Have you seen our new ‘How I treat…’ features? Become a member (it’s free!) and discover the latest advice from fellow clinicians here >>>

How is artificial intelligence changing clinical trial feasibility and recruitment?



How is artificial intelligence changing clinical trial feasibility and recruitment?



Hello, and welcome to this episode of OC talks Podcast Series brought to you by Oncology Central. I am Jade Parker, Senior Editor of Oncology Central, a free online platform that unites all aspects of oncology to support a multidisciplinary approach to progression of the field. In this episode today, we will be speaking to Wout Brusselaers from Deep 6 AI about how artificial intelligence (AI) is changing clinical trial feasibility and recruitment. Thank you for joining us today, Wout.

To start off could you briefly introduce yourself first, Wout? 

Sure. Hi, my name is Wout. I am the CEO and Founder of Deep 6 AI. Deep 6 as the name suggests uses AI on clinical data to help accelerate clinical trial recruitment, site selection, design and construction.

Thank you. Why is clinical trial recruitment something we should be focusing on? 

Well, at the higher level, right, clinical trials play a major role in most, if not all, healthcare innovation. As you know, every new drug, device, procedure or treatment must be tested on real patients in clinical trials to show both that it is safe and that it works. So clinical trials are something like the gatekeeper that allow effective healthcare innovations to go to market and become available to patients. But unfortunately, in the last couple of years, that gatekeeper has become somewhat of a bottleneck and that is because we simply cannot find a sufficient number of eligible patients and recruit them onto trials.

As a result, nearly 90% of clinical trials run significantly over time or over budget. And up to about 40% of trials will fail because they cannot meet their accural goals. This has a tremendous impact on the development of new drugs and cures across the industry. It adds time for new potential lifesaving cures to make it to market. The original gatekeeper has become bottleneck and that is why I think it is super important for us to recognize this, if we want to boost innovation in healthcare again.

Can you tell us how patients are currently selected for clinical trials? 

Sure, there is a whole range of different processes and techniques but most of them have in common that they are all very labor intensive and require a lot of cumbersome browsing of patient records. The most traditional way of trying to recruit patients is probably asking physicians for referrals. As you can imagine, I mean, current clinical trial protocols have easily between 25 and 40 different inclusion-exclusion criteria. For any physician to know all their patients well enough to remember whether they are a match for each of those 40 criteria is just impossible. So it is a kind of a haphazard method and it is hit or miss, and then you still need to validate them. Another method typically is to use the rudimentary IP software capabilities to search the EMR, the electronic medical records, for patients.

The problem that you have there is that most of the salient information is put in by physicians into physician notes or into genomic reports or into pathology reports etc. Most of those documents are in unstructured or at best semi-structured form, either kind of free text without any labels. That makes it really hard to search by traditional software. So again, you have to do the same thing, you search for a couple of the criteria, and then you find probably way too many or way too few patients and you can manually go through and find them. There is also third party services like CROs and social media campaigns that you can use but again all of those bring in certain patients and you still have to go through the records to do the actual matching.

There you mentioned some of the challenges in clinical trials, what are the key challenges in clinical trial recruitment in oncology? 

Well, in oncology more so than in any other therapeutic area, time is a decisive factor. Patients who have been diagnosed with a cancer do not have that much time. I mean disease progression can be quite aggressive. So you want to put them on a clinical trial as soon as possible. The other factor is that if you get diagnosed with the cancer probably, you will start some kind of the treatment whether it is radiotherapy or whether it is chemotherapy or another therapy very soon after diagnosis. That means you have only very small window of time to identify a clinical trial for the patient and put that patient on the trial before we start a competing line of therapy. Once you start another line of therapy, most patients become ineligible for a clinical trial.

In your opinion, what role does AI play in solving some of these challenges that you have mentioned? 

Well, AI has the ability to deal with all of the unstructured data. An AI can learn from what it sees in the data and help fill in the blanks when certain data points are missing. Based on those capacities or capabilities, it can radically accelerate the ability to find basis for trials, and it can actually use its ability to search data across the entire value chain of the clinical trials starting with trial design, feasibility, site selection, patient recruitment and also under study management because AI can also find in data actually real time outcomes and feed them back to the research teams in real time or quasi-real time. So you can see the impact of the device or drug that is part of a clinical study as close as possible to real time. And then make decision in order to basically adapt the trial to the needed outcomes of what you are seeing. It can also help with adaptive trial design.

AI has also been linked to improve trial feasibility, could you explain how? 

Today, as I mentioned before, roughly anywhere between 25–40% of trials fail, mostly due to a lack of patient accrual. In too many cases, wishful thinking cannot drive trial design and trial acceptance by side. If you could actually use AI on your real time, real-world data population, you could figure out which patients actually exist and you could test whether you have patients for the trial the way you designed it today, before you even actually push it out the site before you decide to go further. Some of our clients use our software in a mandatory setting for any PI (principal investigator) to test and run their protocols, make sure to have the patients before the IRB will either look at their study and consider discussing it and allowing it.

Looking at the wider picture of AI, what are some of the common misconceptions you hear about how AI is being used? 

