Our Strategy for Navigating the Regulatory Landscape of AI in Drug Development
By Charles Fisher, Founder and CEO, Unlearn.AI
Sometimes the technologies we’re developing at Unlearn sound like science fiction, but we’ve been working extremely hard to turn them into science fact. Our regulatory strategy for our clinical trial solutions has been a key part of that effort.
Over the long term, we aim to advance artificial intelligence to develop the capability of creating accurate, comprehensive digital twins of individual patients. A patient’s digital twin will allow us to simulate how that patient’s health will evolve under different scenarios, or to drill down into details of the patient’s current health that we cannot directly observe. We want to use AI to reveal the unseen aspects of individual health.
In the short term, we are focused on developing and applying methods that use patients’ digital twins to improve clinical trials. Why?
There are a couple of obvious reasons. First, clinical trials are a key bottleneck in drug development and making them better will help to get new treatments to patients who need them. Second, there are many well known problems with clinical trials from high failure rates, to long timelines, to the reluctance of patient’s to participate in trials. The current way that we run trials isn’t working for either clinical trial sponsors or for patients.
The final reason seems to be less obvious to many people, but thankfully it was obvious to us — there’s a clear regulatory pathway to using patient’s digital twins to improve both phase 2 and 3 clinical trials right now if we do it right.
When we create a digital twin for a participant in a clinical trial, we input data collected from that participant at or before baseline into an AI model trained on historical data (e.g., from previous clinical trials or disease registries). Then, we use our AI model to create a probabilistic forecast describing how that participant’s health (e.g., symptoms, biomarkers, etc) would likely evolve over the course of the clinical trial if that participant happened to be randomly assigned to the control group. That is, the participant’s digital twin provides a probabilistic forecast of their prognosis.
A few years ago, we had an epiphany. Current regulatory guidance supports adjusting for prognostic baseline covariates in randomized clinical trials. The more prognostic a baseline covariate is, the more power you can add to a clinical trial. A prediction obtained from a well-calibrated probabilistic forecast is just about the most prognostic covariate possible. Therefore, we could compute prognostic scores from participants’ digital twins and use them to create much more powerful clinical trials without requiring any changes to regulatory guidance. Moreover, we thought we could take into account the effect of the prognostic scores on the trial power during its design, allowing us to design trials that can achieve high power while using smaller control groups. We called this technique Prognostic Covariate Adjustment (PROCOVA™) and built our TwinRCT solution around it; and we could do it right now.
Don’t believe me?
Check out these excerpts from FDA’s recent guidance Adjusting for Covariates in Randomized Clinical Trials for Drugs and Biological Products.
“Covariate adjustment leads to efficiency gains when the covariates are prognostic for the outcome of interest in the trial. Therefore, FDA recommends that sponsors adjust for covariates that are anticipated to be most strongly associated with the outcome of interest. In some circumstances these covariates may be known from the scientific literature. In other cases, it may be useful to use previous studies (e.g., a Phase 2 trial) to select prognostic covariates or form prognostic indices.” (Emphasis mine.)
“In a trial that uses covariate adjustment, the sample size and power calculations can be based on adjusted or unadjusted methods.”
So, FDA guidance says that one can use data from previous studies to form prognostic scores and account for them in the design and analysis of a trial. Check one.
How about the EMA?
In this case, we actually took PROCOVA through the EMA’s novel methodologies qualification process, culminating in a Qualification opinion for Prognostic Covariate Adjustment (PROCOVA™). Here are some excerpts from the qualification opinion.
“CHMP qualifies PROCOVA as prognostic score adjustment and the proposed procedures, as described in a handbook for trial statisticians, could enable increases in power or precision of treatment effect estimates in controlled randomised clinical trials with continuous outcomes.”
“The assumed reduction in residual variance due to a prognostic score may in principle be taken into account to reduce sample size …”
So, EMA guidance also says that one can use data from previous studies to form prognostic scores and account for them in the design and analysis of a trial. Check two.
For many years now, we’ve believed that our approach — PROCOVA — would be acceptable to regulators; it seemed obvious to us, and that’s why we built our TwinRCT solution around it. In general, our strategy has hinged on the belief that it is much easier to build a solution that satisfies current regulatory requirements than to try to create something new and fight to change the regulations.
TwinRCTs based on PROCOVA were the first application of patients’ digital twins that we brought to market, but they will not be the last. For a few years now, we’ve been working on new clinical methods that use more information from participants’ digital twins than just the expected prognosis to deliver even more powerful trials; for example, we could also use the predicted variability (or uncertainty) in a participant’s prognosis provided by their digital twin. We plan on announcing some of these new clinical trial solutions soon.
In fact, we’ve mapped out a plethora of potential applications of patients’ digital twins, both in clinical trials and in other areas of healthcare. We’ve tried to order these from those we believe have the smoothest regulatory pathways to those that have the most challenging regulatory pathways, and we plan to bring them to market in roughly that order. Our beliefs may not be perfectly calibrated, and we may make some mistakes and encounter some roadblocks on the way, but because we chose to start with a foundational solution that satisfies regulatory guidance right now we now have a solid platform on which to build the future.