A physician’s time should be spent diagnosing and treating patients rather than spending hours pouring over a patient’s test data.
While medical tests like electrocardiograms (or ECGs) are easy to record, they can be complex and time-consuming to interpret due to the many patterns and abnormalities that can be present in the data—and physicians have to remember and recognise each one of them.
But what if artificial intelligence could be leveraged to do most of the heavy lifting instead? Cardiologs, a French startup, has built a cardiology analysis platform powered by artificial intelligence to help healthcare professionals detect heart abnormalities from ECG data quicker and more effectively.
In this episode of HealthRedesigned, we chat with Chiara Scabellone, a Product Manager at Cardiologs who tells us how the company plans to create intelligent cardiology tools to help physicians manage cardiac patients at scale and make high-quality cardiology accessible to all.
Intelligently decoding ECGs
What are Electrocardiograms?
ECGs are the representation of the electrical activity of the heart. They’re very routine examinations and are the starting point for the diagnosis of any kind of cardiovascular disease. They're very easy to record but very hard to interpret.
And why are they difficult to interpret?
When a resting ECG is taken, for example, over 120 different abnormalities can usually be spotted. That's a huge amount of patterns to remember for any person, and if a doctor hasn't been doing this daily, he or she will probably recognise only up to 20 different abnormalities and possibly miss the rest. That's why it's a particularly complicated exam to diagnose.
How is Cardiologs helping to improve this?
Over the span of his or her career, a doctor would have looked at a number of ECGs and gained experience in recognising patterns and abnormalities in the signal, applying a mixture of pattern recognition and memory. We tried to use the same approach when developing the algorithm. What artificial intelligence (AI) brings to the table is the ability to replicate this intuitive manner that people use to recognise patterns.
We trained the algorithm using a database of over 500,000 different ECGs and developed a deep neural network that’s learned to spot patterns in the signal in the same intuitive manner as a human expert would. Commercially available solutions today use rules-based algorithms, but there are so many different abnormalities and combinations of abnormalities, that what they end up with is a huge decision tree.
Since they’re trying to avoid missing anything in their detection, they tend to be oversensitive. This raises a lot of false alarms in their diagnosis and results in a vastly decreased confidence in what the algorithm can output.
Cardiologs is different because we’re using an approach that tries to replicate the way a human does this, to overcome the mistrust.
Our mission is to enable tomorrow’s medical decision making to be as effective and accessible to all. We’re trying to provide healthcare physicians and practitioners with the right tools to better manage cardiac patients at scale.
Bringing better tech to cardiology
How did Cardiologs begin?
Back in 2014, our co-founders, Yann Fleureau and Jia Li came across ECGs and were really surprised by how such an easy and common exam could lead to so many missed diagnoses and how there wasn’t an algorithm available to convince doctors of its output.
To test their idea of using artificial intelligence to interpret ECGs and to begin training the algorithm, they began gathering ECGs and took about a year to gather 100,000 tests, mostly through clinical relationships in the US. That allowed them to validate the first methods and metrics before raising the first round of funding and getting regulatory approval.
Cardiologs now has clients worldwide, especially in Europe and the US. Since our product is a medical device, we are regulated and follow the CE Marking in Europe and have obtained FDA clearance in the US.
What is Cardiologs’ value proposition for health practitioners?
The main value proposition that we offer with our ambulatory solutions and web platform is for doctors to save time. Today, it takes an average of 20 minutes to analyse an abnormal ECG. That can go up to an hour and a half for extended examinations. That’s mainly because the software that’s in use today is built by the hardware vendors to support their offering and there hasn't been much innovation in recent years.
What we offer is a solution that allows them to bring down the analysis time to between five and 10 minutes. Software that’s used today hasn’t been designed with a clear user journey in mind and so we’re bringing a focus to the software solution and the user flow.
By maximising the output from the algorithm as much as possible, we’re able to optimise and streamline the analysis flow.
What has been some of the reactions of doctors who are using it?
The AI output and its presence is definitely the biggest challenge and biggest benefit of our solution. We’ve tried using the AI in a way that isn't intrusive and scary by balancing the amount of optimisation the AI does while giving the physician all the context he or she needs to reach a diagnosis.
The reaction has usually been around the experience itself and the most common one is that the platform is very intuitive and simple. The AI isn't something that’s particularly asked about because the doctor is ultimately interested in the patient outcome. So no matter what tech you use, if you can allow doctors to be comfortable with their diagnosis and speed up their analysis time, that's all that’s important.
Designing medical software with the physician in mind
What were some of the design problems your team encountered when working with artificial intelligence?
The biggest challenge in the design of a platform that is fully based on artificial intelligence, and something that I wasn't accustomed to before, is how visible you make the artificial intelligence. What do you want that to be and what do you want your user to think of the AI? So, the balance of automating the flow as much as possible, making it very sleek and simple and proposing episodes or making automatic reports.
At the same time, you really want to empower the doctor, physician or whoever is doing the analysis to have full control of the diagnosis. So, the extent to which you make the AI present and the balance between validating something that the AI proposes and enabling the person, for example, to do it manually, that's a very fine line, especially when your overall goal is to decrease analysis time.
What’s been your biggest learning along the way?
What I really enjoy at Cardiologs is the fusion of all these different areas of expertise in cardiology, deep learning, engineering, cloud computing and design. Being able to orchestrate the development cycles and different timelines for different teams is something that’s very challenging, but it's also very interesting and we're continuously working on it and learning from it.
What do you hope for the future of healthcare?
We want to make cardiology accessible and equal for all, which aligns with our mission to make high-quality cardiology really accessible by providing the right tools. We do so by trying to create the future of the interaction between the AI and the doctor.
What are you most proud of when you think of what Cardiologs has accomplished?
There's often a certain scepticism when we demo the product because this problem has been around for a really long time and people who have been in this space have seen many companies and software come and go while claiming to solve ECG analysis problem.
But when you actually get someone to try Cardiologs, they’re often delighted because the AI actually works—it recognises an atrial fibrillation, for example, or it spots a signal where the person was expecting it to find noise or rubbish, and that's nice to see. Even at the early stages of development when we had a rough prototype, you could see there was ground and the fundamentals were there to build upon.
Cardiologs loves working with physicians and learning about their problems to continuously validate their solution. If anyone is interested in collaborating with them on design, testing or for clinical research, do get in touch. They’re also currently looking for a product designer. For more information about the position, check out their website.