00:00:00
Voice off: "100 jours pour réussir" is the G_NIUS podcast. Voice-over: "100 days to success" is the podcast from G_NIUS, the Guichet national de l'innovation et des usages en e-santé. Around Lionel Reichardt, meet e-health innovators and key experts to help you succeed in your projects.

00:00:19
Mr. Reichardt: Hello everyone and welcome to the podcast "100 Days to Success", the podcast for innovators and entrepreneurs in digital healthcare, but also for anyone curious about this field. This podcast is produced by G_NIUS, the Guichet national de l'innovation et des usages en e-santé. In this episode, we talk about integrating artificial intelligence into an innovative healthcare solution. To do so, we welcome Alexandre Guenoun, co-founder and CEO of Kiro, a digital, visual and interactive platform that improves the communication of medical biology reports between laboratories, healthcare professionals and patients.

00:00:53
Mr Reichardt: We're also joined by Olivier Clatz, an expert in artificial intelligence and director of the Grand défi "improving medical diagnostics through artificial intelligence", launched by the government. Hello, Alexandre Guenoun. Thank you for sharing your experience with us. Can you tell us a little about your training and background?

00:01:19
M. Guenoun: I graduated from Essec in Paris, en grande école, where I specialized in data science and strategy, and from the University of California, Berkeley, where I followed a dual path in innovation and law. I also studied innovation and the economics of new technologies at Bocconi University in Milan. I worked in private equity at L Catterton, part of the LVMH group, and in strategy consulting at Oliver Wyman in Paris. I started working in artificial intelligence in 2016 for an assignment with Etalab, a government administration, where the aim was to analyze the relevance of reimbursing certain drugs with machine learning.

00:01:55
M. Guenoun: I learned a lot during this experience, and I saw the positive impact we could have with new technologies in the care pathway. It really convinced me that there were things to be done in this field, and that if we wanted to win this game, we had to get down on the ground and not just keep our eyes riveted on the scoreboard. That's what made me decide to set up Kiro a few years later.

00:02:13
Mr. Reichardt: Your company Kiro was founded in 2019. Can you introduce it to us?

00:02:18
M. Guenoun: Yes. At Kiro, we're developing a software platform that makes it possible to render medical biology results, such as, for example, the Covid tests we know today. The aim is to make these results more understandable for patients, and more relevant for healthcare professionals, thanks to artificial intelligence. In concrete terms, we offer patients and healthcare professionals an interface that is interactive, secure, and on which we have completely reworked the current medical biology report that we are all familiar with. At Kiro, once the biology results are obtained and validated by the biologist, we transform the paper report into a completely new experience for all users.

00:02:55
M. Guenoun: With our algorithms, we process the available information to structure it and propose content in real time that is adapted to the user and that we personalize according to each person's biological profile. Today, we're a winner of the i-Lab national innovation competition, and we're also an "investissement d'avenir", with the PIA 3 that we won with the Sud region and the Secrétariat général pour l'investissement.

00:03:17
Mr Reichardt: Artificial intelligence is at the heart of your solution. How would you define this artificial intelligence? What is it? And above all, what is it not?

00:03:26
M. Guenoun: By classical definition, we could say that AI is a set of computing techniques, often derived from mathematics and statistics, that generally enable machines to perform tasks, and possibly solve complex problems that are normally reserved for humans. That's the standard definition, but I think the main problem - and you put it very well in your question - is what it is not. The main problem with artificial intelligence is above all one of semantics. AI could be better defined by what it is not. Artificial intelligence is neither intelligent nor artificial. On the contrary, you could almost say that AI is a bit stupid.

00:04:10
M. Guenoun: In reality, it's really just a series of calculations. The word intelligence is not really adapted to reality here, and this is precisely what produces, today, a little too much fantasy. The only notable difference with humans is speed. The term that could perhaps come closest to artificial intelligence in reality would be "heuristic computing". The problem is that behind the term AI, which is a bit of a catch-all term, many different things are indicated. AI is not something new. The first AIs date back to the 50s, and there were already a lot of expert systems in the 70s and 80s. What is new today, however, is a whole new type of algorithm, which we find under the terms machine learning, and even more recently, deep learning.

