Using AI to help physicians diagnose rare genetic diseases affecting children

Rare disease diagnostics often stagnate because clinical knowledge evolves faster than manual review processes can accommodate. Ignoring these backlogs leaves families without answers, while active reanalysis can resolve cases that have evaded experts for years. By deploying AI as a reasoning layer, institutions can prioritize high-probability leads for human intervention. However, the model cannot replace clinical judgment, as it serves only to generate hypotheses rather than provide definitive medical decisions.

A collaborative team from Boston Children’s Hospital, Harvard University, and OpenAI developed a workflow where the o3 reasoning model processes de-identified genomic and clinical data. The model synthesizes fragmented patient records and updated scientific literature to propose molecular explanations, which are then validated by clinicians. This shifts the diagnostic process from a static, one-time event to a dynamic, iterative cycle of review. The model is currently limited by its inability to perform clinical confirmation, necessitating laboratory-certified validation for every result.

This week

• Audit existing backlogs of undiagnosed genetic cases to identify high-priority cohorts for pilot testing.
• Review internal data privacy protocols to ensure clinical datasets are prepared for secure, de-identified AI processing.

This quarter

• Establish an expert review panel to evaluate AI-generated hypotheses against established ACMG/AMP standards.
• Develop a standardized input template for patient phenotypes to improve the consistency of AI-assisted data analysis.

This year

• Integrate periodic automated reanalysis into standard clinical workflows to ensure patient data remains synchronized with evolving scientific literature.
• Measure the cumulative diagnostic yield improvement to determine the cost-benefit ratio of AI-augmented genetic screening.