Lessons from CardioLogs, the French AI Startup disrupting Cardiology: from Data Acquisition to Business Model & Value Proposition.

Amine Korchi MD
Apr 19, 2018 · 6 min read
Source: https://cardiologs.com/

I listened carefully to Yann Fleureau’s speech during the DATADRIVENPARIS event about his 4 past years as a Co-Founder and CEO of CardioLogs and his journey towards building and selling an AI-based Clinical Decision Support System (CDSS) for Clinicians in the Cardiology space.

CardioLogs is a Paris-based Startup building Deep-Learning Algorithms for ECG (EKG) analysis. They have raised approximately 10M$ to date and have won approval for commercialization in Europe of the first medical grade deep-learning technology in 2016 and the second in the US in 2017. CardioLogs was founded in 2014 by Yann Fleureau, the current CEO. Yann is a graduate from the prestigious Polytechnic School of Paris (X) and passionate about New Technology & Medicine (https://cardiologs.com/).

You can find the whole presentation of Yann Fleureau on YouTube :

CardioLogs is a French Startup based in Paris (source: www.crunchbase.com)

The last 4 years of CardioLogs illustrate well the challenges of implementing an AI-based solution in clinical practice. I decided to write this short article because their experience reflects and resonates perfectly with what I have experienced as a Medical Doctor, Strategic Advisor and HealthTech Entrepreneur during the last decade and especially during the last 3 years with the explosion of the number of AI startups in Medical Imaging and Healthcare.

Here are some “bullet” points unveiling some crucial aspects of building and offering on the market a AI-based Clinical Decision Support System in the Real World:

- Getting access to Data is hard but is definitely not the most difficult task.

- Enriching, Annotating and Labeling the Data represent the most complex task according to my experience. By opposition to the everyday life images which almost everyone can annotate (car, bus, cats and dogs), medical images and data need domain-specific expertise; not only the expertise of a certified medical doctor, or even a board certified radiologist, but the input of fellowship trained and specialized radiologists and cardiologists. These expert annotators are scarce, struggle to handle their own workload in the real life and are not used to collaborate with technology startups and larger tech companies. So how can we replicate the model of MightyAI and Spare5 (https://app.spare5.com/fives/sign_in) where every human who knows how to use a smart phone can label data and contribute to the development of AI. One initiative is trying to allow citizens to become “scientists” and annotate microscopic images of cells; Etch-a-Cell allows “Citizen Scientists” to learn how to annotate Electron Microscopy images and leverage their labeling power at scale. (see illustration below) So far they have enrolled 4 513 volunteers and made 91 250 classifications including 4 597 subjects (https://www.zooniverse.org/projects/h-spiers/etch-a-cell).

From my side and in the Radiology space, I am currently involved in a Zurich-based Venture which dedicates all his time and money to solve this problem for medical images.

Etch a cell: Image from the presentation of Lucy Collinson (Francis Crick Institute) during WiredHealth 2018 in London. We see here that a “citizen scientist” is able to annotate a cell microscopy image as good as an expert.

- Data are coming from different Silos, different countries and clinical cultures, heterogeneous in quality with a staggering lack of standardization. The diversity of data is a good point for bias reduction, but a pain point for preprocessing and building consistent ground-truth.

- When building ground-truth, even human medical experts do not agree on the correct diagnosis. This was one of my biggest frustrations when I was a Radiology and Neuroradiology Resident. Looking for “my ground-truth” as a young medical doctor in training was a challenge and a pain. For the same clinical case, I was frequently confronted to 3 different diagnosis from 3 different Experts Privatdozent and Professors…. And I have been told and reminded countless time that asking for a diagnosis 3 different persons for the same cases lacks collegiality and is not appropriate… I would answer today that it was not my purpose to lack collegiality or to challenge anyone, my sole purpose was to define my ground-truth... Back to CardioLogs, 1/3 of its dataset did not reached consensus from 3 different expert cardiologists…

- IP in the field of Algorithmic Clinical Decision Support remains a gray zone. CardioLogs patents of Deep-Learning models are still pending according to Yann Fleureau, and they used a singular approach to protect their intellectual assets. They did not submit an application to patent only the Neural Networks architecture, but they are looking to patenting the end-to-end CardioLogs approach from EKG to Diagnosis for the Doctors. This holistic approach of Algorithmic Decision Support experience includes also patents for Human Interactions between clinicians and algorithms, the way to display the EKG and the user workflow. So we are now in a different Era comparing to the first patents in the pharmaceutical industry claiming novel active chemical structures or novel processes to manufacture medications.

As Yann Fleureau said in his presentation, the point here in not AI, the point is about the usage. This reflects also at a higher level on the value on an AI-based CDSS; its value is not only based on the core technology of deep neural networks, but especially on its usability in clinical real life and the benefit/risk ratio for the patient.

- Regarding Regulatory Approval, he highlighted that Transparency is key; « Do what you Say, and Say what you Do » regardless of the performance of your algorithms.

- Last but not least, I completely share the point of vue of Yann Fleureau regarding the Value Proposition and Business Model issues in this area. It is not because your algorithm detects heart attacks with the best accuracy or spots lung cancer on CT scans with unparalleled sensitivity and specificity, that you will systematically convert this performance into clinical adoption, better outcome and a commercial success.

In the field of Medical Imaging and AI and reflecting on my last years as a Medical and Strategic Advisors for HealthTech companies operating in this field, I do not see many companies generating strong and sustainable revenue based solely on their AI-backed Computer Aided Detection/Diagnosis Algorithms… it is a bit deceiving for the entrepreneurial optimist that I am… Early in my journey in the Healthtech industry, I was disappointed how the market was not welcoming breakthrough technologies, and with time I understood that precision and « sexyness » of an Algorithm are pointless if it does not translate to scalability, improved patient outcomes and integration into existing clinical workflows. Not talking about the reimbursement models, that’s for another blog post. We are in a complex ecosystem with multiple stakeholders where the Customer is usually not the Payor; drivers and incentives are different. CardioLogs shifted their value proposition from “Precision only” to “Speed & Scale” to improve patient outcomes and allows now massive screening programs for Heart Arythmias. How do we monetize our Value as a HealthTech Company is a hot topic and unsolved problem. But that’s what makes the game challenging and exciting. Not only we should think Technological Excellence, Patient Safety & Better Outcomes, but we should also find our place in a complex multi-stakeholders ecosystem.

Stay tuned for next publications; I will continue on my article series about how AI can transform Medical Imaging in a Realistic Scenario.

Feel free to reach out if my vision and activities resonates with yours !

Amine Korchi MD

Written by

Medical Doctor and Swiss Board Certified Radiologist & Neuroradiologist. Venture Partner at Fusion. Life Sciences, HealthTech & AI.