# Validation and longitudinal trajectory analysis of an AI-based ECG model for aortic stenosis: from community screening to pre-TAVR risk stratification

## TL;DR

An AI-based ECG model can detect early signs of aortic stenosis and predict future risk, even years before treatment.

## Contribution

The study validates an AI model for aortic stenosis detection and shows its ability to predict disease progression and mortality risk.

## Key findings

- The AK-AVS model achieved 79% AUROC in detecting moderate/severe aortic stenosis in community screening.
- AI-ECG trajectory patterns predicted mortality risk up to 4.5 years before TAVR, outperforming traditional risk scores.
- False-positive predictions identified individuals with higher future risk of AS hospitalization and heart failure.

## Abstract

Early aortic stenosis (AS) detection remains challenging, with many patients presenting late when left ventricular dysfunction may be irreversible. We evaluated whether longitudinal AI-enhanced ECG patterns can predict outcomes years before intervention and assessed the community screening potential of the AK-AVS model.

We conducted two complementary analyses: (1) community validation of the AK-AVS model in 3632 cardiovascular disease-free ARIC participants, and (2) longitudinal trajectory analysis of 7860 ECGs from 2040 TAVR recipients collected up to 10 years pre-procedure. Unsupervised clustering identified distinct AK-AVS trajectories, with mortality associations assessed using Cox regression and net reclassification improvement. In community screening (n = 16 moderate/severe AS), AK-AVS achieved an AUROC of 0.79, sensitivity 75%, and specificity 75% for moderate/severe AS. At hypothetical screening prevalences of 1–5%, positive predictive values improved to 3.1–14.3%. False-positive predictions identified individuals at 4-fold increased risk for future AS hospitalisation (HR 4.05, P < 0.001) and 52% increased risk for heart failure (HR 1.52, P = 0.02). In the TAVR cohort, trajectory analysis revealed three distinct patterns: Stable Low (19.3%), Accelerated Progression (23.6%), and Persistently High (57.1%). Elevated trajectory groups were older (78.4 and 77.8 vs. 72.6 years, P < 0.001) with higher pacemaker rates (16.4% and 17.3% vs. 10.7%, P = 0.008), despite similar hemodynamic severity. Both elevated patterns independently predicted mortality (Accelerated: HR 1.40, P = 0.03; Persistently High: HR 1.48, P = 0.005) and significantly improved risk reclassification beyond traditional risk scores (NRI 0.069–0.074).

Longitudinal AI-ECG trajectory patterns detect disease progression up to 4.5 years before TAVR and enhance mortality prediction beyond traditional risk scores. Community validation shows potential screening utility with ‘false-positives’ identifying future risk.

Graphical Abstract