Preoperative AI-ECG analysis predicts paravalvular leak after transcatheter aortic valve replacement

Background Paravalvular leak (PVL) following transcatheter aortic valve replacement (TAVR) remains associated with adverse outcomes and increased mortality. Current risk stratification relies primarily on anatomic features from cardiac CT, which incompletely capture the complex structural and functional remodelling influencing valve seating. The AccurKardia Aortic Valve Stenosis (AK-AVS) score is an artificial intelligence-enabled electrocardiographic (AI-ECG) biomarker validated to detect moderate-to-severe aortic stenosis by integrating electrical consequences of structural changes including left ventricular hypertrophy, fibrosis, and conduction abnormalities. Purpose To determine whether preoperative AK-AVS scores independently predict PVL. Methods We analysed 1,456 patients who underwent TAVR at our institution (2013-2024) after excluding patients with permanent pacemakers or missing preoperative AK-AVS scores. The AK-AVS score was derived from standard 12-lead ECGs obtained within 30 days prior to TAVR. PVL was graded according to Valve Academic Research Consortium-3 criteria as none, mild, or moderate-to-severe. Multivariable logistic regression models were adjusted for AK-AVS score (scaled per 0.10 increase) and established predictors: mean gradient, aortic valve area, aortic annulus diameter, valve size, aortic valve calcium score, device type (balloon-expanding [BEV] vs. self-expanding [SEV]), and STS risk score. Receiver operating characteristic curves were generated and area under the curve (AUC) values were compared using DeLong's test. Results Of 1,456 patients, 232 (15.9%) developed PVL. Patients with PVL had higher preoperative AK-AVS scores (0.71 vs. 0.67, p=0.011), lower ejection fraction (51.6% vs. 54.5%, p<0.001), larger valve size (27.3 vs. 26.2mm, p<0.001), higher calcium burden (2,743 vs. 2,413 Agatston units, p=0.001), and higher STS scores (6.4% vs. 5.4%, p=0.003) (Figure 1). Each 0.10 increase in AK-AVS score independently predicted PVL (OR 1.10, 95% CI 1.03-1.22, p=0.028) after adjusting for other predictors. Addition of AK-AVS to the base clinical model improved discrimination, increasing AUC from 0.643 (95% CI 0.609-0.677) to 0.686 (95% CI 0.653-0.719), p=0.007 (Figure 2A). In ordinal regression examining PVL severity, each 0.10 increase in AK-AVS score was associated with 12% higher odds of more severe PVL (OR 1.12, 95% CI 1.03-1.22, p=0.01). A clear dose-response relationship was evident: mean AK-AVS scores were 0.67 (SD 0.20) for no PVL, 0.71 (SD 0.19) for mild PVL, and 0.78 (SD 0.18) for moderate-to-severe PVL (Figure 2B). Conclusion Preoperative AK-AVS scores independently predict PVL development and severity following TAVR, providing incremental value beyond traditional clinical, echocardiographic, and procedural risk factors. This universally available AI-ECG biomarker may aid preprocedural risk stratification and guide valve selection or procedural planning to minimize PVL risk.