NEW YORK – The Icahn School of Medicine at Mount Sinai seeks to advance pediatric medicine by leveraging artificial intelligence (AI) to connect genotypes and phenotypes.
Through this initiative, housed at the recently launched Center for Artificial Intelligence in Children's Health, Mount Sinai researchers hope to develop better diagnostic tools and personalized therapies and to improve overall healthcare delivery for newborns and children.
Benjamin Glicksberg, an assistant professor of genetics and genomic sciences at Mount Sinai and leader of the center, said that in general there has been a significant increase in the collection and use of real-world pediatric patient data but that many of the systems used to gather, analyze, and report that data are not interconnected, creating inefficiencies in its use.
Glicksberg said that the primary aim of the center will be to characterize pediatric patients both molecularly and phenotypically, combining DNA and RNA sequencing with computer vision analysis to better understand how disease presents in newborn and pediatric populations and to identify some of the "more subtle" signs of disease that might often be missed otherwise.
Many pediatric samples will be collected as part of the larger Mount Sinai Million Health Discoveries Program initiative and protocol, Glicksberg said, adding that "other multiomic samples will be collected on a clinical trial-by-trial basis."
The Mount Sinai Million Health Discoveries Program is an initiative that is sequencing the genomes of up to one million Mount Sinai patients over the next five years, with the goal of integrating genetics into clinical care.
Investigators with that program, launched in 2022 in collaboration with biotechnology company Regeneron, are conducting exome sequencing and whole-genome genotyping-by-sequencing analysis on patient DNA samples, as well as whole-genome sequencing on a large subset of those samples.
The new center is also a participating site in NICUnet, a Mount Sinai-headquartered research consortium of 16 US academic institutions dedicated to unraveling the science behind rare events such as unexplained neonatal deaths.
Although the center intends to participate in multiple clinical trials, Glicksberg said that the specifics of these plans will come later, after the center fully works out protocols and consent procedures and after it has further developed its computational approaches.
"We're in the building and foundation phase right now," he said.
One initiative that the center has begun work on involves integrating real-time video monitoring in the newborn intensive care unit (NICU) with genomic and clinical data to assist caregivers in more rapidly diagnosing disorders such as rare diseases and neurological disorders that can be difficult to identify in newborns.
Glicksberg said that real-time monitoring systems already installed in some newborn and pediatric intensive care units are "gold mines" of information that are presently being underutilized.
"[This] initiative," he said, "uses video feeds from the neonatal intensive care unit to come up with an alert system where certain neurological changes –– dysfunction –– are reflected in movement."
Although this project will seek to identify any potential neurological issues, Glicksberg said that it will also be applied more specifically to rare disease diagnosis and discovery and to the discovery of personalized rare disease treatments.
Still a relatively young field, computationally assisted deep phenotyping has attracted the interest of some researchers as a more granular way to correlate phenotype with genotype, especially in the rare disease camp, where diagnosis is complicated by both the lack of knowledge regarding many rare disorders and by their often overlapping clinical features.
Computer vision techniques such as pose analysis and facial analysis, Glicksberg explained, can "tease apart" some of the things that are difficult for the unaided human eye to see and to help researchers shine more light on the mechanisms of these disorders.
A greater understanding of those mechanisms, Glicksberg said, is needed to identify potential drug targets that can be used to develop new treatments and to get a better understanding of the natural history of rare disorders –– of how their phenotypes change over time.
"That may also be a clue into trajectories of diseases [and] outcomes that, without this kind of refined understanding, is really hard to piece together," Glicksberg said.
Beyond how these data will be analyzed at the center, Glicksberg said that merely gathering such data on younger patient populations will be of immense value, as less publicly available physical analysis data exists, particularly with respect to newborns.
Most pose models, for example, are trained on people performing natural movement actions such as walking and jumping, which Glicksberg said doesn't translate well to neonatal populations, for whom little such data exists.
Part of this has to do with obtaining regulatory and IRB permission to gather data, but Benedikt Hallgrímsson, deputy director of the Alberta Children's Hospital Research Institute and whose research focuses on applying facial recognition technologies to neonatal rare disease diagnosis, said that the bigger barrier for him has often been finding time to incorporate this research into a busy medical practice.
"Even when you have motivated physicians," he said, "they have a lot going on and it can be a distraction to do something that isn't necessarily clinically useful for them in that moment."
Deep phenotyping remains a somewhat nascent field, and although a few products have made their way to market, such as Face2Gene, from Boston-based FDNA, clinical utility for the technology has yet to be broadly established.
Hallgrímsson is optimistic about this outcome, however, saying that research into the diagnostic applications of facial analysis technology continues to show significant promise.
Two applications that he finds particularly interesting are in assessing individuals with unknown genetic diseases and in more quantitatively measuring disease severity.
The facial similarities between a person with an unknown disease to others with known disease can guide hypotheses about the underlying genetics.
"We know from our analyses that symptoms that are involved [in/with] particular pathways tend to cluster together," Hallgrímsson said. "So there's information there that we don't really know yet how to process."
Computer-assisted facial analysis may also improve physicians' ability to discriminate between rare disorders with overlapping clinical features, Hallgrímsson added, although this bears some nuance because current analytical methods have their own blind spots.
Although algorithms have outperformed humans in correctly identifying rare diseases in some cases, Hallgrímsson said that "as you try to train the system to recognize more and more syndromes, the potential for confusion among syndromes that are similar increases."
While Glicksberg and his colleagues have many plans for the new center, the details of what they can accomplish will naturally depend heavily on their available funding.
Glicksberg said that although Mount Sinai has "generously" provided some of the initial funding, the center is gearing up for a fundraising campaign in the next two to six weeks, with the goal of raising $10 million.
"I'm very confident that we're going to be doing great things," Glicksberg said.