00:04:58
M. Guenoun: It's really machine learning that has made a complete paradigm shift from the previous generation of AI that were expert systems. In other words, the aim of machine learning is not really to acquire already formalized knowledge, but rather to understand the structure of data and integrate it into models to automate tasks. When we talk about AI today, it's precisely these new learning systems that are becoming truly remarkable, enabling us to do things that were thought completely impossible just a short while ago, such as, for example, translating languages, recognizing speech, or, for healthcare, reading medical images.

00:05:34
Mr. Reichardt: We sometimes talk about several levels of artificial intelligence: strong artificial intelligence and weak artificial intelligence. Is this distinction relevant?

00:05:44
M. Guenoun: Today, in my opinion, this is not necessarily what is the real heart of the matter. AI is more focused on optimizing the resolution of a problem, more than augmenting a human being, and still far from the cyborg human, half-man, half-machine. The question that AI answers today is: is the value we produce efficient, optimal and appropriate? That's why, at Kiro, we like the idea of giving superpowers to people who are first and foremost human, or even everyday heroes. All we do is provide a tool, without replacing anyone else. When we talk about strong AI or weak AI, we're talking about whether, one day, artificial intelligence will be able to surpass human intelligence.

00:06:33
M. Guenoun: Today, I think we're still a bit far from it. To consider this dichotomy is sometimes a little too quick to absolve man of his intellectual and ethical responsibilities towards the machine. The real risk, when we talk about strong and weak AI - which is not a current reality - is that we attribute to the machine an intelligence that does not come from it; or even worse: that we take for intelligence what is not.

00:07:00
M. Reichardt: For a project leader, how do you integrate AI into your project? How do you work on AI with healthcare professionals and institutions?

00:07:10
M. Guenoun: Before choosing to integrate artificial intelligence into your project - it's a real choice, in my opinion - it's very important, beforehand, to make sure that what you're talking about is understood, and above all to check that the use of this technology is relevant and adapted to the problem you're trying to solve. These days, AI is often talked about as a fantasy. The fact remains that it's a technological tool like any other. I think it's very important to demystify this, before even trying to integrate artificial intelligence into your project. AI has already begun to transform the healthcare sector and other sectors in particular; there's no longer any real doubt about that.

00:07:53
Mr. Guenoun: With AI and data analysis, the transformation of the medical sector will enable - and is already enabling - more preventive, personalized medicine, at the service of patients, doctors and healthcare organizations. To integrate AI into your project, you need to understand the capacity of this technology in general - and AI in particular - to be relevant to helping individuals in their care and access to information as part of the care pathway. The first thing is to avoid exaggerated science fiction. Then, understand that it's technology at the service of the problem and the need. Once you've said that, there are several ways of integrating AI into your project.

00:08:33
M. Guenoun: First, for healthcare professionals, what's important is to get trained in these subjects, without becoming an expert in data science. It's about understanding what's at stake, and differentiating the concepts between expert systems, machine and deep learning, the notions of precision and recall to see performance, as well as the differences between training, test and cross-validation data. For facilities, the other challenge in integrating AI, especially when you're an entrepreneur and want to work with a healthcare facility, is to first build the structures that will enable artificial intelligence to be implemented, and in particular to have, as an underlying requirement, information systems that are scalable.

00:09:18
M. Guenoun: For that, we're talking about standardization and interoperability, which still need to be improved for many. Integrating AI involves this first and foremost. We can't ignore LOINC, FHIR or HL7. For us, it's something we're already doing upstream, with establishments and healthcare professionals, to integrate AI into their projects. For an entrepreneur, once it's been said and we've worked with healthcare professionals who understand what we're talking about, and we have relevant structures behind us, we have to start at the beginning. Starting at the beginning in AI means knowing how to define the problem. Today, far too many people still see technology, and AI in particular, as an end in itself. That's not necessarily the idea.

