Key result
ECGFounder outperforms comparator models in ECG diagnosis, achieving a ~0.93 AUC for subclass classification.
Why the study?
Does the choice of foundation model architecture and fine-tuning strategy improve diagnostic performance and physiologic saliency alignment in ECG analysis?
Does the choice of foundation model architecture and fine-tuning strategy improve diagnostic performance and physiologic saliency alignment in ECG analysis?
Domain-aligned ECG foundation models provide stable and efficient diagnostic performance and physiologic saliency alignment compared to general time-series models.
Answers are grounded in this paper and cite their sources.
Supports domain-specific ECG models for AI pipelines; leaves open prospective clinical validation and cross-dataset generalizability.
Foundation models have recently emerged as powerful tools for medical artificial intelligence, but the optimal architecture and adaptation strategy for electrocardiogram (ECG) analysis remain unclear. In ECG-based clinical AI, predictive performance alone may be insufficient, because models should also focus on physiologically meaningful waveform regions. We aimed to present an evaluation framework for selecting and adapting signal-based foundation models and for quantitatively assessing physiologic saliency alignment. We evaluated three pretrained foundation models with distinct pretraining domains and architectures: ECGFounder, an ECG-specific convolutional neural network; MOMENT, a general-purpose time-series Transformer; and PaPaGei, a photoplethysmography-specific convolutional neural network. Models were adapted to the PTB-XL dataset using multiple fine-tuning strategies, including linear probing, partial fine-tuning, low-rank adaptation, and convolutional low-rank adaptation. Diagnostic performance was assessed using weighted area under the receiver operating characteristic curve across diagnostic subclass, form, and rhythm categories, along with a separate atrial fibrillation classification task. To evaluate physiologic saliency alignment, we developed a Relative Saliency Score (RSS) framework that quantifies saliency density within clinically relevant ECG segments using automated physiologic segmentation. ECGFounder demonstrated the most robust overall performance, achieving high diagnostic performance even with linear probing, with weighted AUCs of 0.932 for subclass classification and 0.983 for atrial fibrillation. MOMENT showed lower baseline performance but improved substantially with low-rank adaptation, reaching a subclass AUC of 0.909 and atrial fibrillation AUC of 0.947. Quantitative saliency analysis revealed distinct architecture-dependent saliency patterns. ECGFounder showed consistently elevated RSS values in relevant physiologic regions, whereas MOMENT showed greater saliency concentration in dominant waveform components, particularly the QRS complex, after low-rank adaptation. PaPaGei showed moderate cross-modal transferability, with generally weaker diagnostic performance and less consistent physiologic saliency alignment. This study presents an evaluation framework for selecting and adapting signal-based foundation models for ECG analysis. Successful adaptation of foundation models for ECG analysis depends on the interaction between model architecture, pretraining domain, and fine-tuning strategy. Domain-aligned ECG foundation models provide stable and efficient performance, whereas general time-series Transformers may benefit substantially from parameter-efficient adaptation. Quantitative assessment of physiologic saliency alignment offers complementary information beyond conventional performance metrics and may support more comprehensive characterization of foundation models.
No takes yet. Share an insight, caveat, or question.
HAN et al. (2026) studied Electrocardiogram (ECG) analysis (n=18,885). ECGFounder vs. MOMENT and PaPaGei was evaluated on Weighted area under the receiver operating characteristic curve (AUC) for diagnostic subclass classification. ECGFounder demonstrated the most robust overall diagnostic performance for ECG analysis, achieving a weighted AUC of 0.932 for subclass classification and 0.983 for atrial fibrillation with linear probing.
Synapse has enriched 10 related papers. Consider them for comparative context: