The Role of AI in Clinical Development

[00:00:00] Alexander: Welcome to another episode of The Effective Statistician. Today, I’m super excited to speak with Francois about AI. Francois has been in the industry for a very, very long time, has seen many, many different things and has [00:00:20] built Some amazing things. So who can better introduce France rather than France by himself. So over to you, Francois. 

[00:00:28] Francois: Yeah. Thank you, Alexander. So very nice to be here with you. So just so that people know about me a little bit more, I am a statistician by training. I have a PhD in statistics. I’ve been in the industry [00:00:40] for almost 30 years now. Work, started my career as a clinical statistician in a big pharma company for more than 12 years.

[00:00:49] And then there was some reorganization as there always is in this kind of company. And I decided to leave and to set up my own business, Clean Bay, which is a consulting [00:01:00] CRO in statistics. And, and the company has been operating for 17 years now. So we, we primarily work on, on, on statistics analysis and reporting of clinical research.

[00:01:13] And I had last year, well, no, sorry before the covid I wanted to expand a [00:01:20] little bit into the ai area and the machine learning area. But then they created another business at that time, but then the COVID could not really do anything. So I delayed the investment by two years and last year I set up a spin out of this clean big company called produce.

[00:01:39] That [00:01:40] is dealing with some AI and machine learning developments. And I wanted it to be both for pharma and healthcare, as well as in other areas like hospitality and, and things like that. But the company did not really succeed as a standalone company. I, I, we had, we weren’t able to complete the [00:02:00] developments in the other sectors than healthcare.

[00:02:03] Take care. So most recently I decided to terminate that business and then to focus on myself, on clean bay and transition all of the AI and machine learning assets we had produced to clean bay for, for, for the farm out [00:02:20] services. 

[00:02:20] Alexander: Yeah. And it’s quite an achievement to establish a company and then run it for 17 years.

[00:02:28] Well, most companies, you know. don’t survive the first year, let alone more than 10 years, let alone more than 17 years. 

[00:02:37] Francois: So, well, you know, these kinds of [00:02:40] companies they usually either get bought by bigger ones. This is not what I want. I don’t want to kind of get an acquisition by anyone else. And to survive in such a company, you need to have a good team of people.

[00:02:56] And so I spent the early days in my career [00:03:00] as a CEO of that startup company trying to, to sell the company, to sell myself and to set up a group of people around me. And it took me a while. And one of the Customer I had with which was basically one of the the founder of Scytel Nitin Patel told me once [00:03:20] you’re gonna have a business once the business is going to be able to work without you and and it’s really what I try to do, for Sometime it took me about maybe 10 years to, to achieve a level in the company where the company is kind of self-sufficient.

[00:03:35] And, and I’m, I’m not a CEO of the company anymore. We, I, I [00:03:40] created in, in 2019 and, and there is a leadership team in that company. It’s running all of the operations. We do have meetings every quarter when I basically review things. But I do not really put my hands anymore into the day to day activities.

[00:03:55] And, and the people over there, they really feel empowered into their [00:04:00] job. And they like having a perspective of having a long term relationship with their clients, with me, within the company. And, and I think this is It’s one of the achievement that I’m probably the most proud of is to be able to have a team like that, that can really run the things without me [00:04:20] being too much involved.

[00:04:20] Alexander: Awesome. Exactly. That is leadership. You help others to become leaders themselves. So you said they don’t need you anymore so much. 

[00:04:32] Francois: Okay. Maybe. I don’t know. But that’s, yeah. So that was one of the hints that I received from that person, which is [00:04:40] try to have your company work on its own without you being too much involved, and then it’s going to be a success.

[00:04:46] And, well, I don’t pretend it’s a success, but at least it has been running like that for a couple of years. So far, so good. 

[00:04:53] Alexander: I would definitely think that’s a success. If you just think about Simon Sinek [00:05:00] and the infinite game. Yeah. There is no judgment about who wins this game. Who wins this game is who stays in the game and what your company is doing.

[00:05:14] So that’s awesome. Yeah. So but we want to talk. Also about [00:05:20] your AI involvements. And in this episode, we want to talk specifically about all the things happening in clinical research. So if you look at the current landscape of AI in clinical research. What do you see? 

[00:05:39] Francois: Well, I [00:05:40] think there are a lot of people who are looking into that right now.

[00:05:42] Okay. And since the GPT thing came live beginning of 2003 and 23, so more than a year and a half now, there is, has been a lot of, Research and activity and investment into the field. I haven’t really seen [00:06:00] anything being completed yet. I think we are still in the learning curve and the development curve.

[00:06:08] But there are a lot of activities going on. There is a lot of investment. There are a lot of interesting features that are being brought up. And [00:06:20] yeah, it’s a very nice time just to get involved, especially as a statistician, because It’s, it’s a modeling activity. 

