AI ECG Analysis to Predict Deadly Arrhythmias and ICI Myocarditis (ELDORA) — a clinical trial (ClinicalTrials.gov)

1The faint warning signs in an ECG

An electrocardiogram (ECG) records the heart's electrical activity as a waveform. It is inexpensive, non-invasive, and generated in enormous volumes.

Yet early warning signs of life-threatening arrhythmias — such as long QT syndrome and Torsades-de-Pointes — or signs of myocarditis triggered by the immune checkpoint inhibitors used in cancer immunotherapy can appear as subtle shifts in the trace that are easy for the human eye to miss.

Artificial intelligence (AI), which can learn fine-grained patterns from large numbers of ECGs, is hoped to help surface such hard-to-see risks earlier, and this study aims to develop and validate predictive models toward that goal.

2An observational study of existing data

A defining feature of this work is that it is observational rather than interventional: it analyzes already-collected ECGs and de-identified clinical information instead of testing a new treatment.

ECGs that arrive in different formats from different sites are standardized into a common form and assembled into a harmonized database named ECGInsight, after which the project aims for interpretable models whose reasoning a human can follow.

Because clinical adoption of medical AI depends on transparency and verifiability as much as on accuracy, the emphasis on data harmonization and explainability addresses the practical hurdles this field actually faces.

3Unlocking medical data that usually sits idle

Seen more broadly, the study is also an example of unlocking medical data that often sits idle, pooling and reusing multiple cohorts across countries rather than relying on a single site.

As cancer immunotherapy becomes more widespread and monitoring for cardiac side effects grows more important, using AI to extend the reach of a cheap, ubiquitous test could help with the public-health challenge of detecting rare but fatal events early. Even so, what is described here is the study's aim and hypothesis; it does not mean the effectiveness or safety of the AI has been established.

4Aligning formats before solving anything

Pooling data across institutions requires first bringing formats into line. This study builds that step into the design.

1. 1Formats differ by institutionECG formats vary from site to site
2. 2Standardise to a common formatAssembled into a harmonised database named ECGInsight
3. 3Build an interpretable modelAimed at a form in which why a judgement was made can be followed
4. 4Secure transparencyAcceptance in clinical use requires verifiability, not accuracy alone

The ECG is cheap, non-invasive and available in volume, while the precursors of lethal arrhythmia and the signs of myocarditis from immune checkpoint inhibitors appear as slight changes in waveform easily missed by eye. The study also pools and reuses multiple cohorts at home and abroad rather than resting on a single centre.