Stability and Individuality of ECG Foundation Model Embeddings in a Longitudinal Case Study

Portable ECG devices enable frequent, real-world cardiac monitoring, yet the longitudinal behavior of modern ECG foundation model representations derived from such data remains poorly characterized. In this work, we present a 20-day longitudinal case study examining the temporal stability and individual specificity of ECG representations obtained from a pretrained foundation model applied to consumer-grade six-lead ECG recordings. Daily resting ECGs were collected from two healthy adults using a portable device. For each recording, both latent embedding vectors and task-level probability outputs produced by the model were analyzed. Principal component analysis revealed clear subject-specific clustering in both representation spaces. Temporal drift analysis demonstrated that intra-subject variability remained consistently smaller than inter-subject separation over time. A nearest-centroid distance-margin analysis further showed robust subject-specific separability without classifier training. Together, these results indicate that ECG foundation model representations derived from portable recordings are stable over time and encode persistent individual characteristics, supporting their potential utility for longitudinal and personalized ECG analysis.