AI’s Impact on Diagnosing Rare Diseases in Children

Every day, more than a thousand children go through the doors of Boston Children’s Hospital. Many of them leave with specific diagnoses and treatment plans, but a smaller group, affected by rare illnesses, often leaves without answers. However, recent advancements in artificial intelligence (AI) are beginning to change that.

New findings from a collaboration between the hospital’s rare diseases center and OpenAI indicate that readily available AI tools can assist in pinpointing genetic errors potentially linked to these children’s ailments. The results, published in the New England Journal of Medicine’s AI-focused section, highlight how OpenAI’s o3 model was able to clarify diagnoses for 18 children who had previously been unable to identify the sources of their conditions.

“It’s revolutionary,” commented Catherine Brownstein, one of the lead researchers and the scientific director of genetic investigations at the Manton Center for Orphan Disease Research at Boston Children’s Hospital. The research team reviewed several hundred genomes from patients who hadn’t received answers regarding their rare diseases.

Brownstein noted that they uncovered nearly 5% new diagnoses through this research—a figure that might seem small at first glance but represents a significant breakthrough given the number of prior analyses. Each successful identification provides vital insights for the families involved.

The Manton Center collaborates with over 3,500 individuals worldwide, spanning all 50 states, who are impacted by rare diseases, working closely with hospitals and health centers across the globe. Brownstein explained that while the hospital regularly screens patient genomes against newly discovered genes, these screenings often yield no new findings.

Identifying the genetic basis of a disease is no easy task. There are roughly 20,000 protein-coding genes in the human genome. Although sequencing is relatively straightforward nowadays, establishing clear cause-and-effect relationships in the complex genetic data can be challenging.

Suyash Shringarpure, a technical researcher at OpenAI focusing on health applications, shared that researchers are often constrained by time when dealing with individual cases. A case might remain unresolved initially, but new findings could emerge later that clarify the genetic links.

Today’s advanced generative AI systems excel at sifting through extensive data to identify these elusive relationships. Brownstein and the Manton Center researchers turned to OpenAI’s capabilities, wondering if commercial AI products could assist geneticists in associating sickness with specific genes.

During their research last year, the team processed the genomes of 376 undiagnosed patients through the o3 system, which was then the leading AI tool available. They fed the model clinical notes, symptom descriptions, and targeted lists of potential genes to search for diagnostic hints. A human research team then reviewed the AI’s outputs to confirm any diagnoses.

From these 376 cases across four different disease categories, they discovered new diagnoses for ten children with rare neurodevelopmental conditions, four with neuromuscular disorders, two with unspecified sudden deaths, and two diagnosed with early childhood psychosis.

One such case involved Kyra Benton, who received an unexpected diagnosis due to this research. At nine, her mother observed changes in her movements, prompting consultations with specialists, yet no conclusive answers emerged. After years of deteriorating health, Benton visited Boston Children’s Hospital, where they were also baffled by her condition.

By the time Benton was thirteen, she faced severe heart issues that led to a tracheotomy. She had resigned herself to the possibility of never knowing what was wrong—until last year, when Manton Center researchers contacted her with a diagnosis: myofibrillar myopathy, a genetic neuromuscular disorder that results in muscle fiber breakdown.

“I didn’t expect to receive news like that after 15 years,” Benton recalled. “The researcher’s call was a surprise just before my 20th birthday, and it felt like a weight had lifted.”

Brownstein expressed her astonishment that a commercial system like o3 could identify new insights in genomes that had been previously analyzed numerous times—analyses that would typically take human experts days to navigate, if they could reach similar conclusions at all.

“The workload is immense, and the LLM doesn’t fatigue,” she pointed out, emphasizing the shortage of geneticists and analysts capable of sorting through the intricate genetic data to determine the causes of patients’ symptoms.

Dr. Adam Rodman, an AI medicine expert at Beth Israel Deaconess Medical Center who wasn’t part of the study, found the paper to be an exciting example of AI’s potential in aiding diagnoses when employed by physicians. “A 5% diagnostic yield is significant and could help manage existing case backlogs more effectively,” he noted.

The study reiterated findings from prior research about the ability of large language models to sift through genomes for genes discovered after patients’ initial hospital visits. However, the authors emphasized that this new research illustrates how doctors nationwide can utilize commercial AI systems to expedite their work and enhance patient access to vital health information.

Chunhua Weng, a bioinformatics professor at Columbia University who wasn’t involved in the study, praised the research as a “valuable” addition to the field. Yet, akin to the study’s authors, she cautioned that the outputs from LLMs should always undergo thorough human assessment. “Trustworthiness is crucial when using LLMs for diagnosis,” she emphasized.

The research also pointed out that seven of the newly identified diagnoses were categorized as “rediscoveries”—diagnoses made by other teams that hadn’t been shared globally. Brownstein underscored the importance of these rediscoveries for ensuring that patients can access new therapies as they arise.

OpenAI’s health team lauded the results as evidence of how current AI technologies can profoundly benefit patients’ lives. However, the researchers were careful to clarify that their findings are not a catch-all solution. Diagnosing a specific condition is often just the beginning of a longer journey toward finding effective treatments, and LLMs should not be viewed as a substitute for professional medical advice.

“We don’t want to exaggerate this,” remarked Ashley Alexander, OpenAI’s head of health. “But it’s essential that people recognize the potential of the tools, even the ones that are easily accessible today.”

Benton added her own thoughts on the AI aspect of her diagnosis, revealing a mix of skepticism and recognition of its advantages. “I’ve historically been hesitant about AI, yet I see its benefits, especially in cases like mine that can significantly improve lives.”