[00:06:28] It’s a, yeah, basically you, you, you make a prediction.

[00:06:32] It’s, it’s a, I mean, the GPT is a kind of an autoregressive model. You want to predict the next word in the sentence. And, but the [00:06:40] data is different. You know, you are not playing with the structural data that we used to. We are playing with, I mean, anything can become data, really, provided it’s processed by a computer.

[00:06:52] And that’s really the interesting part. And just to go to your question, which is about [00:07:00] the application of this technology within the clinical research area. I basically see two trends. I mean, as far as I’m involved, The first is related to this digi, the, the Di Digital Twin Initiatives, which is basically using machine learning methods in order to [00:07:20] predict what is going to be the response of a given patient.

[00:07:23] A given is, is baseline profile and, and the the treatment and the inter intervention being, so there has been a lot of modeling and simulation efforts that were conducted beforehand, but it was models that were trying to [00:07:40] understand the things in order to make predictions. And you know, The better you understand the, the, the less predictivities and vice versa with the machine learning models.

[00:07:50] We are not really so much into the understand thing about things, it’s more about the prediction of things. But, but the predictions are relatively good [00:08:00] and there are, there are ways to, to monitor the accuracy whether your model is overfitting or not and this kind of thing. And so. Overall, that gives you the possibility to basically simulate patient without having the patient responses during the [00:08:20] trial, without having the patient to come physically, and then to basically do better designs for your study and to optimize the development of your patient.

[00:08:31] Of your clinical research and in that front, we have been very, very active at at clean bay and produce. We, we have basically developed some automated machine [00:08:40] learning techniques. So these are techniques where you basically don’t have to do all of these checks and processing of the data testing multiple models, optimizing your models, et cetera.

[00:08:53] It’s done automatically in a pipeline manner. So you basically have your, your input [00:09:00] data. You define what is your response, what kind of response, what are the covariates, what type of covariates, and then it does everything for you. Up to the end, and it’s find the best, the best model according to some, some performance criteria.

[00:09:14] And it’s really helping us to streamline this kind of analysis and its application in [00:09:20] clinical research. And what we are promoting a lot in this field is That’s all of the face, all of the efficacy studies, I think, should go through a machine learning pipeline now. Basically, when people are running efficacy trials, they want to make comparisons of treatments.

[00:09:37] So they want to compare groups. They look at [00:09:40] cross sectional type of approaches. Whereas the machine learning thing has the ability to look longitudinally in a patient and make predictions for the future. And I think we should basically take advantage of the data that has been collected in the clinical research.

[00:09:58] Run this machine [00:10:00] learning methodologies and use this in order to help us design the next phase in the development. And the companies who are able to do this kind of things are going really to get the competitive advantage when it’s going to come to design the next phase and then to improve the success rate of their development.

[00:10:18] So [00:10:20] that’s probably a long, long answer. 

[00:10:23] Alexander: Yeah, I think this is a super exciting area to work on. And yes, currently there’s a lot of hype about generative AI and then people completely forget about there’s also other things within machine [00:10:40] learning that are not generative AI. 

[00:10:42] Francois: So yeah, with respect to that, it’s really the first point is really recycling the existing clinical data that we have in order to look at the data in a different way that is going to be helpful for you to design your [00:11:00] next phase.

[00:11:00] This is the goal. Now, with respect to generative AI initiatives, and, and there is also a lot of things that can be done. And there are a lot of people who are doing a lot of things using that. But I think with respect to that, the biggest change is going to be [00:11:20] through the processing of the data in a study.

[00:11:24] So the, we, we used to basically. Manipulate the patient experience in the study into and to basically do it into a multi part type of thing. We basically got the patient going [00:11:40] to the doctor’s appointment. The patient is reporting his symptoms, whether he feels better or not, etc. The doctor puts that in a case report form and then it’s encoded into a database.

[00:11:53] You make the mapping for the adverse events into a dictionary and at the end of the day, you analyze data [00:12:00] that basically you claim are really mapping exactly what the patient meant to do, but which are very structured. And there has been a lot of information lost and a lot of activities happening from the time that the patient met the doctor.

[00:12:17] To the time you do the analysis. And I [00:12:20] think the big change that is, we are going to see is that we are going to be able to analyze directly the recording between the patient and the, and the doctor. 

[00:12:31] Alexander: Okay. So do you think we could come back? So let’s say the patient and the doctors, I [00:12:40] have a, have a discussion and say like typically what happens every day.

[00:12:47] In the clinic. Yeah. You come to your physician. You get asked lots of different questions. You explain about your background. All kind of different things. And what’s [00:13:00] currently happening is the physician or the site nurse or whoever takes that information. Thank you. And captures everything that is in the case report form accordingly.