AI’s Role in Diagnosing Rare Diseases

Online chatbots can sometimes provide questionable health advice, and in extreme cases, they might even lead individuals to the emergency room. Yet, when these tools are utilized by medical professionals, they can achieve outcomes that exceed typical diagnostic capabilities.

A recent study conducted by Boston Children’s Hospital reveals that an AI tool successfully diagnosed 18 children with conditions that had previously gone unrecognized.

These advancements offer hope for children suffering from rare diseases, enabling quicker identification of their ailments.

Published in NEJM AI, the research, conducted by the hospital’s Rare Disease Center alongside OpenAI, demonstrated that AI can pinpoint errors in patient genomes.

Utilizing OpenAI’s o3 Deep Research model, researchers analyzed the genomes of hundreds of patients with undiagnosed rare diseases, discovering that approximately 5% of these evaluations led to new diagnoses.

With around 20,000 protein-coding genes present in the human genome, it’s generally a laborious task for humans to sift through data from numerous patients in search of the genetic basis of a condition.

The research team explored the genomes of 376 patients lacking diagnoses, integrating clinician notes, patient symptom descriptions, and a focused list of genes potentially related to their conditions.

This effort led to new diagnoses for 10 children with rare neurodevelopmental diseases, four with neuromuscular disorders, and two experiencing childhood psychosis.

While specific details were not shared, it was noted that the two children who tragically passed away had been experiencing health issues.

“This is a game changer,” commented Katherine Brownstein, the principal investigator for the study. “Considering how frequently these cases have been analyzed, each new diagnosis signifies an answer for a family.”

Brownstein likened the search through extensive genomic datasets to finding a needle in a haystack, emphasizing that, while machine learning models can significantly aid this process, clinicians still need to sift through countless genes to arrive at a solution.

Previous research had looked into similar methodologies for employing large-scale language models in the genetic analysis of rare disease patients.

Some healthcare experts have praised the findings of this latest study, suggesting that commercial AI tools could expedite diagnostic processes for doctors nationwide.

However, many caution that these AI systems still require meticulous human oversight.

Despite the risks of misinformation, a growing number of individuals are seeking health advice from AI models, as tech leaders and specialists warn of their potential dangers.

“Keeping people informed is vital; it serves as a crucial safety measure,” noted Andy Kurzig, CEO of the AI-powered search engine pearl.com.

While OpenAI expressed enthusiasm regarding the study’s outcomes, the research team highlighted that these findings do not represent a solution for all undiagnosed conditions, stressing that patients should refrain from relying on AI for self-diagnosis.

Instead, physicians can utilize this tool to navigate complex medical data that would otherwise take considerable time to analyze.