Pushing the boundaries of machine learning

To date, nearly 6,000 patients worldwide suffering from cardiac arrhythmia have undergone procedures lasting two hours instead of the average of five, with a 50% higher success rate. This significant progress is down to Cardiologie Numérique Personnalisée, a team made up of three Inria researchers (one of whom is now an entrepreneur), and two doctors (see inset).

Our story goes back to 1996, when the team set out with their initial aim to develop a digital model of the heart. “We were a long way away from medical decision-support in surgery. The aim was to get a clearer idea of how much blood is pumped by the heart with each beat”, recalls Hervé Delingette, Inria researcher and coordinator of a research initiative on this subject. Since then this initial work has been continually expanded upon. A major breakthrough came when the researchers were successfully able to model the biomechanics and the electrical circuitry of the heart, before bringing these two physical phenomena together. More features were then added to the model over time, including the measurement of blood flow and the supply of oxygen and energy (ATP) to the myocardium.

The project moved up a gear in 2008 when the researchers decided to integrate conditions such as arrhythmia and heart failure into their model of a “healthy” heart. The idea was that this model could be used as a decision-support system for various clinical applications, including to provide guidance for electrophysiology procedures linked to heart rhythm disorders such as ventricular fibrillation, which is responsible for 40,000 deaths each year in France. “These types of operation can take several hours of near-blind exploration using a catheter to look for the areas affected by the condition”, explains Hervé Delingette, “before the cardiologist can then cauterise the diseased tissue.” 

This latest ambition came about following discussions with the two hospital doctors and researchers from Bordeaux (see inset) who had recently joined the team.

These doctors set out with the aim of enriching Inria’s model using data specific to each patient (from electrocardiograms, medical imaging, and so on) in order to create a digital twin: a computer copy of the heart that is to be operated on. Displayed on a large screen for the surgeon to view, this digital twin would show which areas are diseased, providing real-time confirmation that each cauterisation has been successfully carried out.

It would take the best part of ten years to reach this goal. The team began by combining patient data with their model of the heart, before automating this process in order to be able to treat thousands of patients. They then worked on a detailed analysis of key parameters such as the cardiac conduction system, and further improved their model using artificial intelligence.

But what was the most important stage? “The major milestone in our project came with the launch of the startup inHEART in 2017 to produce and market our technology”, says Hervé Delingette. “That was the culmination of years of hard work, proving that academic and clinical research are complementary and can bring about medical breakthroughs.”

The award of this Inria prize is ample recognition for what is a unique project, having spanned more than 25 years:

Starting out with a very simple model, we managed to develop a solution with the capacity to change the lives of thousands of patients. Things have come full circle, and this Inria prize is a huge source of satisfaction for us.

In silico testing - a new frontier for this digital heart

The startup inHEART now has no fewer than 50 partner hospitals across Europe and the USA, and they don’t intend to stop there: over the past few years Cardiologie Numérique Personnalisée has turned its attention to the in silico planning (i.e. using a computer) of other surgical procedures on real-life patients. “Let’s take the fitting of pacemakers, for example”, says Hervé Delingette. “A decision has to be made as to whether or not this will actually be beneficial for the patient, as well as on the location and the optimal settings. Having digital twins of patients will help clinicians when making these choices.” It looks as though this “small” team might well have some more big results in store for us.