Standard defibrillation
A shock delivered through the standard anterior–lateral pad configuration.
A foundation model for ventricular fibrillationTowards
Towards
personalized
defibrillation.
Can the ECG help us understand how a patient will respond to a shock, even while chest compressions are ongoing?
Read the signal.
Understand the response.VF-FM
Understand the response.VF-FM
ECG during chest compressionsElectrical activity
ImpedanceCompression information
WaveformRepresentationResponse
Funded research · Preprocessing beginningLawrence Leroux · Principal investigatorClinical medicine × Signal learning
01 / The clinical questionA different signal.
A different signal.
A different response?
Ventricular fibrillation is a disorganized heart rhythm treated with defibrillation. Yet the response to a shock differs between patients and over the course of a resuscitation.
VF-FM asks whether learned ECG representations can reveal some of that difference. Its central challenge is to retain useful information in a signal distorted by the very compressions keeping circulation going.
The longer-term question is whether that information can help teams choose the timing and strategy of defibrillation more precisely.
Survival to hospital discharge
405 patients with refractory VF · Cluster-randomized trial
Standard13.3%
Vector change21.7%
Double sequential30.4%
These are outcomes from the published trial after refractory VF—not VF-FM predictions or evidence for selecting a patient’s first shock.
Cheskes et al. · NEJM, 2022Vector change
A different current pathway, using an anterior–posterior pad configuration.
Double sequential
Two defibrillators delivering shocks in rapid sequence through two pad configurations.
02 / The protocolFrom learning a waveform
From learning a waveform
to asking a clinical question.
A retrospective model-development study using existing recordings. The five phases below describe the planned work, from general ECG pretraining to strategy-specific predictions and calibration.
Where we are now
We are beginning data preprocessing as the remaining data-sharing agreements are finalized. Model training and validation are the next stages of the study.
Phase 01 · Study plan
Learn the language of the ECG.
Create a general foundation model for single-lead ECG signals. Pretrain a convolutional–Transformer model on diverse ECG recordings using self-supervised learning to learn reusable representations of electrical activity across rhythms and recording conditions.
General single-lead ECG recordingsA general ECG foundation model
- Masked signal modeling
- Hide parts of a recording and learn to reconstruct them from the surrounding signal.
- Contrastive learning
- Bring representations of related signal views closer together while distinguishing different recordings.
What this phase asksCan self-supervised learning produce a general foundation model for single-lead ECG signals?
Study designRetrospective AI development
Previously collected recordings; no new patient enrollment in this phase.
Primary clinical outcomeReturn of spontaneous circulation
Organized activity with evidence of perfusion, distinguished from VF termination alone.
Signal of interestPre-shock ECG during CPR
Four-second windows, compared with segments recorded during pauses.
03 / The dataDifferent datasets.
Different datasets.
Distinct jobs in the study.
The protocol connects general ECG learning with prehospital recordings and randomized defibrillation strategies. Cohort sizes below are planning figures; the final usable sample depends on access, linkage and signal quality.
General ECG repositories
PTB-XL · MIMIC-IV · Chapman · CODE-15 · MIT-BIH · I-CARE
Self-supervised pretraining across rhythms and recording conditions.
Multiple repositoriesFinal pretraining corpus to be assembled
Urgences-santé
Montréal · Prehospital resuscitation
Adaptation to VF and chest-compression artifacts using ECG and impedance.
Continuous recordingsFinal cohort to be established
ARREST registry
Amsterdam · Out-of-hospital cardiac arrest
Annotated standard shocks for response prediction.
Multiple thousands of shocksFinal cohort to be established
DOSE VF
Canada · Randomized defibrillation strategies
Adaptation and evaluation across standard, vector-change and double sequential shocks.
2,853 shocksPlanning inventory from the Sunnybrook presentation
DOSE VF · Planned shock inventory
2,086Standard
443Vector change
324Double sequential
Multiple shocks can belong to the same patient. The smaller alternative-strategy cohorts make transfer learning and uncertainty particularly important.
