Synapse
⌘+K
Synapse
PulseExploreJournal ClubResearchersJournals
Instagram
HomeJournal ClubExplore
September 18, 2026BMC Medical Informatics and Decision MakingOpen Access

Evaluation of diagnostic performance and quantitative physiologic saliency alignment of time-series foundation models for electrocardiogram

View Full Paper
Ask AI
Bookmark
Share

Population

Electrocardiograms from the PTB-XL dataset

Comparison

Pretrained foundation models adapted using… vs Comparison between different foundation models…

Design

Other

Key result

ECGFounder achieved higher diagnostic performance (subclass AUC 0.932, AF AUC 0.983) compared to the general-purpose MOMENT model (subclass AUC 0.909, AF AUC 0.947).

Authors

CHC HANHLHyeokjong LeeJKJaewon Kim

Discussion

Loading...

Member takes

Overview

Supports ECG-specific models for diagnostics; leaves open prospective validation versus adapted general-purpose alternatives.

Key Points

  • To establish an evaluation framework for selecting and adapting signal-based foundation models for electrocardiogram analysis, focusing on both diagnostic performance and quantitative physiologic saliency alignment.
  • Evaluated three pretrained foundation models across distinct architectures and pretraining domains: ECGFounder (ECG-specific CNN), MOMENT (general time-series Transformer), and PaPaGei (photoplethysmography CNN).
  • Adapted models to the PTB-XL dataset using linear probing, partial fine-tuning, low-rank adaptation, and convolutional low-rank adaptation.
  • Assessed classification performance across diagnostic subclass, form, rhythm, and atrial fibrillation using weighted AUC, alongside a novel Relative Saliency Score framework to quantify attention within clinical waveform segments.
  • ECGFounder demonstrated the highest baseline and overall performance via linear probing, achieving weighted AUCs of 0.932 for subclass classification and 0.983 for atrial fibrillation, while consistently focusing on clinically relevant waveform regions.
  • MOMENT substantially improved with low-rank adaptation, achieving a subclass AUC of 0.909 and atrial fibrillation AUC of 0.947, with saliency predominantly concentrated on the QRS complex.
  • PaPaGei exhibited weaker cross-modal diagnostic performance and less consistent physiologic saliency alignment relative to domain-specific architectures.

Structured PICO

P
Population
Electrocardiograms from the PTB-XL dataset
I
Intervention
Pretrained foundation models (ECGFounder, MOMENT, PaPaGei) adapted using multiple fine-tuning strategies (linear probing, partial fine-tuning, low-rank adaptation, and convolutional low-rank adaptation)
C
Comparator
Comparison between different foundation models and fine-tuning strategies
O
Outcome
Diagnostic performance (weighted area under the receiver operating characteristic curve across diagnostic subclass, form, and rhythm categories, and atrial fibrillation classification) and physiologic saliency alignment (Relative Saliency Score)surrogate

Domain-aligned ECG foundation models provide superior diagnostic performance and physiologic saliency alignment compared to general time-series models, though the latter can improve with parameter-efficient adaptation.

Cite This Study

HAN et al. (2026) studied Electrocardiogram (ECG) analysis. ECGFounder vs. MOMENT and PaPaGei was evaluated on Diagnostic performance (weighted AUC) and physiologic saliency alignment (Relative Saliency Score). ECGFounder achieved higher diagnostic performance (subclass AUC 0.932, AF AUC 0.983) compared to the general-purpose MOMENT model (subclass AUC 0.909, AF AUC 0.947).

synapsesocial.com/papers/6aacf5ed0c46fbdff987dd17https://doi.org/10.1186/s12911-026-03821-6
View Full Paper
Ask AI
Bookmark
Share