker to prospectively monitor ATTR-CM disease progression in a RCT (ATTRibute-CM).
Methods
Eligible adult participants (≤90 years) with an established diagnosis of ATTR-CM based on either an endomyocardial biopsy or positive results on technetium-99m scintigraphy were enrolled in the trial. Standard 12-lead ECGs were obtained on all participants every 3 months during the study. The AI-ECG computer vision algorithm was deployed in a blinded fashion to all ECGs with intrinsic rhythms after they were fully anonymized and randomly organized. The output from the AI-ECG is a prediction score from 0 to 1. After re-linking participant identifiers and visit timing, the mean baseline AI-ECG score and the subsequent change from baseline (CFB) at month 30 (M30) were summarized. The difference in the M30 CFB AI-ECG scores between acoramidis and placebo were compared using a 2-sample t-test.
Results
A total of 516 baseline ECGs were analyzed. Among these, 199 acoramidis-treated and 91 placebo-treated participants had both baseline and Month 30 (M30) ECGs available for longitudinal analysis. Across study arms, baseline AI-ECG scores were higher in those with variant/mutant ATTR-CM vs wild-type (0.43 vs 0.35) and those with elevated NT-proBNP >3000 pg/mL vs <3000 pg/mL (0.39 vs 0.34) (Table). Mean (SD) baseline AI-ECG scores were 0.38 (0.21) and 0.31 (0.23) in the acoramidis and placebo arms, respectively (p-value = ns). Over 30 months of follow-up, there was a CFB in the AI-ECG score of -0.03 in the acoramidis arm versus +0.02 in the placebo arm (mean difference 0.05, p<0.05) (Figure).
Conclusion
An image-based AI-ECG algorithm demonstrated discrimination across clinical subgroups at baseline and detected differential longitudinal changes between acoramidis and placebo over 30 months. These findings support the potential role of AI-ECG-derived prediction scores as a scalable digital biomarker for disease monitoring in clinical trials and potential routine cardiovascular care.