There is quite a few. There is misconception about AI across society in so many capacities and probably even more so in healthcare than anywhere else, because it does hold a lot of promise but it is easy to overstate or to, again, add a lot of wishful thinking to that promise. One thing about AI there is no near term, Dr AI right, there is no near term capability for AI to take over taking care of patients etc. I think we are very far away from that. One reason for that is that the data quality is still an issue. Even if you would have a perfect AI or set of algorithms that could do lots of detection or prediction and recommendation, etc., the data that would be used to feed those recommendations is still pretty varied and sometimes even downright contradictory because data that it is in a patient record can take many shapes and many forms. There is the formal structured data that is a list of symptoms and treatment and medication demonstrations.

Then there is a lot of the unstructured data that goes into genomic, that goes into physician, etc. Sometimes that data is old, there may be hypotheticals in there, which may or may not be that easy to identify even by AI. There are various types of deviations that may actually take certain or contain certain grammatical errors, which also makes them harder to content all of those are solvable problems. On top of that, some of the data is just downright contradictory. We have seen sometimes that genomic reports will list a certain genetic mutation for patient and then a recent physician report would actually transcribe that incorrectly or say its positive instead of negative, so long as we don’t solve the data reliability, AI is kind of limited.

You mentioned patient data. I think one of the big questions in this remit is how do you protect patient data when developing such advanced software? 

It is a great question. I can only talk for ourselves, many companies know that data protection and patient health record protection is crucially important and is basically the foundational layer of playing in this field. There are different ways that specifically in an AI setting, you have to safeguard that well. Of course, there is access to data and potential breaches and the ability again to use protected health information (PHI) or to actually leak PHI which is something that is similar for any type of software, specific for AI as well, just to make sure that the learnings from your data can somewhat be identified as well if you learn from a limited set of data points, which in AI it is hard, you mostly try to get as much data as you can.

If you are dealing with rare diseases, you have to make sure that you do not have any indications of the data that you use to derive certain conclusions or altering your algorithm. On top of that, when you are dealing with massive amounts of data, which is the other side of this, which is also more typical for AI, you have to make sure that you can do your training in a safe environment without contamination and without potential breaches. For us, we had to design both our technology, our platform, and our business model to minimize the risks and focus on security.

One key component for us is that we do not want to own any data. We want our clients to own their own data, we will ingest the data, create a new environment for them to make sure that they actually can put their security and their instructions in place as well. We follow their typical protocols and their BAAs (Business Associate Agreement) on top of the platform that we create to make sure that we live up to the higher standard but it is an ongoing thing. I feel like we are by virtue of coming out of the US intelligence community, we are very trained, and we are very aware of the many data protection issues. I think healthcare as a whole is aware of that as well with risks you have to keep on investing in technology and the processes and training your people to make sure that this is one of the key focal areas of anybody involved who touches data.

Do you have any closing comments you would like to add? 

First of all, let me thank you because I am grateful for your efforts to highlight the most of the clinical trials within healthcare as that driver and gatekeeper of innovation and healthcare. I believe that the downstream effects for clinical trial recruitment are enormous both in terms of getting actually lifesaving treatments to patients faster, but also the major factor in the ever-rising drug prices. If we can actually become more efficient with clinical trials, we can actually get new drugs to market faster and cheaper, which will hopefully also stop the inflation in drug prices.

At the same time, I also want to state that clinical trial acceleration is a great use case for AI. It really is a specific scenario where there is lots of cumbersome manual labor being used today to find and recruit patients onto clinical trials. By using AI, you are not going to eliminate all of that labor, but you will make it so much more effective from trial design to site selection, recruitment etc. I do believe this can have a major impact on healthcare, not just for the sponsors to get their drugs to market faster. Also, patients get access to the drugs but also for health systems. They can use AI to run the research efforts much more efficiently and effectively and turn what is typically a loss leader into more revenue and a profit center. I believe that is a triple win for the sponsors, patients and providers. Again, it can help accelerate the clinical trial process and at the micro level, the rate of innovation in healthcare as a whole.


Thank you for joining us Wout. We look forward to seeing how AI can advance oncology.


Wout Brusselaers
CEO + Co-Founder

Wout Brusselaers is the CEO and co-founder of Deep 6 AI. Wout started his career as a diplomat in the Middle East before joining global management consulting firm, McKinsey & Company, serving a wide variety of industries across Asia. Next, Wout founded an extreme sports adventure company in Singapore and then spent five years in the high-stakes world of international security, growing and leading a 22,000-headcount organization across four continents into a $500M global player.

Wout co-founded Deep 6 AI as a cutting-edge AI platform for the intelligence community, arguably the most complex data environment in the world. Since 2016, Deep 6 AI has focused exclusively on healthcare, working with industry leaders such as Cedars-Sinai Medical Center. Deep 6 AI now uses its expertise to accelerate patient recruitment exponentially and get life-saving cures to people faster. Wout is a frequent speaker on artificial intelligence in healthcare and works closely with hospitals, pharma companies and research institutes on patient recruitment programs.