Development of an Artificial Intelligence Algorithm to Detect Pathological Repolarization Disorders on the ECG and the Risk of Ventricular Arrhythmias

Recruiting

• Conditions: Cardiac Disease; Ventricular Arrythmia

• Registration Number: NCT05829993
• Lead Sponsor: Assistance Publique - Hôpitaux de Paris

Brief Summary

Torsades de Pointes (TdP) are potentially fatal ventricular arrhythmias that are promoted by prolonged ventricular repolarization (Long QT, LQT). The different forms of LQT result from inhibition of cardiac potassium currents (IKr and IKs) or activation of a late sodium current (INaL). These alterations may be either congenital (3 types: cLQT-1: IKs, cLQT-2: IKr, cLQT-3: INaL) or drug-induced (diLQT, via inhibition of IKr).

More than 100 medications have received marketing authorization despite a known risk of TdP, due to a favorable benefit-risk ratio (e.g., hydroxychloroquine).

QTc, which represents the duration of ventricular repolarization (in milliseconds) - defined as the time from the beginning of the QRS complex to the end of the T wave, corrected for heart rate - is prolonged in all forms of LQT.

Specific T-wave abnormalities, depending on the altered ion currents, have been described and can help differentiate the various types of congenital or drug-induced LQT.

However, screening for LQT and TdP risk, both at the individual and population levels, currently relies mainly on isolated QTc evaluation and genetic testing, which often takes considerable time to return.

Thus, limiting ECG analysis to QTc measurement alone offers low predictive value, as the ECG contains a wealth of additional information beyond a single interval.

The investigator recently demonstrated that artificial intelligence (AI)-based ECG analysis using deep-learning convolutional neural networks can detect more discriminative features of the ECG for predicting the type of LQT and the risk of TdP, going beyond QTc alone.

Using these techniques, the investigator developed a model with probabilistic modules capable of: predicting TdP risk, identifying LQT subtypes (scores ranging from 0 to 100%), and quantitatively measuring ECG parameters such as QTc, heart rate, PR, and QRS duration.

The objective of this project is to prospectively validate our model in real-world conditions across various departments within AP-HP, for:

Automatic measurement of QTc, and Identification and classification of LQT types and TdP risk.

Detailed Description

Background Torsades de Pointes (TdP) are rare but potentially fatal ventricular arrhythmias promoted by a prolongation of ventricular repolarization, observed on electrocardiogram (ECG) as Long QT (LQT). This prolongation may result from either genetic or acquired alterations in cardiac ion channels.

The congenital forms of LQT (cLQT) have an estimated prevalence of 1 in 2,000 to 3,000 and are caused by mutations affecting specific ionic currents: cLQT-1, associated with decreased IKs current (KCNQ1 gene), cLQT-2, with decreased IKr current (KCNH2 gene), and cLQT-3, with increased late sodium current INaL (SCN5A gene).

Acquired or drug-induced LQT (diLQT) typically results from the inhibition of IKr by medications. Over 100 drugs currently on the market, including hydroxychloroquine and azithromycin, are known or suspected to prolong the QT interval, with some associated with up to 7% incidence of TdP.

QTc, measured from the beginning of the QRS complex to the end of the T wave and corrected for heart rate, is the principal biomarker used to assess repolarization duration and risk of TdP. A QTc ≥500 ms is associated with a significantly increased risk.

However, QTc measurement is subject to high inter- and intra-observer variability and lacks sufficient predictive performance to differentiate cLQT types or detect TdP risk in routine clinical settings.

Rationale Current clinical practices rely primarily on QTc measurement and genetic testing, which are not optimal in real-world conditions. Given the volume of digital ECGs now available, manual interpretation is increasingly limited.

Artificial Intelligence (AI), and particularly deep-learning models, have demonstrated promising results for ECG analysis beyond QTc alone.

