PulseTrendingJournal ClubResearchersJournalsExplore
Instagram
HomeTrendingJournal ClubExplore
Synapse
⌘+K
Synapse
August 2, 2026European Heart Journal - Digital HealthOpen Access

Explainable Electrocardiogram Interpretation using Deep Learning-based Semantic Segmentation

View Full Paper
Ask AI
Bookmark
Share

Why the study?

Accurate ECG waveform delineation, rhythm classification, and median beat generation are interdependent steps whose joint modeling improves consistency for downstream computerized diagnostic tasks.

Population

1,931 annotated ECGs, 33,093 ECGs with diagnostic statements, and internal (n = 988) and external (n = 1,303) test sets

Comparison

DeepLabV3-based neural network model vs reference medians from Marquette 12SL and University of Glasgow

Design

Model development and validation study

Key result

A deep learning model for ECG analysis achieved mean delineation errors of -0.9 to 1.8 ms and downstream classification performance superior to Uni-G medians for 7 of 10 diagnostic labels.

Authors

BABauke ArendsBSBas B.S. SchotsPKParmenion Koutsogeorgos

Discussion

Loading...

Member takes

Overview

May enhance ECG diagnostic consistency via joint segmentation; leaves open prospective validation before clinical adoption.

Key Points

  • This study aims to develop a lead-agnostic model for electrocardiogram analysis, improving accuracy in waveform delineation and rhythm classification.
  • Developed a DeepLabV3-based neural network to segment ECG leads into 20 classes using 1,931 annotated ECGs and 33,093 with verified diagnostics.
  • Evaluated performance on internal (N=988) and external (N=1,303) test sets, comparing median beats with existing references.
  • An interactive web tool was launched to support ongoing research and usability.
  • In the external test set, mean errors were -0.9 ± 10.4 ms for PQ-interval, 0.2 ± 7.4 ms for QRS-duration, and 1.8 ± 16.0 ms for QT-interval.
  • Achieved weighted F1 scores of 0.94 for P waves and 0.89 for QRS complexes in rhythm classification.
  • Downstream classification was statistically superior to Uni-G medians for 7 out of 10 diagnostic labels.

Structured PICO

P
Population
37,315 ECGs used to train and evaluate a deep learning model for waveform delineation, rhythm classification, and median beat construction.
E
Exposure
DeepLabV3-based neural network for lead-agnostic segmentation of ECG leads into 20 waveform and rhythm classes
C
Comparator
Reference medians from Marquette 12SL and University of Glasgow (Uni-G)
O
Outcome
Delineation errors for PQ-interval, QRS-duration, and QT-interval, rhythm classification F1 scores, and downstream classification performancesurrogate

A deep learning-based semantic segmentation model provides robust, vendor- and lead-agnostic ECG analysis with performance comparable or superior to standard commercial algorithms.

Cite This Study

Arends et al. (2026) studied Electrocardiogram (ECG) analysis (n=37,315). DeepLabV3-based neural network vs. Marquette 12SL and University of Glasgow (Uni-G) medians was evaluated on Delineation errors and downstream classification performance. A deep learning model for ECG analysis achieved mean delineation errors of -0.9 to 1.8 ms and downstream classification performance superior to Uni-G medians for 7 of 10 diagnostic labels.

synapsesocial.com/papers/6a6eeaa21b0468a7eeab2f65https://doi.org/10.1093/ehjdh/ztag122
View Full Paper
Ask AI
Bookmark
Share