LAEF: A Lead-Agnostic ECG Foundation Model Towards Point-of-Care Diagnostics

Existing ECG foundation models are architecturally rigid, requiring a fixed 12-lead input. This creates a significant performance gap when these models are applied to point-of-care devices (like smartwatches) that typically capture only 1 or 2 leads. The authors ask whether an ECG foundation model can be designed to be natively lead-agnostic, allowing it to maintain high diagnostic accuracy regardless of the number of leads available at inference time.

To solve this, the authors introduce LAEF (Lead-Agnostic ECG Foundation), a 7M-parameter model that treats ECG signals as variable-size spatiotemporal graphs. Instead of relying on fixed-length tensors, LAEF represents ECG leads as nodes in a graph, with edges encoding both intra-lead temporal dependencies and inter-lead physiological synchrony. The model is pre-trained on 9.2 million 12-lead ECGs using a masked node modeling objective combined with stochastic lead sampling. This training strategy forces the model to learn robust representations of cardiac dynamics from arbitrary lead subsets, effectively treating lead absence as a structural property of the input rather than a missing data problem requiring zero-padding.

LAEF demonstrates that lead-agnosticism does not come at the cost of clinical performance. When tested on 18 downstream datasets, LAEF performs on par with specialized 12-lead models at full lead availability. More importantly, in reduced-lead settings, LAEF significantly outperforms zero-padded baselines, achieving higher AUROC scores on 17 out of 18 datasets with a single lead and 14 out of 18 with two leads. The authors' analysis suggests that this performance gain stems from the model's architectural flexibility, which avoids the input corruption caused by padding missing leads with zeros.

This work bridges the gap between clinical-grade diagnostic models and the reality of modern, accessible cardiac monitoring. By enabling a single model to function effectively across both high-fidelity clinical environments and resource-constrained point-of-care settings, LAEF provides a scalable path toward more reliable, ubiquitous cardiovascular diagnostics.