AI Aids in Diagnosing 18 Children With Previously Undiagnosed Diseases

Researchers at Boston Children's Hospital have utilized artificial intelligence tools to diagnose 18 children with rare diseases that had previously eluded identification. The study, published in NEJM AI, demonstrates AI's potential to significantly shorten the time it takes to find answers for children suffering from rare illnesses and complex symptoms.

By employing OpenAI's o3 Deep Research model, the team analyzed hundreds of patient genomes, identifying new diagnoses in nearly 5% of cases. This AI-driven approach proved capable of identifying errors in patient genomes, a task that can be labor-intensive for humans. Analyzing the approximately 20,000 protein-coding genes in the human genome to pinpoint the genetic cause of a rare disease can take days when performed manually.

The research team processed the genomes of 376 patients who had not received a diagnosis, alongside clinical notes and symptom descriptions. The AI system successfully identified new diagnoses for 10 patients with rare neurodevelopmental diseases, four with neuromuscular disorders, and two with early childhood psychosis. Additionally, the AI helped identify the causes of sudden, unexplained deaths in two children.

Lead study researcher Catherine Brownstein described the development as a "game changer," noting that each new diagnosis offers crucial answers to families. She compared the process to finding a needle in a haystack, highlighting that while clinicians face extensive gene data, AI tools do not experience fatigue.

This work builds on previous research exploring the use of large language models to scan sequenced genomes for rare disease markers. While other health experts have lauded the findings, emphasizing the potential for faster diagnoses across the country with commercial AI tools, they also stress the continued necessity of thorough human review. Experts caution against using LLMs for self-diagnosis, emphasizing that these tools are intended to assist medical professionals in navigating complex information.

OpenAI has expressed enthusiasm for the study's outcomes, but the research team reiterated that these AI tools are not a universal solution for all undiagnosable conditions and should not be used by consumers for independent medical diagnosis. Instead, the technology is presented as a powerful aid for physicians to more efficiently process and interpret intricate medical data.