PaperlayerView exact revision on arXiv

Source-linked AI summary

Wearable ECG Quality Assessment: A Deep Learning and Ambulatory Context-Awareness Approach

Xiaopeng Mao, Marike Weisbjerg, Sadasivan Puthusserypady

arXiv:2609.29396v1cs.AIeess.SP

TL;DR

Ambulatory ECG quality assessment matters because free-living recordings have particularly severe noise compared with in-hospital recordings. This paper trains a deep-learning signal-quality model on contextualized ambulatory ECG data and finds that normal-PQRST rhythms are more likely to be predicted correctly than pathological waveforms with different morphologies.

  • Problem

    Free-living ambulatory ECG recordings have particularly severe noise compared with in-hospital recordings, making it important to prevent noise from obscuring the ECG.

    Source: §I · p. 1

  • Method

    The study trains a deep-learning signal-quality model on contextualized ambulatory ECG data, evaluates it on public databases, and analyzes ECG noise alongside physical context.

    Source: §I · p. 1

  • Results

    ECG rhythms with normal PQRST character were more likely to be predicted correctly, whereas pathological waveforms with different morphologies were more likely to lead to wrong predictions.

    Source: §IV · p. 5

  • Takeaways & Limitations

    Identifying physical context in relation to ECG signal quality marks a small beginning in signal-context research.

    Source: §IV · p. 5

    Because the clean and noisy ECGs were not reviewed by a cardiologist, the contextualized analysis results warrant caution.

    Source: §IV · p. 5

Abstract

from arXiv · show

This paper presents and evaluates a Deep Learning-based (DL-based) Signal Quality Assessment (SQA) model to distinguish between clean and noisy ambulatory Electrocardiograms (ECG). The model is trained on Copenhagen Center for Health Technology-Contextualized Arrhythmia Database (CACHET-CADB), which, to the best of our knowledge, is the first ambulatory ECG database with both physical and patient-reported contextual data. The model shows stable performance on different databases such as MIT-databases and the latest PyhsioNet/Cinc Challenge 2021 databases. Subsequently, the paper demonstrates how complicated ECG noise can be investigated by the SQA model and the physical contextual data.

I. INTRODUCTION · II. METHODS · A. Datasets

The study evaluates a deep-learning ECG signal-quality model trained on contextualized ambulatory data across public databases and explores how physical context can inform ECG-noise analysis. It also defines the CACHET-CADB labels and the clean/noisy test task.

I. INTRODUCTION

  • The study evaluates a deep-learning signal-quality model trained on CACHET-CADB across MIT-BIH-Arrhythmia and PhysioNet/Cinc Challenge 2021 databases.It also aims to assess the model’s strengths and weaknesses across databases.Source: §I · p. 1
  • The study explores whether combining the model with physical ECG context can deepen insight into noise and help improve the recording environment.Source: §I · p. 1
  • Free-living ambulatory ECG can have more severe noise than in-hospital recordings, making it important to prevent noise from obscuring the signal.Source: §I · p. 1

A. Datasets

  • CACHET-CADB contains 259 days of continuous ambulatory ECG from 24 people: 21 with atrial fibrillation and three healthy individuals.Signals were sampled at 1,024 Hz from one precordial lead.Source: A. Datasets · p. 1
  • Two cardiologists independently annotated 1,602 ten-second ECGs as normal sinus rhythm, atrial fibrillation, noise, or other.Source: A. Datasets · p. 1
  • For signal-quality assessment, the labelled ECGs serve as the test set: normal sinus rhythm, atrial fibrillation, and other rhythms are “clean,” while noise is “noisy.”Source: A. Datasets · p. 1

B. Continuous Wavelet Transformation

The section describes converting 1D ECGs into 2D CWT scalograms, whose time-frequency representation enables image classification of clean versus noisy signals. It also specifies the frequency scaling used and the InceptionV3 classifier used in this work.

B. Continuous Wavelet Transformation

C. Model architecture

The model uses frozen InceptionV3 feature extraction followed by dense layers, with dropout added to stabilize training.

C. Model architecture

  • The 224 × 224 × 3 input passes through frozen InceptionV3 modules to produce 5 × 5 × 2048 features, which are flattened before dense layers; dropout in Dense layer 1 stabilizes training.Freezing all InceptionV3 parameters is used to prevent overfitting.Source: C. Model architecture · p. 2

D. Contextualized analysis

The contextualized analysis uses device-measured physical parameters alongside ECG quality assessment, with ECG preprocessing, segment generation, and model training also specified.

