The genetic basis of cardiovascular disease remains largely unresolved. Existing deep-learning applications to electrocardiograms (ECGs) focus on single phenotypes or isolated timepoints Radhakrishnan et al. (2023); Libiseller-Egger et al. (2022); Wang et al. (2023), introducing biases. Here, we present Delta ECG, a metric based on pairs of embeddings that captures patient-specific changes in cardiovascular state over time. We demonstrate that this measure reliably differentiates within-patient variation from population-level differences and aligns with genome-wide significant loci for cardiovascular disease (CVD) and its risk factors