00:10:01
M. Guenoun: It's really just a means that can prove extremely relevant to solving certain concrete issues, and perhaps less interesting for others. The challenge is to clearly define, with healthcare professionals and users in the field, the why, the how, and above all, behind this, with artificial intelligence professionals, to be able to work to answer and ensure that AI can have the answer to this question.

00:10:24
Mr. Reichardt: Fairly, if AI can have the answer, how did this approach come about at Kiro?

00:10:30
Mr. Guenoun: We thought about this question for a long time. We realized that for the problems we're trying to solve, and in particular when it comes to medical biology, machine and deep learning, much more than artificial intelligence as such, are particularly well suited. Let me give you an example: when you have a medical biology check-up, on average twenty-five to forty individual parameters vary simultaneously, along several dimensions. In general, the dimensions of variation are the patient, the analysis itself, and time; and this, for each of the parameters of the workup. It's a very complex problem for the human brain to grasp, especially in a time frame that can be constrained.

00:11:10
M. Guenoun: On the other hand, it's much simpler to formalize for the machine, which can perform several operations in a fraction of a second. We realized that machine and deep learning could be very interesting for answering questions in medical biology, and in particular for helping healthcare professionals and patients with results. That's why we chose to use AI in the project, and it has proved extremely relevant in helping healthcare professionals and patients in the field.

00:11:39
Mr. Reichardt: How do you assess whether a company is really doing AI? It's a buzzword. What are the pitfalls to avoid if you want to work well in this field?

00:11:47
Mr. Guenoun: It's true that, recently, a study by the Financial Times - or reported by the Financial Times - indicated that there are forty percent of European startups that are called artificial intelligence, but which, in the end, don't use artificial intelligence at all. This is a real point, and we need to be very vigilant. There are lots of ways to tell if you're dealing with a company that really does AI. In my opinion, there are two main factors to watch out for. Paradoxically, the first is the human element, and the second is the company's or startup's approach to data. First of all, on the question of people: an AI company is first and foremost a company, so it's representative of the people who make it up.

00:12:27
M. Guenoun: There is no AI, in the sense of machine and deep learning, without a data scientist, and data scientists are not sorcerer's apprentices, they are engineers who have been trained with solid skills, in mathematics and statistics, to work with data. Working with data can't be invented. Today, eighty percent of a data scientist's work consists of having clean data - in the sense of "clean" -, and adapting the data for the analysis we want in relation to the problems we have defined. Only twenty percent of the time is actually spent on the analysis itself. This implies a need, more than a simple desire, to work with people who are qualified for this.

00:13:08
M. Guenoun: In the same way that data scientists are qualified to work with data, we need to work hand in hand with healthcare professionals in the field to do relevant things. AI and medicine are professions that are different, and it's only through this cooperation that we can understand each other's issues and respond to problems with an answer tailored to the questions that are essential for users in the field. Then there's the question of data beyond the human. We can't say it often enough: if the initial data is bad, artificial intelligence is useless. A company that uses AI, and that is relevant in this respect, works on the quality of its data. This is the concept of GIGO - garbage in, garbage out - according to which input data that is faulty or absurd will produce absurd outputs and results.

00:13:59
Mr. Guenoun: What you have to realize is that statistical analysis is always possible, even with incorrect data. Nothing will stop you from processing them, and nobody will tell you that what you're doing doesn't make sense; and that, you risk discovering too late if you do nothing; hence the importance of knowing the quality of the data upstream. The worst thing about working with AI is thinking that just because you have a lot of data per hour, you can turn it into super-algorithms and therefore super AI. Let's face it: ninety percent of existing data has been created in the last two years. This means that today, the quantity of data is probably no longer an issue. Those who are well prepared to work with AI are focusing on quality, rather than quantity.

00:14:46
M. Guenoun: International AI research competitions - such as Kaggle - offer datasets ranging from fifty thousand to one million. It's not uncommon to see scientific publications in robust AI based on a few tens of thousands of cases, so it all depends on the problem you're trying to solve. The quality and relevance of the data are determined on the basis of that problem, the data professionals and the understanding of the issues in the field. Hence the importance, in the healthcare field in particular - but not only - of interoperability and the use of standard nomenclatures to have homogeneous concepts. Without that, there's not much point in accumulating data without really understanding what's going on. I think that the human element on the one hand, and the quality of the data on the other, at least allow us to detect whether we're dealing with a company that's really doing AI.

