Identification of patients at risk of torsade de pointes, a life-threatening arrhythmia, using ECGs and deep learning – DeepECG4U
Identification of patients at risk of Torsade-de-Pointes, a life threatening arrhythmia using ECG and deep learning
Some cardiovascular diseases (congenital long QT syndrome, cLQTS) or drug-induced long QT syndrome (diLQTS), can cause Torsade de Pointes (TdP), a life-threatening arrhythmia. On the ECG, the QT interval is prolonged, but the waveforms have specificities, thus being poorly predictive of TdP. New predictive approaches are needed to improve physician assessment and reduce the risk of TdP events.
This project aims to develop a personalized patient TdP risk prediction tool using artificial intelligence.
Automated personalised prediction of the risk of TdP in cLQTS or diLQTS patients, can improve the accuracy of the physician's assessment and reduce the risk of cardiac events. This project aims to develop such a tool using artificial intelligence, which is rapidly reaching medical practice. Deep learning in particular has brought about a radical change in the field of pattern recognition and machine learning itself, improving on most previous models such as image classification and natural language processing. Specifically, in cardiology, deep learning has recently been used for several applications, including the detection of various types of common arrhythmias such as atrial fibrillation, myocardial infarction, etc. However, its use in predicting TdP events in a drug and congenital context has not yet been explored. In this project, we use these algorithms to develop robust models to improve not only prediction accuracy, but also to provide clinicians with new patient stratifications. Finally, we seek to improve the interpretability of these models and consequently the understanding of the molecular mechanisms underlying TdP.
The methodological research work is spread over several axes, namely i) TdP risk prediction, ii) robustness of deep models, iii) interpretability of models and predictions and iv) stratification of ECGs and patients. Within the consortium, we have explored different types of deep neural networks in the context of Torsade de Pointes risk prediction. We have successfully used convolutional neural networks (CNNs) as well as auto-encoders (VAEs) and recurrent neural networks (RNNs). Among the architectures that are best suited to the problems introduced above are Unet-type networks. The approaches explored to increase the robustness of the models consist in filtering the quality signal and training the models with noisy data as well as the use of more varied data. Finally, among the methodological developments to make the models more interpretable, are the introduction of prototypes, the use of occlusion algorithms and the application of constraints on the models while training.
The preliminary results of these different axes are very encouraging and publications are in progress. We have first developed a tool called «ecgtizer«, which allows to convert paper ECG data into XML format usable by deep learning models. This allows increasing the size of the datasets used and improving the performance of the models. In the context of the robustness of deep models, we have developed powerful models that allow to clean and segment the ECG signal and to evaluate its quality. This information helps us to perform a more robust and efficient learning process of the TdP risk prediction models. Finally, we have worked on a new approach based on the notion of «concepts« and genetic algorithms that allows us to improve greatly the interpretability of the predictions made.
The goal of the project is to advance the methodological research topic and create a translational application deployed in several pilot cardiology departments. Key outcomes will include (i) integrated data repository, (ii) deep network models predicting TdP, (iii) improved patient stratification and model interpretability, and (iv) clinical application. These results will then need to be validated in a dedicated clinical project.
Prifti E, Fall A, Davogustto G, Pulini A, Denjoy I, Funck-Brentano C, Khan Y, Durand-Salmon A, Badilini F, Wells QS, Leenhardt A, Zucker JD, Roden DM, Extramiana F, Salem JE. Deep learning analysis of electrocardiogram for risk prediction of drug-induced arrhythmias and diagnosis of long QT syndrome. Eur Heart J. 2021 Oct 7;42(38):3948-3961. doi: 10.1093/eurheartj/ehab588. PMID: 34468739.
Schwartz PJ, Tan HL. Long QT syndrome, artificial intelligence, and common sense. Eur Heart J. 2021 Oct 7;42(38):3962-3964. doi: 10.1093/eurheartj/ehab611. PMID: 34508622.