Which recordings enter the analysis?
Define the cohort before fitting the model.
The proposed shock-outcome analysis includes out-of-hospital cardiac arrest with VF and a recorded pre-shock ECG, with impedance where available. Non-shockable rhythms, traumatic arrests and recordings with missing or corrupted required data are excluded.
04 / EvaluationA useful model needs
A useful model needs
more than a good score.
The evaluation plan examines discrimination, probability calibration and performance under the conditions in which the model is intended to operate.
Distinguish outcomes.
Assess discrimination alongside sensitivity, specificity and predictive values. Precision–recall analysis helps describe performance when successful shocks are uncommon.
Check the probabilities.
Use reliability plots and the Brier score to examine agreement between predicted probabilities and observed outcomes.
Test the difficult signal.
Compare performance with and without compression artifacts, across model architectures and defibrillation strategies.
Keep each patient together.
Group a patient’s recordings and repeated shocks within the same partition. Examine performance across age and sex, alongside overall results.
From prediction to patient benefit
A retrospective response model would be a step toward prospective research on whether using its predictions improves resuscitation decisions and outcomes.
Research informing VF-FMThe work
The work
we build on.
Selected clinical and machine-learning studies that inform the project’s question, methods and early experiments.
The clinical question
Defibrillation Strategies for Refractory Ventricular Fibrillation
The DOSE VF trial provides the clinical starting point for studying response to standard, vector-change and double sequential defibrillation.
Cheskes et al. · New England Journal of Medicine · 2022Learning during CPR
A method to predict ventricular fibrillation shock outcome during chest compressions
Prior work demonstrated that shock-outcome prediction is possible during CPR. It provides an important reference for evaluating how well VF-FM handles compression artifacts.
Coult et al. · Computers in Biology and Medicine · 2021The foundation-model approach
ECG-FM: An Open Electrocardiogram Foundation Model
An open model using contrastive and generative self-supervised learning. Its approach informs the move from general ECG representations to a specific clinical prediction task.
McKeen et al. · arXiv · ECG-FM · 2025 revisionA public starting point
ECG waveform dataset for predicting defibrillation outcome in out-of-hospital cardiac arrested patients
Cardially contains recordings from 260 patients, with pre-shock and post-shock waveforms. This public dataset supports early waveform experiments and comparison with published methods.
Benini et al. · Data in Brief · 202105 / People & supportClinical experience.
Clinical experience.
Engineering expertise.
A shared question.
The project brings together resuscitation clinicians, engineers and prehospital researchers across Montréal, Toronto, Amsterdam, Paris and British Columbia.
Lawrence Leroux
Université de Montréal
Concordia University
Sheldon Cheskes
Sunnybrook Centre for Prehospital Medicine
University of Toronto
Academic supervision
Jamal BentaharInformation Systems Engineering
Concordia University
Concordia University
Lyes KademCardiovascular Engineering
Concordia University
Concordia University
Yiorgos Alexandros CavayasHôpital du Sacré-Cœur de Montréal
Université de Montréal
Université de Montréal
Collaborators
Lionel LamhautSAMU de Paris · Université Paris Cité
Michiel HullemanAmsterdam UMC · ARREST registry
Ian DrennanUniversity of Toronto · Sunnybrook
Paul DorianUniversity of Toronto · Li Ka Shing Knowledge Institute
Linda TurnerSunnybrook Centre for Prehospital Medicine
Jonathan KwongSunnybrook Centre for Prehospital Medicine
Michael FeldmanSunnybrook Centre for Prehospital Medicine
Jacob HuttonBC Emergency Health Services · University of British Columbia
Awarded research support
Making the work possible.
Two complementary grants support VF-FM. Lawrence Leroux is principal investigator on both awards.
Réseau santé numériqueCAD 40,000Self-supervised ECG models for defibrillation-response prediction.
Build knowledge
others can build on.
The dissemination plan includes model weights, training code, evaluation scripts and documentation. As the work progresses, this page can bring together research updates, publications and resources for collaborators.
Discuss the research