A glowing 3D organ with pulsing veins spins on a flashing dashboard. This generatively trained AI simulation, known as a digital twin, is a virtual replica of a person. It has the unusual ability to mimic the biological complexities of humans with meticulous precision – a powerful tool when treating patients, for example running simulations to predict the outcomes of different treatment regimens. And in research, experts say digital twins could help streamline clinical trials and alleviate racial and ethnic disparities in data.
Worth $1.17 billion in 2022, the global healthcare digital twin market is expected to exceed $38 billion by 2032, according to a 2024 study. report by Towards health care. When digital twins model patient outcomes, they generate synthetic or artificial data that is often used to supplement real-world data, said Rashidi Hooman, the new associate dean of AI in Medicine at the University of Texas Medical School. the University of Pittsburgh. And this, he continues, is beginning to change the landscape of life sciences beyond clinical practice. Digital twins can also be used directly in clinical trials, predicting participants’ outcomes if they had been in different treatment groups.
“We’re going to have an ‘AI tsunami,'” Hooman said BioSpace in an email. “We haven’t even had a glimpse of it yet. This will happen in three to five years.
“(Digital twins) are now beginning to address our diverse needs (e.g., drug development or biomarker discovery) and are helping to accelerate our areas of research and development,” he continued.
Pharmaceutical industry adopts digital twins
An early adopter of digital twins in the pharmaceutical sector is Bayer. Sai Jasti, Head of Data Science and Artificial Intelligence at Bayer Research and Development, said: BioSpace in an email, he sees digital twins as a “transformational tool” in the pharmaceutical giant’s arsenal of resources that he hopes will drive innovation in drug development.
“Leveraging digital twins allows us to improve the accuracy of clinical decision-making and streamline trial processes,” he said. “This data improves our understanding of patient responses and treatment effectiveness, providing a more comprehensive view that is often lacking in traditional methodologies. By employing these models, we aim to improve patient stratification and increase the overall efficiency of our trials.
In 2023, Bayer and AstraZeneca reached an agreement collaboration with Toronto-based AI company Altis Labs to help accelerate and improve cancer trials using AI-generated digital twins. As part of the agreement, the companies were granted early access to test and implement Altis’ digital twin technology.
Founded by Felix Baldauf-Lenschen in 2019, Altis intends to advance precision medicine by partnering with health systems to train and validate its AI models on anonymized historical patient data. The company trains its models on real-world historical data to predict patient outcomes under standards of care “so that in a clinical trial we can predict what a patient’s outcome would be if they received the standard treatment care,” he said. said BioSpace.
Altis’ approach was presented at the European Society for Medical Oncology (ESMO) 2024 Conference, where a poster presentation highlighted its AI prognostic models in non-small cell lung cancer. According to the poster, Altis’ AI model predicted overall survival and showed the potential to improve quantification of treatment effect beyond tumor size measurements that have historically been used as a basis. to assess treatment response and markers used in regulatory approvals.
Reducing data disparities
Digital twins also have the potential to help address data disparities among underrepresented patient groups, Baldauf-Lenschen said.
The FDA recently released draft guidelines intended to standardize the collection and reporting of race and ethnicity data in clinical trials. AI and digital twins could help in these efforts, Baldauf-Lenschen added.
Subpopulations such as pregnant women, people of color and children are “extremely underrepresented.” . . in terms of evidence for treatments being evaluated in clinical trials, and that’s where there’s a really exciting opportunity to start using AI to improve evidence generation,” he said.
Demographic differences between clinical trial populations and real-world populations may result in different efficacy profiles, Baldauf-Lenschen explained. For example, there might be a more attenuated treatment effect in older, sicker patients compared to what was observed in a clinical trial population, he said.
One area where digital twin models could improve AI performance is in people of color battling skin cancer, according to the Melanoma Research Alliance. Although people of color are less often diagnosed with melanoma, they are up to four times more likely to be diagnosed with advanced melanoma and 1.5 times more likely to die from the disease.
A study published in JMIR Dermatology in 2022, they noted a “white lens phenomenon”, leading to the underrepresentation of images of dark skin conditions in dermatology resources, which the authors believe has disadvantaged people of color by having AI diagnostic systems trained with light skin color images. They added that AI had the potential to solve this problem.
“Deep learning approaches can generate realistic images of skin lesions that enhance the skin color diversity of dermatological atlases. The diverse image bank, used here to train a (convolutional neural network), demonstrates the potential for developing generalizable skin cancer diagnostic applications using artificial intelligence,” the authors wrote.
Challenges in the matchmaking space
The experts who spoke with BioSpace However, everyone recognized that there remained a multitude of challenges to overcome.
Hooman said that while he sees vast innovation in the area of digital twins, the technology is not foolproof.
“None of these frameworks are perfect. They may seem like magic, but it’s not magic. They track basic statistical parameters and predictive analysis,” he said. “There is a built-in error component no matter what AI you use. That being said, knowing what they can do, we also know a lot of their limitations.
One of the key challenges, Jasti said, is “adapting to regulatory requirements, which requires early engagement with authorities to ensure compliance”. Second, the quality of the historical data collected, which “is vital”, particularly in terms of image annotation. Scientific feedback and collaboration helps Bayer address these issues, he said.
Another challenge is the amount of data collected. Digital twins require an “excessive amount of patient data” to create accurate representations, says a report by Cromos Pharma. “The complexity and volume of data required can be a significant barrier. »
Baldauf-Lenschen agreed, noting that it becomes more difficult to train AI models when dealing with smaller subsets of populations, such as underrepresented patient groups.
Hooman said that in healthcare, when AI makes a mistake, the consequences can be deadly. “In our world, things are much more sensitive. If AI makes a recommendation on Amazon. . . On the other hand, if the AI makes the wrong recommendation for your chemotherapy treatment, it could hurt or kill someone.
Still, he said he was hopeful. “I believe the number of positives will be far greater than the number of negatives,” he said.