“The emergence of artificial intelligence is a chance that cardiologists must seize if we want to preserve the quality of our exercise and respond to the demand for care that will not stop growing”, underlines Dr. Thierry Garban, member of the steering committee of the “Artificial Intelligence” circle of the French Cardiology Society. Indeed, cardiovascular diseases are increasing, cardiological demography in tension, and consequently the time available per patient is reduced. “The AI can release medical time, secure medical responses and improve the conditions of exercise of cardiologists”, underlines the cardiologist.
Applications in imaging and rhythmology
The AI has reached a particularly high level of maturity in the field of medical imaging, especially in the analysis of the cutting images (cardiac scanner, MRI, 3D echocardiography). These advances mainly concern the automation of segmentation, functional evaluation (ejection fraction, ventricular volumes), the detection of structural or ischemic pathologies and, more recently, prognostic stratification.
At the same time, in cardiac electrophysiology, progress is notable, especially in rhythmology. The AI allows today, from an electrocardiogram (ECG) in a sinus rhythm, to identify patients at high risk of developing atrial fibrillation (FA). Several studies have shown that networks of deep neurons (deep learning) could detect electrophysiological signatures invisible to the naked eye of the clinician, predictive of a future occurrence of FA.
Mainly approaches to the state of evidence of the concept
However, these approaches are still mainly in the concept of proof of concept and are not integrated into clinical recommendations, or in daily practice. The same goes for prediction, from a resting ECG, the risk of acute coronary events (acute coronary syndrome type), which is the subject of research, promising but also exploratory.
These AI devices must be differentiated from ECG reading software for reading ECG, based on decision -making rules or trees, which work well for simple anomalies (BBD, BBG, atrial hypertrophy, conduction anomalies, sometimes early repolarization syndrome) but are frequently taken in complex diagnoses.
Automatic or deep learning
But when we talk about AI, what are we talking about? According to the World Health Organization, it can be defined as “A branch of IT, statistics and engineering that uses algorithms or models to perform tasks and adopt behavior such as learning, decision -making and prediction.” In a way, a machine capable of performing human beings tasks, which is not yet the case today in cardiology.
The term encompasses the field of « machine learning » (automatic learning), which designates algorithms whose performance improves as they are exposed to more data, and that of the « deep learning » (deep learning), subset of « machine learning », in which multilayer neural networks learn from a large amount of data.
Three types of data
Whatever they are, all the algorithms used in the medical field are based on three types of data: data from practice; artificial data, entirely from algorithms; and synthetic data, associating the two.
Currently, more than half of the medical devices having benefited from an FDA marking (Food and Drug Administration) and which embark on AI in the broad sense are based on data of artificial or synthetic origin, which raises the question of the absence of clinical validation. “Doctors must become aware of the essential nature of securing their patients and the need for compliance with French and European regulations as to their secure storage, respecting the” IA Act “, which will come into force soon”, specifies Dr Garban. Thus, when they use diagnostic assistance tools proposed by different firms, they must learn about the origin of the data (real life, synthetic or artificial) and on the clinical validation of the system. As in clinical trials, there can be selection biases: not all models are generalizable. However, increased recovery generation systems (RAG) make it possible to provide more targeted data based on data, with a lower risk of “hallucinations” (false or misleading responses), therefore more suitable for medical practice.
Interview with Dr. Thierry Garban, secretary general of the National Union of Cardiologists