00:15:35
Mr. Reichardt: Humans and data quality represent an investment. What is the return on investment of using artificial intelligence for a project owner?

00:15:43
M. Guenoun: It's really going to depend on the problem we've identified. Inevitably, in AI, there's a question of standardizing the response to that problem. Today, the question of cost is linked to the degree of complexity involved in standardizing the problem. It's true that the objective of AI is to respond to a problem; in the case of Kiro, it's the supply of care versus the demand for care. We want to respond to a high level of service that frees up medical time for healthcare professionals - this is important in the AI we're setting up - because the aim behind it is to provide better services and better, more accurate, safer and more effective care.

00:16:26
M. Guenoun: This is a luxury we want to offer our users, healthcare professionals and patients, as part of a more comprehensive service offering. Yes, it can be quite costly, but the idea is that it can also have a significant return on investment, because we often realize that the technical operational measures that would have to be implemented to be able to meet the same demand with the same level of quality are much greater; often by a ratio of one to ten. So, this is an important point if you want to consider the return on investment of artificial intelligence in a project.

00:17:03
Mr. Reichardt: Is artificial intelligence complicated to implement within a company? Does it represent a high cost?

00:17:09
M. Guenoun: There are different levels. A bit like in many fields, if you want to have a quality service, you need to call on data professionals - data scientists and software engineers - who have the skills to be able to meet these challenges. These are rare professions, and therefore inevitably expensive. Today, they are favored by most GAFAs, with salaries that are quite substantial. It's a project that starts from the principle that AI has to be integrated into it; it can be expensive, but the added value, in general, is there. Why is it expensive? Because it requires an expertise in understanding data that is quite unique. We're not talking about statistics: we're going beyond statistics, or biostatistics in healthcare, with real challenges in machine and deep learning.

00:18:04
M. Guenoun: They have in-depth mathematical knowledge to be able to do deep neural networks or diagnostic trees. I think it's also expensive for another reason: at the outset, there's a significant investment to be made. That's why, today, it's startups that are leading this kind of project, because there's also risk involved. They're the ones able to take the risk, thanks to investors who can make it possible. In any case, I think there are a number of steps that need to be taken, regardless of cost, to ensure the relevance of a project like this. The first thing is to define the problem, clarify the objectives, and validate that we have the database to respond to the right problems; that takes time.

00:18:46
M. Guenoun: The second thing is to ensure the quality of the data, define what's expected, and work with healthcare professionals in the field to evaluate data sources and sanitize the data we're going to retrieve. At the end of this phase, not that much good data remains. That's what can be so costly, the time spent slanting, fetching data in parallel, and completing datasets. The third point that also explains this rather high cost is the training of models, with all the biases and limitations that need to be kept in mind when running them. This is why a well-trained data scientist, with experience, will be able to see the biases and limitations better than a data scientist or someone who is not trained in these subjects.

00:19:27
M. Guenoun: So, in addition, we're going to have to have the models reviewed by peers and industry professionals; if possible the best, to have relevant algorithms. You can't gamble on quality; at least in healthcare and diagnostics in particular. Quality management systems have to be put in place, and all this comes at a cost. Since we are standardizing a large proportion of the needs and resources that could be realized, the ROI is inevitably there, when we take the trouble to look into it.

00:19:57
Mr. Reichardt: To conclude: what advice would you give to an entrepreneur who wants to integrate artificial intelligence into their project?

00:20:07
Mr. Guenoun: There are three things to keep in mind that we've talked about. First, it's the definition of the problem. Second, there's the question of quality. Finally, there's the human factor. We really mustn't forget the human element when we talk about artificial intelligence. On the question of the problem, the thing to keep in mind for the entrepreneurs listening to us is to understand that AI is not an end in itself, it's just a means to an end. It's very important to define upstream the problem for which you want to use AI. Today, the first mistake AI entrepreneurs make is to start with technology. At Kiro, machine and deep learning are extremely useful for the problems we're trying to solve, but you really have to start from the problem. Only if AI is relevant to answering it should it be used.

