The Center for Data Innovation spoke with Charles Fisher, chief executive officer and founder of Unlearn.AI, a U.S. firm that uses AI to build digital replicas of patients in clinical trials. Fisher discussed how this approach can address patient recruitment, one of the main challenges associated with trials for diseases like Alzheimer’s.
Eline Chivot: What has led you to build Unlearn.AI? How is your approach to clinical trials different from traditional solutions?
Charles Fisher: I’m a biophysicist by training. I did a PhD at Harvard University, studying different ways of applying machine learning problems in biology, and then did a couple of postdocs including as a researcher at Ecole Normale Supérieure in Paris. I worked at Pfizer back in Boston, as a machine learning scientist in clinical trials, and later at a virtual reality company in San Francisco. We were thinking about ways to develop new machine learning technologies to improve healthcare. Instead of using approaches developed for and applied to other areas, we started to look at the problems specific to medicine, as these are quite different, and using the kinds of data we encounter in medicine, to build new machine learning technologies. That’s how we started Unlearn.AI, three years ago.
Typically, clinical trials don’t use all the information that is available about a disease. In addition to patients in the trial, there is data from tens of thousands of patients, maybe millions, that you can collect from all kinds of medical records or previously developed clinical trials. We thought about ways in which we could incorporate all of that additional data about a disease into a clinical trial. In engineering, people think of digital twins as the computer model of a device that enables you to run simulations about how that device functions. Because we don’t have blueprints for a person like we do for a real device, we apply machine learning methods to build the computational model that’s able to generate digital twin simulations of all these patients rolling in the trial, and their corresponding virtual medical records (which match people according to biomarkers, lab tests, and demographics). We then incorporate that information into the analysis of a clinical trial, and provide predictions of what would happen to them if they were able to receive a placebo or existing treatments. The long-term goal would be to reduce the number of real patients needed to enroll in a trial. The efficiency or time gains that you would get from that would be enormous.
Chivot: How does your machine learning model help researchers design better results for clinical drug trials and make them more efficient? Why is that becoming so critical?
Fisher: Clinical trials are an incredibly long and expensive process. In the area of Alzheimer’s disease, for instance, running those trials can take eight years and costs $500 million. This process slows down the entire pace of innovation in medicine. What can we do to speed that up—if you only have to recruit half as many patients, a trial would be twice as fast. Even with fewer participants, researchers would then compare the simulation to their observation of trials and their effect on the patient, and generate better evidence using statistical analysis.
Most clinical trials compare some new treatments to an existing treatment—which is called the standard of care. With the standard of care, we can get an enormous amount of historical disease data—from all over the world, from previously run clinical trials and electronic medical records. We can ingest it, and apply what it learns to the person’s medical history and personal information, creating a digital twin, building models to predict how patients will respond to that standard of care.
In the future, we could use the computational model for every patient receiving a new treatment, and simulate the behavior for one patient at a time. In order to get there, there’s an enormous amount of validation that needs to happen. So right now, we are applying a hybrid method, where you have some patients receiving the control treatment and placebo in the study, as well as these simulated digital twins that provide these simulated controls for all the patients.
Chivot: On which disease have you been testing your technology, and what are the accomplishments you are most proud of?
Fisher: We focus on complex diseases, where you have many symptoms that are changing over time. That is the case of a lot of neurological diseases, and these are the most difficult to study. In the United States, there hasn’t been a new drug approval for Alzheimer’s disease since 2003. There’s been hundreds of clinical trials that have been run since then, but our ability to develop new therapies for Alzheimer’s has been very poor. We saw there was an enormous unmet need there.
In September, we published a peer-reviewed paper in Nature Scientific Reports that describes our work on Alzheimer’s disease. We’ve started working with a partner to run a phase 3 pivotal clinical trial with digital twins. Passing a phase 3 trial successfully means that they would be able to file for approval by the U.S. Food and Drug Administration (FDA).
We’ve had a lot of really great discussions with the FDA about our approach and how to apply and use it for clinical trials. We have also done a lot of things under the radar in research and development. Because we use new methods, we have to use new, custom software, and we want to write it with really high-quality standards: We put a lot of care into this, and that is in itself something that we can be proud of.
Chivot: What are some of the processes that you use for scientific validation?
Fisher: There are many different levels to validation. The basic level is technical: Does the software work, and do what you say it does? Is it sufficiently tested? For data scientists and machine learning people, validation means: Have you tested this algorithm and this method on new datasets? For this, we use cross-validation, such as by taking 2,000 patients and training the model on half of them and then on the other half. Then there is retrospective validation, we can look at clinical trials run in the past and reanalyze that to see if the approach we’ve taken is adding value to those trials and does not disrupt them. The last level is prospective validation: We run clinical trials and make predictions. Most of the work we’ve published so far on validation has been for technical validation and machine learning data science cross-validation—we have ongoing retrospective and prospective studies which will be published by the end of this year.
Chivot: How do you expect your approach could further impact the delivery of healthcare in the future?
Fisher: In the next five years, we want to run these hybrid clinical trials, incorporating the digital twins into randomized trials. As we start to build up more and more evidence, we would like to be able to roll out more of the data from these digital twins over time. Further ahead, in five to 10 years’ time, we should measure our progress in terms of its impact on the field, by the average time it takes to run a trial. We’d like to decrease that significantly, to accelerate drug discovery processes.
In addition, the idea of a digital twin—running simulations of people and what could happen to them by applying a particular treatment—could guide patient care in the future. Imagine going to your doctor’s office: They have your digital twin on their computer, can ask questions from your twin, and use that information to guide your treatment. That is something very far in the future, but I think we’re currently trying to lay a technological foundation for that.