D. Contextualized analysis

  • The analysis uses barometer- and accelerometer-measured physical parameters, including body positions and activity classes, for contextualization.The parameters are listed in Table III.Source: D. Contextualized analysis · p. 2, Table III · p. 2
  • ECGs are bandpass-filtered at 0.5–50 Hz, smoothed, and resampled to 256 Hz; 20,000 clean and 20,000 noisy 10-second segments are generated, with the method adjusted for AFIB.Segment labels use decision rules combined with template matching from.Source: D. Contextualized analysis · p. 2
  • Model training uses Adam, cross-entropy loss, Python and TensorFlow, and three nodes of four Tesla V100 32 GB GPUs with NVlink.The optimizer and loss are cited as and, respectively.Source: D. Contextualized analysis · p. 3

III. RESULTS · A. Performance on CACHET-CADB test set

On the CACHET-CADB test set, the paper presents the model’s performance as a confusion matrix and reports stable convergence of both loss curves under the optimized hyperparameters.

A. Performance on CACHET-CADB test set

B. Performance on MIT-BIH-NSR, MIT-BIH-Arrhythmia and MIT-BIH-Noise Stress Test Database

The model was evaluated on MIT-BIH-NSR, MIT-BIH-Arrhythmia, and MIT-BIH-Noise Stress Test data using Lead 2 ECG. It tended to fail on arrhythmia records with abnormal morphology, including paced beats combined with PVC.

B. Performance on MIT-BIH-NSR, MIT-BIH-Arrhythmia and MIT-BIH-Noise Stress Test Database

  • Evaluation used three randomly chosen MIT-BIH-NSR records, the entire MIT-BIH-Arrhythmia database, and three pure noise signals, with Lead 2 ECG.The noise signals represented baseline wander, muscle artifact, and electrode motion artifact; Lead 2 was selected because it most clearly expresses the QRS complex.Source: B. Performance on MIT-BIH-NSR,… · p. 3
  • The model tended to fail on arrhythmia records with abnormal morphology, such as paced beats combined with PVC.Source: B. Performance on MIT-BIH-NSR,… · p. 4

C. Performance on PhysioNet/CinC challenge 2021

Evaluations on the PhysioNet/CinC challenge 2021 databases show that ECG morphology deviations pose more difficulty for the model than changes in beat frequency.

C. Performance on PhysioNet/CinC challenge 2021

  • Chapman-Shaoxing evaluation used 10-second clean ECGs from 10,646 heart patients, with model performance categorized by rhythm.The database was sampled at 500 Hz; 10,247 ECGs were available from the challenge.Source: C. Performance on PhysioNet/CinC… · p. 4, Table VI · p. 4
  • On Ningbo, the largest 2021 challenge database, evaluation covered all ECG types using manually identified annotations across 40,258 10-second ECGs from different patients.34,905 ECGs were available from the challenge, and recordings were sampled at 500 Hz.Source: C. Performance on PhysioNet/CinC… · p. 4
  • The model is more error-prone when ECG morphology departs substantially from normal sinus rhythm, while sinus bradycardia or tachycardia does not severely affect performance.Source: C. Performance on PhysioNet/CinC… · p. 5

D. Contextualized analysis

Context histograms show noisy ECGs concentrated during upright sitting/standing and unknown context combinations, while upright sitting/standing is also most frequent for clean ECGs. The analysis suggests investigating unknown contexts by tracing their time periods and asking patients what was happening.

D. Contextualized analysis

IV. DISCUSSION

The discussion finds that normal-PQRST ECG rhythms were more likely to be predicted correctly than pathological morphologies, and cautions that contextualized analyses rely on clean/noisy labels not reviewed by cardiologists.

  • Sinus-wave predictions were particularly accurate, but predictions for abnormal atrial waves such as AFIB were difficult.Source: §IV · p. 5
  • Abnormal ST segments, ST intervals, and cardiac blocks were difficult to predict; CACHET-CADB lacked patients with diseases inducing these rhythms, consistent with MIT-BIH-Arrhythmia results in Table V.Source: §IV · p. 5
  • Normal-PQRST ECG rhythms were more likely to be predicted correctly, while pathological waveforms with different morphologies more often produced wrong predictions.Source: §IV · p. 5
  • Because cardiologists did not review the clean/noisy ECGs, the contextualized analyses in Figures 6 and 7 warrant caution.Source: §IV · p. 5
  • Identifying physical context in relation to ECG signal quality marks a small beginning in the broader area of signal-context research.Source: §IV · p. 5
Loading 2609.29396v1…