AI-powered ECG Analysis for Deadly Arrhythmias and ICI Myocarditis (ELDORA)

"ELDORA (Efficient Deep Learning Approaches for the Rapid and Interpretable Detection of Deadly Arrhythmias in ECG Data) is an observational, non-interventional project focused on ECG-based artificial intelligence. Its overarching objective is to develop and optimize clinical-grade AI-powered tools for: (1) digitizing, standardizing and analyzing heterogeneous ECG signals, including real-life analog/paper-derived and digital recordings; and (2) supporting clinical decision research for two sudden-cardiac-arrest-prone conditions: Torsades-de-Pointes (TdP) risk prediction in established long QT syndrome, whether congenital or drug-induced, and diagnosis, prognosis and risk prediction for immune checkpoint inhibitor-induced myocarditis.

The project will consolidate diverse ECG and clinical datasets into ECGInsight, a harmonized database planned to include approximately 49 national and international ECG cohorts, around 127,000 subjects and up to about 10 million 10-second ECG equivalents. Cohorts cover a broad spectrum of health states and cardiovascular conditions, including healthy volunteers, congenital and drug-induced long QT/TdP populations, cancer patients treated with immune checkpoint inhibitors with or without myocarditis, heart transplant, diabetes, obesity and hormonal phenotyping cohorts. Data include raw ECG waveforms, automatic and expert annotations, scanned paper ECGs where applicable, demographics, clinical characteristics, laboratory results, drug exposure and hormono-metabolic assessments near the time of ECG acquisition.

Data curation will include mapping of cohort variables and clinical concepts into an ELDORA glossary, using controlled terminologies where appropriate, including ICD-10, MedDRA, OMOP and ATC for drug exposure. ECGs will be standardized using the project toolkit and integrated in a secure, GDPR-compliant infrastructure. Access is intended to be controlled and limited to approved researchers/clinicians under the project governance. The study involves no treatment allocation, no investigational medicinal product and no direct AI-driven change to patient care. Model performance will be evaluated using standard classification and regression metrics, including AUC, sensitivity, specificity, F1 score, accuracy, MAE, RMSE, R2 and Bland-Altman analyses, as appropriate to each task."