An AI-based framework to identify duplicate ECGs in clinical trials
How the ECG Mirror Participant tool helps sponsors detect anomalous ECG data, investigate potential duplicate recordings, and strengthen data integrity oversight across clinical trial sites.
Authors:
- Alain Gay, M.D.
- Jean-Philippe Couderc, MBA, Ph.D.
- Luc Dekie, Ph.D.
- Todd Rudo, M.D.
- Vickas Patel, M.D., Ph.D.
- Hui Zhao, Ph.D.
Summary
This white paper introduces an AI-based framework for identifying duplicate or mislabeled ECGs in clinical trials. By learning participant-specific electrophysiologic signatures and calculating dissimilarity scores across ECG recordings, the approach helps detect unexpectedly similar data between participants and surfaces potential data integrity risks for sponsor review.
Why ECG data integrity matters
Clinical trial sponsors rely on accurate ECG collection to evaluate drug safety, but conventional monitoring may not reveal every protocol deviation or anomalous recording pattern across geographically dispersed sites and large datasets.
Mislabeled, duplicated, or fabricated ECGs can obscure true cardiac safety signals, distort efficacy and safety assessments, delay access to effective therapies, and increase regulatory and ethical risk.
What is a “mirror participant”?
A mirror participant refers to a single individual being enrolled as multiple study participants in a clinical trial, resulting in highly similar ECG morphologies appearing under different participant IDs.
How the ECG Mirror Participant framework works
The white paper maps the process into a practical review workflow that combines AI-generated similarity scoring with expert interpretation.
01
Learn participant-specific ECG signatures
The model analyzes 12-lead ECG morphology and learns stable, participant-specific patterns rather than relying on superficial visual similarity.
02
Calculate dissimilarity scores
ECG strips are compared within and across study participants. A low dissimilarity score indicates that two tracings are unexpectedly similar. This is true only when comparing ECGs across study participants. Within the same participant, a low dissimilarity score is expected.
03
Generate study-level heatmaps
Study-level maps summarize the minimum dissimilarity between participant pairs, making suspicious cross-participant similarity patterns easy to spot.
03
Drill into intra-participant maps
When suspicious data are found, intra-participant maps help reviewers isolate ECG strips that appear inconsistent with the rest of a participant’s longitudinal record.
Figure 1, Panel A: Study-level map showing minimum dissimilarity scores across participant pairs. Light cells indicate unexpectedly similar ECGs across different participant IDs.
Figure 1, Panel B: Intra-participant map used to identify ECG strips that appear inconsistent with the rest of a participant’s longitudinal record. Note: Dark cells indicate unexpectedly different ECGs from the same participant.
Why this matters for sponsors
Strengthen oversight
Review more ECG tracings than a manual process alone could realistically handle while maintaining consistency across large studies.
Increase data confidence
Identify anomalous patterns earlier and build greater confidence in the evidence submitted to regulatory authorities.
Prioritize expert review
Use AI-generated outputs to focus cardiologist attention on the sites, participants, and ECG strips most likely to require investigation.
Explore the full white paper
Get the complete white paper or get started and continue the conversation.
FAQs
How does the tool detect duplicate or mislabeled ECGs?
The framework calculates dissimilarity scores across ECG pairs, highlights suspicious cross-participant similarity at the study level, and then drills into participant-specific maps to locate inconsistent recordings.
Does this replace cardiologist review?
No. The AI framework helps surface suspicious patterns efficiently, while cardiologists review flagged ECG tracing pairs and participant-level outputs to confirm whether further investigation is warranted.
Can thresholds vary by study type?
Yes. The paper notes that thresholds may be adjusted for certain study types, including long-term follow-up studies where physiologic change within a participant may affect ECG morphology over time.