00:20:54
M. Guenoun: We all dream of doing super sexy things, but I think it's even more important to do useful things, especially in healthcare. There are many more uses in which we can integrate less ostentatious AI to make users' lives easier in the healthcare system, and which are really useful, if not more so than the magical AI we hear here and there. There's no secret to it: you have to work on the problem, understand the problem you're trying to solve, and to do that, you have to get as close as possible to the business and work with professionals in the field to understand the problems you're trying to solve, and respond to them. Then, and only then, is tech useful. So that's mistake number one. Don't start with the technology, start with the problem.

00:21:35
M. Guenoun: The second thing - this is public enemy number one in healthcare - is low-quality data. There are steps to follow - we've given them. We need to be sure that the data we have is relevant to the problem we're trying to solve. Finally, the third point is not to minimize the importance of people. In fact, it's a good thing that humans are very important when it comes to AI. That's why I find the concept of the human guarantee so rich - it's an opportunity we have in France -: we recognize that AI can't be perfect, because it's very human behind the scenes, but that, for all that, it can be useful to us in many ways. This is why we must always keep supervision and human intelligence - the intelligence that data scientists and healthcare professionals can have - over algorithms.

00:22:21
Mr. Guenoun: It's really more than necessary. For the entrepreneurs listening to us, it's important to understand that this is what will make the difference between a quality company and another. In healthcare, in particular, you can't cut corners on quality. It's fundamental. To conclude: after all, what could be more logical than for a company to be valued by the people who make it up? That's really what you need to remember when you want to integrate AI into your project. It's a quality issue at the service of health.

00:22:55
Mr. Reichardt: Thank you, Alexandre Guenoun, for your testimony.

00:23:03
Mr. Reichardt: Are you wondering how to integrate artificial intelligence into your project? Elements of an answer with Olivier Clatz, artificial intelligence expert and director of the Grand défi "amélioration des diagnostics médicaux par l'intelligence artificielle", launched by the government. Hello, Olivier Clatz. Thank you for answering our questions. Can you tell us about your background and training?

00:23:26
Mr Clatz: Hello. I'm an engineer who did a thesis in medical image processing, at INRIA. Following this thesis, I did a post-doc in Boston, in the United States. I came back in a research position, and was a researcher for six years on the same themes - medical image processing. Then I founded a company that I ran for six years, called Therapixel, which does mammography-based breast cancer screening. Finally, it's been just under two years that I've been leading the Grand défi "improving medical diagnostics through AI".

00:23:58
Mr. Reichardt: You're in charge of the Grand défi "improving medical diagnostics through artificial intelligence". Can you remind us of its main objectives?

00:24:06
Mr. Clatz: The Grand défi is structured around three axes. The first is to support projects that are maturing and don't yet have their algorithm completely defined, to provide them with the means to both develop their algorithm and acquire the data. We're working on two fronts. We launched an initial call for projects - for example, with the Health Data Hub - offering a package to project leaders, including R&D funding and support for data access. The second axis aims to support projects that are a little more mature and fund them to experiment and validate their value proposition.

00:24:44
M. Clatz: For us, The objective is to take on projects that are still at risk, that already have a technology that is mature - fixed, when we're talking about artificial intelligence algorithms -, but that hasn't yet proven itself with a sufficient number of patients or healthcare establishments. What we are funding, via calls for projects, is this phase of evaluation, which costs money; a cost that is often underestimated by project leaders. We finance the new patients who enter the study, the statistics, possibly an arc, or the engineers who can work in the back office. In short, everything that's needed to demonstrate the value of these algorithms in the field. The third axis is a little more systemic, and aims to facilitate data exchange within the care pathway.

00:25:38
M. Clatz: The aim is to accelerate, on the medical imaging theme that I know well, data exchanges, and in particular medical images. This is an area that began with the challenge, but which is now part of a much larger initiative, the Ségur de la santé and its digital side.

00:25:57
Mr. Reichardt: Is there a link between the Grand défi "improving medical diagnostics through artificial intelligence" and the Ségur de la santé?