Vulgarisation article in French: « Améliorer la prévention des morts subites associées à la prise de médicaments grâce à l'intelligence artificielle » en.ird.fr/node/10536
Vulgarisation article in French « DeepECG4U : l’intelligence artificielle au service de la santé cardiaque » www.ird.fr/deepecg4u-lintelligence-artificielle-au-service-de-la-sante-cardiaque
Edi Prifti, Fabrice Extramiana. Intelligence Artificielle appliqué à l’ECG, 8ème édition des journées annuelles de la filière nationale Cardiogen réservées aux professionnels de santé. Bordeaux 31/03/2022
Joe-Elie Salem. Sexual dimorphism, anti-hormones and cardiac arrhythmias: a focus on QT. Copenhagen Meeting on Cardiac Arrhythmia. September 2022. Denmark
Joe-Elie Salem. Anticancer drug induced atrial and ventricular arrhythmias. European Society of Cardiology 2022. Barcelona. Spain
Some cardiovascular diseases (such as congenital long QT syndrome, cLQTS) or drug-induced long QT syndrome (diLQTS), can cause a particular form of ventricular arrhythmia called Torsade de Pointes (TdP). While often self-terminating, TdP can degenerate leading to death. There are three main forms of cLQTS: type 1, caused by mutations in cardiac channel genes leading to IKs current blockade; type 2, IKr blockade, and type 3, INaL activation. On Electrocardiogram (ECG), QT is prolonged in all these latter conditions, but ECG waveforms carry specificities including T-wave morphology abnormalities specific of each type of cLQTS. Most drugs responsible for diLQTS and eventually TdP, can be identified by assessing any of these mechanisms on the ECG. Therefore, diLQTS and cLQTS type 2 carry similarities in their ECG footprints. Regulatory agencies require new drugs to undergo thorough QT studies. It is however established that limiting the ECG evaluation to QT measurements is poorly predictive of TdP. Prediction of TdP risk and the characterization of the molecular mechanisms involved is of major interest for patients suspected to carry cLQTS as it is for other patients receiving drugs that may cause TdP. This is also a major issue for the pharmaceutical industry when developing new drugs. Finally, most physicians prescribing these drugs are unable to correctly quantify QT or evaluate TdP risk and do are not able to immediately consult with expert cardiologists.
Automatized personalized prediction for TdP risk of cLQTS or diLQTS patients, who may be unknown, can improve the accuracy of physician assessment and lower the risk of adverse events. In this project, we aim to develop such a user-friendly tool using artificial intelligence, which is rapidly reaching medical practice. Deep learning (DL) in particular has brought a radical change in the field of pattern recognition and machine learning (ML) itself, improving most of the earlier models devoted to learning tasks such as image classification and natural language processing. Specifically, in cardiology, DL has recently been used for several applications, including the detection of various types of common arrhythmia such as atrial fibrillation, myocardial infarction, and cardiac contractile dysfunction. However, the use of DL in predicting TdP events in a drug-induced and congenital context has not yet been explored. In this project, we will use such algorithms that provide models that will not only increase precision but also provide clinicians with interpretable novel features and representations that could improve patient stratification. We expect that ECG annotation with the signal features that most influence the prediction would improve the understanding of the molecular mechanisms underlying TdP. Within the consortium, we have already explored the DL hypothesis and our preliminary results are very encouraging. The objective of our 3.5 year-long project is to advance this research topic, transform it into a translational application in several pilot cardiology departments during a first phase and design validation clinical trials during a second phase for widespread use in and out of hospital settings.
This project will involve 5 teams. Teams 1, 4 are experts on AI and Interpretability, and Teams 2,3,5 are world-class experts in cardiology (cLQTS, diLQTS, and TdP). Moreover, they come with unique valuable datasets. The project will last 42 months. Key results will include (i) an integrated data repository (ii) DL models that predict TdP (iii) patient stratification and interpretability and (iv) a clinical application. The project is in full coherence with the government’s and the ANR’s objectives in accelerating AI-based translational applications in medicine and will most likely strengthen the position of France in the international arena.
Monsieur Edi Prifti (Unité de modélisation mathématique et informatique des systèmes complexes)
The author of this summary is the project coordinator, who is responsible for the content of this summary. The ANR declines any responsibility as for its contents.
IBISC Informatique, BioInformatique, Systèmes Complexes
CIC-1901 CIC PARIS-EST (Centre d'investigation clinique)
UMMISCO Unité de modélisation mathématique et informatique des systèmes complexes
UMR ICAN Unité de recherche sur les maladies cardiovasculaires, du métabolisme et de la nutrition
Roden's Lab Vanderbilt University Medical Center / Roden's Lab
Help of the ANR 570,517 euros
Beginning and duration of the scientific project: March 2021 - 42 Months