The medication you took this morning has come a long way from the lab to your pill pack. First, there is extensive laboratory research. Then animal testing. But before a drug can be approved for use, it must be tested on humans — in an expensive, complex process called a clinical trial.
In its simplest form, a clinical trial works something like this: Researchers recruit patients who have the disease that the experimental drug is targeting. Volunteers will be randomly divided into two groups. One group gets the experimental drug; the other, called the control group, receives a placebo (a treatment that appears identical to the drug being tested but has no effect). If the patients receiving the active drug show greater improvement than those receiving the placebo, that is evidence that the drug is effective.
One of the biggest challenges in designing a study is finding enough volunteers who meet the precise criteria for the study. Physicians may not be aware of trials that might be appropriate for their patients, and patients who are willing to enroll may not have the characteristics required for a particular trial. But artificial intelligence could make this job a lot easier.
Digital twins are computer models that simulate real objects or systems. Statistically, they behave in much the same way as their physical counterparts. NASA used a digital twin of the Apollo 13 spacecraft to make repairs after an oxygen tank exploded, requiring engineers on Earth to carry out repairs from 200,000 miles away.
Given enough data, scientists can create digital twins of humans using machine learning, a type of artificial intelligence in which programs learn from large amounts of data rather than being programmed specifically for the task at hand. Digital twins of patients in clinical trials are created by training machine learning models with patient data from previous clinical trials and from individual patient records. The model predicts how the patient’s health would evolve over the course of the study if given a placebo, essentially creating a simulated control group for a given patient.
Here’s how it would work: One person, let’s call her Sally, is assigned to the group that gets the active drug. Sally’s digital twin (the computer model) is in the control group. It predicts what would happen if Sally didn’t receive the treatment. The difference between Sally’s response to the drug and the model’s prediction of Sally’s response if she were to take the placebo instead would be an estimate of how effective the treatment would be for Sally.
Digital twins are also created for patients in the control group. By comparing the predictions of what would happen if digital twins get the placebo to the people who actually get the placebo, researchers can spot any problems in the model and make it more accurate.
Replacing or augmenting control groups with digital twins could help both patient volunteers and researchers. Most people who take part in a trial hope to get a new drug that could help them when approved drugs have failed. But there is a 50/50 chance that they will be included in the control group and receive no experimental treatment. Replacing control groups with digital twins could mean more people have access to experimental drugs.
The technology, while promising, is not yet widely adopted — perhaps for good reason. Daniel Neill, PhD, is an expert in machine learning, including its healthcare applications, at New York University. He points out that machine learning models depend on having lots of data, and obtaining high-quality data about individuals can be difficult. Information about things like diet and exercise is often self-reported, and people aren’t always honest. They tend to overestimate the amount of exercise they get and underestimate the amount of junk food they eat, he says.
Accounting for rare adverse events could also be an issue, he adds. “Most likely these are things you didn’t model in your control group.” For example, someone might have an unexpected adverse reaction to a medication.
But Neill’s biggest concern is that the prediction model reflects what he calls “business as usual”. Let’s say a major unexpected event – something like the COVID-19 pandemic, for example – changes everyone’s behavior patterns and people get sick. “That’s something these governance models wouldn’t account for,” he says. These unforeseen events, which were not taken into account in the control group, could falsify the results of the study.
Eric Topol, founder and director of the Scripps Research Translational Institute and an expert on the use of digital technologies in healthcare, thinks the idea is great
, but not ready for prime time. “I don’t think clinical trials are going to change anytime soon as it requires multiple layers of data beyond health records, such as: E.g. a genome sequence, gut microbiome, environmental data and so on.” He predicts that it will be years before large-scale studies can be conducted with AI, especially for more than one disease. (Topol is also the editor-in-chief of Medscape, WebMD’s sister site.)
Gathering enough high-quality data is a challenge, says Charles Fisher, PhD, founder and CEO of Unlearn.AI, a start-up developing digital twins for clinical trials. But, he says, solving these kinds of problems is part of the company’s long-term goals.
Two of the most commonly cited concerns about machine learning models—privacy and bias—have already been addressed, says Fisher. “Privacy is easy. We only work with already anonymized data.”
As for the bias, while the issue isn’t resolved, it’s irrelevant — at least for the outcome of the trial, Fisher said. A well-documented problem with machine learning tools is that they can be trained on biased data sets—for example, those that underrepresent a certain group. But, says Fisher, because the studies are randomized, the results are immune to bias in the data. The study measures how the tested drug affects the subjects in the study based on a comparison to the controls and adjusts the model to better match the real-world controls. So, according to Fisher, even if the selection of subjects for the study is biased, and the original dataset is biased: “We are able to design studies to be insensitive to this bias.”
Neill doesn’t find that convincing. You can remove bias in a strict randomized trial by adjusting your model to correctly estimate the treatment effect for the study population, but you will only reintroduce that bias if you try to generalize beyond the study. Unlearn.AI “does not compare treated individuals to controls,” says Neill. “It compares people treated with model-based estimates of what the person’s result would have been if they had been in the control group. Any errors in these models, or events they fail to predict, can result in systematic biases — that is, over- or underestimates of the treatment effect.”
But unlearn.AI is moving forward. It is already working with pharmaceutical companies to develop studies for neurological diseases such as Alzheimer’s, Parkinson’s and multiple sclerosis. There’s more data on these diseases than many others, so it was a good place to start. Fisher says the approach could eventually be applied to any disease, which would greatly reduce the time it takes to bring new drugs to market.
If this technology proves useful, these invisible siblings could benefit patients and researchers alike.