00:26:04
M. Clatz: The first vocation of the Ségur de la santé is to finance the digitization of the healthcare pathway for patients and citizens. At first glance, there's not necessarily a direct link between artificial intelligence, algorithms, and the action we've launched through the Ségur, which targets the care pathway, i.e. when you go to see the doctor, when you go to the pharmacy, or when you go to the bio lab or the radio surgery. On the other hand, we're quite convinced that when this data is truly dematerialized in the care pathway, we'll then be able to deliver value much more easily on the basis of this data.

00:26:47
M. Clatz: The aim is to move from paper to digital. As soon as we manage to take this step, we're able to plug in algos much more easily, or even be able to fetch data that's much more easily accessible, because it's better organized, better interoperable, and accessible with secure means; which isn't the case when we work with paper. Even if, at first glance, there's no direct link, since Segur focuses more on routine care, and artificial intelligence focuses more on new technos, we think the two will come together fairly quickly.

00:27:19
Mr Reichardt: There's a lot of talk about artificial intelligence. Can you give us your definition of it?

00:27:24
Mr. Clatz: That's the two-million-dollar question. It's always asked of AI specialists. The idea is to make decisions based on decisions that have already been made in the past. The algorithms are given data that is known and for which the decision has already been made, whether it comes from the doctor or from additional tests. So we really know what happened to the patient. Then there's data of the same nature as that used for training, but for which we don't have the answer. In this case, the artificial intelligence algorithm will make the decision that is statistically closest to the dataset it was shown the first time. These algorithms aren't very smart: they reproduce decisions they've seen in the past.

00:28:19
Mr. Clatz: Today, what's fairly new is that we're achieving decision-making capabilities on data that are high-dimensional, i.e. we can grind datasets that are gigantic; millions, even hundreds of millions or billions of MRIs, electrocardiogram signals, or biology data. We weren't able to do that until not too long ago. There's one powerful feature of these algorithms. Before, we had to describe to the algorithm where the information was, i.e. we would tell it, in an image, that an anomaly might be a little whiter. For example, in the case of a pulmonary nodule, it might be a little spherical in shape, or conversely, with a texture known as "frosted glass". In short, we tried to qualify and explain to the algorithm what it was looking for.

00:29:12
M. Clatz: The big advantage these new algorithms have is that we no longer need to develop them in the form of descriptors; we just come and reindicate in the image where the anomaly is, and it's they who make their own representation of what they're looking for. These are the two breaks that have occurred recently. Firstly, we're dealing with much bigger data. Secondly, algorithms no longer need to explain and characterize anomalies very precisely. They're the ones who discover them themselves in the data.

00:29:43
Mr Reichardt: What can you say to a project owner who wants to integrate artificial intelligence into their solution?

00:29:50
Mr. Clatz: The best advice is not to ask the question vis-à-vis artificial intelligence. From my point of view, what will interest both the patient, but eventually the ministry when it comes to funding products, is the value it can bring for the healthcare system and for the patient. The first question to ask, before asking algorithmic questions, is: what value do I bring? The second question is: how can I create a business model around this value? Because sometimes you can create value, but it's hard to get it from the market. The idea is to know how I can monetize the value I can create for this patient or this healthcare system.

00:30:32
M. Clatz: After that, there are probably lots of questions that are pretty standard for entrepreneurs, like the quality of the team, i.e. knowing who I'm going to have in my company and who's going to accompany me on my adventure. Techno shouldn't be at the heart of the entrepreneur's preoccupations. Instead, they should be concerned with the classic questions of entrepreneurs - and questions that are much more specific to healthcare, such as: what value do I bring to my patient or my healthcare system?

00:31:08
Mr. Reichardt: Our episode is coming to an end. Thank you for listening to us. We thank our two guests for their availability. Feel free to subscribe to the podcast on the listening platforms. We look forward to seeing you very soon for a new episode of "100 days to success".

00:31:27
Voiceover: Those who are making e-health today and tomorrow are on the G_NIUS podcast. All the solutions for success are on gnius.esanté.gouv.fr.