The project team has developed and patented (AP-HP patent: WO/2020/245322) an AI model capable of:

automatically quantifying ECG parameters (QTc, PR, QRS, heart rate), generating probabilistic scores (0-100%) for congenital (cLQT-1, -2, -3) or drug-induced LQT, and estimating the risk of TdP.

This model showed strong retrospective performance. A prospective validation in clinical practice is now necessary.

Study Design Multicenter, cross-sectional, non-interventional study across 14 AP-HP hospitals over 42 months, enrolling 5,000 participants undergoing ECG as part of care or research. Data collected include digital ECGs (.xml), clinical history, medications, and genetic data if available.

Workflow Digital ECGs are anonymized and stored centrally. QTc values are determined using triplicate ECGs, threshold method, and Fridericia correction. AI analysis is then compared with reference. Associated clinical and genetic data enrich model evaluation.

Expected Impact If validated, this AI model could be integrated into ECG acquisition devices to enable real-time QTc measurement, risk classification, and better triage. This would improve detection and management of arrhythmogenic risk, especially in settings lacking rhythmology expertise.

Study Timeline

• Study First Submitted: 2023-04-13
• Study First Submitted QC: 2023-04-13
• Study First Posted: 2023-04-26
• Study Start: 2023-11-28
• Status Verified: 2025-07
• Last Update Submitted: 2025-07-28
• Last Update Posted: 2025-07-31
• Primary Completion: 2027-05-28
• Study Completion: 2027-06-28

Recruitment & Eligibility

• Status: RECRUITING
• Sex: All
• Target Recruitment: 5000

Inclusion Criteria

• Age ≥ 18
• Patients or subjects taken care in recruiting centres for which an ECG is indicated
• No opposition to participation in the study

Exclusion Criteria

• Medical contraindication for ECG
• Subjects with pacemaker-driven QRS

Study & Design

• Study Type: OBSERVATIONAL
• Study Design: Not specified

Primary Outcome Measures

Name	Time	Method
Concordance of QTc measurement between the reference method and deep-learning model at 500 ms threshold	Day 0	Evaluate the concordance (Kappa coefficient, Κ) of QTc measurement between the reference method (triplicated averaged 10-second ECG complexes, "threshold" technique, Fridericia correction) and the deep-learning model in patients classified as having QTc ≥500 ms versus \<500 ms.

Secondary Outcome Measures

Name	Time	Method
Diagnostic performance of AI-generated scores for congenital long QT types 1, 2, and 3	Day 0	Evaluate sensitivity, specificity, positive predictive value, and negative predictive value (with 95% confidence intervals) for three AI-generated scores (0-100%) estimating the probability of congenital long QT types 1, 2, and 3 (cLQT-1: IKs KCNQ1, cLQT-2: IKr KCNH2, cLQT-3: INaL SCN5A). Scores will be compared between genetically confirmed patients and control subjects without cardiovascular disease, accounting for QT type, ongoing treatments, and time relative to major clinical events (e.g., Torsades de Pointes). Patients with AI scores \>50% without prior diagnosis will undergo genetic testing, which serves as the reference standard to confirm congenital long QT.
Diagnostic performance of AI score for drug-induced long QT	Day 0	Evaluate sensitivity, specificity, positive predictive value, and negative predictive value with 95% confidence intervals for an AI-generated score (0 to 100%) estimating the probability of drug-induced long QT, defined by a QTc increase of more than 60 ms from baseline or drug-induced Torsades de Pointes, compared to control subjects without cardiovascular disease.
Accuracy of AI-derived quantitative ECG measurements	Day 0	Compare AI-generated measurements of QT, QTc, heart rate, PR interval, and QRS duration with values obtained by reference standard methods at CIC-Paris Est.
Evaluation of standardized feature importance profile (FIP) for ECG segment discrimination	Day 0	Evaluate the standardized feature importance profile (FIP) derived from the deep-learning model to identify ECG segments most discriminative among different patient groups.

Trial Locations

Locations (1)

Centre d'Investigation Clinique Paris-Est/Hôpital Pitié-Salpêtrière; 🇫🇷 Paris, France