Key result
The compact ECGFM-U-Lite model achieved an overall AUC of 0.862 compared to 0.918 for the teacher model, while decreasing the parameter count from 86.4 M to 655.7 K.
Why the study?
Recent ECG foundation models extend analysis to cardiac structure-function but have high computational and memory demands limiting use in resource-constrained settings.
Does a lightweight ECG foundation model (ECGFM-U-Lite) reduce computational and memory requirements while preserving diagnostic performance in cardiac disease screening?
Does a lightweight ECG foundation model (ECGFM-U-Lite) reduce computational and memory requirements while preserving diagnostic performance in cardiac disease screening?
Absolute Event Rate: 0.862% vs 0.918%
The ECGFM-U-Lite model significantly reduces the computational and memory footprint of ECG foundation models while preserving high diagnostic accuracy, enabling practical deployment on resource-constrained hardware.
Lite ECG foundation models may enable deployment in resource-limited settings; extends prior work but leaves open prospective clinical validation.
Recent electrocardiogram (ECG) foundation models have extended ECG analysis from rhythm and morphology recognition to cardiac structure-function assessment, but the computational and memory requirements limit their use in resource-constrained settings. This study aimed to reduce the size and inference cost of a unified ECG foundation model while retaining its multitask performance and clinically relevant representations. We developed the ECGFM-U-Lite, a compact 1D-MobileViT student, using a three-stage compression procedure. Task outputs and intermediate representations were firstly transfered from teacher through structure-function consistent distillation, redundant channels were than removed by using physiologic attention-guided pruning while retaining anatomically informative pathways, and the structure-function-aware quantization strategy achieves adaptive precise quantification by leveraging the sensitivity discrepancies across different modules finally. The model was evaluated on six locked external cohorts. Its computational profile was examined on a laptop GPU and under Android- and Jetson-Nano-oriented ONNX Runtime simulations. ECGFM-U-Lite achieved an overall AUC of 0.862, compared with 0.918 for the teacher, and retained 93.8% and 93.5% of teacher AUC for rhythm/morphology and structure/function tasks, respectively. ECG-to-Text and Text-to-ECG R@1 values were 0.766 and 0.762 in the teacher-derived semantic consistency probe. The parameter count decreased from 86.4 M to 655.7 K, model size from 345.6 MB to 1.7 MB, and peak RAM from 320 MB to 4.4 MB. The inference process was 3.6-fold faster on the laptop GPU and approximately 3.7-fold faster in the simulated edge-oriented profiles. U-Lite substantially reduces the resource requirements of a unified ECG foundation model while preserving most of its external discrimination and teacher-derived cross-modal representation structure, providing a practical basis for subsequent evaluation on resource-constrained hardware.
No takes yet. Share an insight, caveat, or question.
Ma et al. (2026) studied Cardiac disease. ECGFM-U-Lite (compact 1D-MobileViT student model) vs. Teacher ECG foundation model was evaluated on Overall AUC across tasks. The compact ECGFM-U-Lite model achieved an overall AUC of 0.862 compared to 0.918 for the teacher model, while decreasing the parameter count from 86.4 M to 655.7 K.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: