• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Thursday, November 19, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Health

ECG Foundation Models Require Targeted Fine-Tuning for Accurate ICU Arrhythmia Detection

by
November 19, 2026
in Health
Reading Time: 6 mins read
0
ECG Foundation Models Require Targeted Fine-Tuning for Accurate ICU Arrhythmia Detection

ECG Foundation Models Require Targeted Fine-Tuning for Accurate ICU Arrhythmia Detection

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

Pre-trained electrocardiogram foundation models fail to generalize out of the box to continuous monitoring in intensive care units, but targeted fine-tuning on curated public and institutional datasets achieves robust detection of sinus rhythm and atrial fibrillation, according to a study in Intensive Care Medicine Experimental. Conducted by researchers at Amsterdam University Medical Center and the AF Expert Panel, the investigation examined whether self-supervised transformer architectures could overcome acute domain shifts that undermine automated arrhythmia detection in critical care. Continuous electrocardiographic monitoring is standard in intensive care units, yet deploying deep learning algorithms at the bedside has long been hindered by signal disparities between static outpatient recordings and dynamic critical care telemetry. To evaluate foundation model transferability, the team developed a three-stage experimental protocol: assessing dataset annotation consistency across public archives using cross-validation, evaluating an off-the-shelf foundation model on real-world intensive care telemetry, and fine-tuning the architecture on curated public datasets with and without weakly labeled hospital data.

Atrial fibrillation is a common arrhythmia characterized by disorganized atrial activity and irregular heartbeats, occurring frequently among critically ill patients with severe conditions. Within intensive care units, atrial fibrillation is linked to adverse clinical outcomes, including stroke, heart failure, and overall mortality. Automated detection systems hold considerable promise for streamlining diagnosis and enabling earlier clinical interventions to protect vulnerable patients. However, reliable algorithmic classification during continuous bedside monitoring remains challenging. Adding to these difficulties, existing arrhythmia detection algorithms have predominantly been trained on ambulatory monitor data collected from non-ICU patients performing routine daily activities. Because critical care patients display fundamentally different physiological profiles, models developed on ambulatory cohorts suffer pronounced performance degradation when applied to continuous hospital telemetry.

Recent deep learning advances introduced electrocardiogram foundation models (ECG-FMs) that employ self-supervised learning to extract generalizable representations from large datasets without manual annotations. The researchers examined an open-source transformer pre-trained on approximately 1.5 million recordings from the PhysioNet Challenge 2021 and MIMIC-IV-ECG databases. Built upon a wav2vec 2.0-style design, the model utilizes a convolutional feature extractor to transform raw signals into latent embeddings, processed by a BERT-Base-like transformer encoder to capture waveform shapes and temporal structure across five-second segments. To provide resilience against missing channels, the model incorporates Random Lead Masking during pre-training, dropping leads randomly to accommodate arbitrary lead subsets at inference. In contrast, prior critical care studies by Chen and colleagues trained convolutional neural networks from scratch on roughly 10,000 recordings without pre-training. Those earlier systems merged atrial fibrillation and atrial flutter, achieving F1 scores of 0.64 to 0.67 and 55 percent precision, highlighting the promise of foundation models.

In Experiment 1, the authors evaluated annotation quality across public databases from the PhysioNet Challenge 2021, which consolidated international cohorts including CPSC, CPSC-Extra, INCART, PTB, PTB-XL, Georgia, and Chapman-Shaoxing. Because Challenge datasets employed heterogeneous annotation methods and variable quality control, the researchers excluded collections containing fewer than 200 cases of atrial fibrillation or sinus rhythm, eliminating Ningbo, INCART, CPSC-Extra, and PTB. To objectively gauge annotation reliability without exhaustive manual audits, the team fine-tuned ECG-FM independently on each remaining cohort—PTB-XL, Chapman-Shaoxing, Georgia, and CPSC—using stratified five-fold cross-validation. The authors posited that superior within-dataset classification performance would serve as an effective proxy for consistent internal labeling. PTB-XL and Chapman-Shaoxing delivered consistently superior F1 scores for both target rhythms, whereas Georgia and CPSC demonstrated lower, more erratic performance. These quantitative outcomes aligned closely with dataset documentation transparency: Chapman-Shaoxing and PTB-XL utilized rigorous physician review workflows, whereas Georgia lacked formal procedural documentation. Consequently, PTB-XL and Chapman-Shaoxing were chosen for downstream fine-tuning.

To evaluate models on real-world critical care data, the team curated two datasets from the Amsterdam University Medical Center institutional data warehouse. Bedside monitoring signals are continuously stored in the warehouse, but systems capture only the subset of leads actively selected for monitor display by clinical personnel. Concurrently, intensive care nursing staff chart cardiac rhythms hourly in the Epic electronic health record across 16 predefined categories. For model development, the researchers assembled the Amsterdam UMC-weak dataset by temporally pairing nursing observations with ten-second signal segments within a maximum matching window of one minute. Recordings displaying flatlines or missing signals across all leads were excluded. The resulting training dataset comprised 8,500 ten-second segments from 119 patients. The authors classified these annotations as weak supervision because they stem from routine clinical charting rather than dedicated electrocardiographic adjudication. Furthermore, nursing documentation in Epic is typically recorded hourly, introducing temporal uncertainty between charted timestamps and actual physiological rhythm assessment.

To establish an expert test standard, the researchers created the Amsterdam UMC Expert dataset from a distinct cohort of 349 patients over a separate timeframe, preventing patient overlap and data leakage. From 25,714 available samples, they selected 600 ten-second recordings via stratified random sampling to preserve the clinical distribution of the 16 documented rhythm categories while maximizing patient diversity. Each segment was processed with a zero-phase third-order Butterworth bandpass filter (1 to 47 Hz) and assigned to six physician annotators working in paired batches. Reviewers assessed signal quality and classified rhythms into sinus rhythm, atrial fibrillation, or other categories. Disagreements regarding noise, diagnostic classification, or annotator uncertainty were adjudicated by a senior cardiology clinician. Pairwise inter-rater agreement yielded median Cohen’s kappa values of 0.60 for sinus rhythm and 0.49 for atrial fibrillation. After excluding eight uninterpretable segments, the final expert benchmark contained 592 validated samples from 226 patients, comprising 365 sinus rhythm recordings, 117 atrial fibrillation cases, and 111 other rhythms.

Data preprocessing across all experimental phases involved upsampling signals to 500 Hz, applying zero-phase Butterworth filtering, and performing z-score normalization. In Experiment 2, the researchers evaluated the out-of-the-box generalization of an off-the-shelf ECG-FM model hosted on Hugging Face, fine-tuned on MIMIC-IV-ECG for 17 classes, directly against the Amsterdam UMC Expert dataset. This pre-existing model demonstrated poor transferability, obtaining F1 scores of 0.68 for sinus rhythm and 0.52 for atrial fibrillation despite high atrial fibrillation recall. In response, Experiment 3 tested targeted fine-tuning by appending two linear classification layers to ECG-FM for non-exclusive multi-label prediction on five-second input segments. The authors evaluated two training regimes: fine-tuning on curated public datasets alone (PTB-XL and Chapman-Shaoxing), and joint fine-tuning combining curated public data with the in-house Amsterdam UMC-weak dataset. Training utilized binary cross-entropy loss, with decision thresholds optimized on validation folds to maximize F1 scores across five random seeds, reporting bootstrap 95 percent confidence intervals and Cliff’s delta effect sizes.

Targeted fine-tuning on curated public data achieved robust generalization to the Amsterdam UMC Expert benchmark, producing F1 scores of 0.814 for atrial fibrillation and 0.914 for sinus rhythm. Incorporating weakly labeled hospital recordings provided further improvements on the expert set, raising F1 scores to 0.830 for atrial fibrillation and 0.920 for sinus rhythm. The addition of weak in-house data enhanced sensitivity and recall across test sets, accompanied by modest reductions in precision of 0.012 for atrial fibrillation and 0.014 for sinus rhythm. Performance gains were most striking on the Amsterdam UMC-weak test set, where atrial fibrillation F1 rose from 0.743 to 0.829 and sinus rhythm F1 increased from 0.957 to 0.970. Performance remained highest on the held-out public test set, where the publicly trained model attained F1 scores of 0.937 for atrial fibrillation and 0.977 for sinus rhythm, and the jointly trained model attained virtually identical values, indicating that weak local supervision preserved underlying model competencies.

The fine-tuned foundation model demonstrated notable adaptability across heterogeneous input leads, facilitated by Random Lead Masking during pre-training. Within the Amsterdam UMC Expert cohort, most recordings contained either two or eight available leads, with leads II and aVR appearing most frequently while precordial leads occurred intermittently. Stratified evaluations revealed that model classification accuracy remained stable between recordings with one to four leads and those with five to eight leads. Repeating evaluations on the second five seconds of each ten-second recording demonstrated consistent temporal reproducibility. An in-depth review of misclassified cases revealed distinct error patterns related to ICU complexity: atrial flutter was occasionally mistaken for atrial fibrillation, certain pacemaker rhythms were classified as sinus rhythm, and intra-segment rhythm transitions caused discrepancies between five-second model inputs and ten-second clinical annotations. Physical movement artifacts, ongoing bedside interventions, and potential training data mislabeling contributed to remaining false classifications, emphasizing critical care clinical complexity.

Model robustness to reduced lead configurations could also decrease reliance on full 12-lead electrocardiograms for rhythm verification. Nevertheless, several limitations warrant consideration. The investigation evaluated a single foundation model architecture, and internal consistency metrics may partly reflect architectural characteristics rather than absolute label quality. In addition, the expert-annotated Amsterdam UMC cohort comprised 592 segments from 226 patients, which constrained statistical power for demographic subgroup analyses across sex and age groups. Missing lead subsets in data storage, temporal rounding in documentation timestamps, and single-center data collection represent further operational constraints. The authors noted that translating five-second window inference to continuous bedside monitoring introduces an inherent latency. Future work will focus on multi-center data collection, federated validation paradigms, lead configuration harmonization, and expanding classification capabilities across broader critical care arrhythmias.

Subject of Research: Evaluation and fine-tuning of electrocardiogram foundation models for sinus rhythm and atrial fibrillation detection in continuous intensive care unit monitoring

Article Title: Leveraging ECG foundation models in critical care for sinus rhythm and atrial fibrillation classification

Article References: Galanty, M., Hulleman, M., Reuland, M. C., Ster, B. V. D., Vlaar, A. P. J., Sánchez, C. I., & the AF Expert Panel (2026). Leveraging ECG foundation models in critical care for sinus rhythm and atrial fibrillation classification. Intensive Care Medicine Experimental, 14(1), Article 111. https://doi.org/10.1186/s40635-026-00949-7

Image Credits: AI Generated

DOI: 10.1186/s40635-026-00949-7

Keywords: ECG foundation models, atrial fibrillation, sinus rhythm, intensive care unit, critical care monitoring, self-supervised learning, transformer architecture, domain adaptation, Amsterdam UMC, PhysioNet Challenge 2021, PTB-XL, Chapman-Shaoxing, weak supervision, random lead masking, arrhythmia classification, bedside telemetry, MIMIC-IV-ECG, inter-rater agreement

News Source: Ophelia Keating. (November 19, 2026). ECG Foundation Models Require Targeted Fine-Tuning for Accurate ICU Arrhythmia Detection. Scienmag.

Tags: Amsterdam UMCarrhythmia classificationAtrial Fibrillationbedside telemetryChapman-Shaoxingcritical care monitoringdomain adaptationECG foundation modelsintensive care unitinter-rater agreementMIMIC-IV-ECGPhysioNet Challenge 2021PTB-XLrandom lead maskingSelf-Supervised Learningsinus rhythmtransformer architectureweak supervision
Share12Tweet7Share2ShareShareShare1

Related Posts

Mechanosensitive Gene Signature Predicts Glioma Patient Survival Across Multiple Clinical Cohorts

Mechanosensitive Gene Signature Predicts Glioma Patient Survival Across Multiple Clinical Cohorts

November 19, 2026
Periodontitis Reveals Immune Heterogeneity Through Fusobacterium-Reactive Antibody Profiles

Periodontitis Reveals Immune Heterogeneity Through Fusobacterium-Reactive Antibody Profiles

November 19, 2026

Bedside Airway Ultrasound Enables Rapid Identification of Rare Neonatal Tracheal Agenesis

November 19, 2026

Compounding Discrimination and Isolation Burden Diverse LGBTQIA+ Dementia Caregivers

November 19, 2026

POPULAR NEWS

  • Virtual Reality Cycling Proves Feasible and Potentially Safe in ICU Pilot

    Virtual Reality Cycling Proves Feasible and Potentially Safe in ICU Pilot

    29 shares
    Share 12 Tweet 7
  • Clinical Documentation and Imaging Practices for Positional Vertigo Vary After Implementation

    29 shares
    Share 12 Tweet 7
  • Green Tea Compound Boosts Irinotecan Efficacy and Curbs Angiogenesis in Mice

    29 shares
    Share 12 Tweet 7
  • Gradual Dexamethasone Dose Reduction Manages Chemotherapy-Related Delirium in Lung Cancer Patients

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Mechanosensitive Gene Signature Predicts Glioma Patient Survival Across Multiple Clinical Cohorts

Periodontitis Reveals Immune Heterogeneity Through Fusobacterium-Reactive Antibody Profiles

Migratory Ungulates Face Elevated Extinction and Speciation Rates in Macroevolutionary Analysis

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 88 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.