OpenAI has rolled out updated health intelligence capabilities for ChatGPT, centered on a clinician-gated diagnostic workflow designed to support rare disease identification. Built on the company’s o3 Deep Research model, the tool is explicitly positioned as a support layer for existing genomic analysis pipelines, not a replacement for clinical judgment. Per official OpenAI documentation, the workflow delivered a 4.8% additional diagnostic yield when tested across 376 previously unsolved rare disease cases that had already been reviewed by multiple commercial and institutional genomic pipelines and multidisciplinary clinical teams. All outputs are framed as evidence-linked diagnostic leads for specialist review, with hard guardrails prohibiting use as a substitute for clinical adjudication OpenAI.
The workflow functions as an explanation-first integration layer atop existing genomic analysis pipelines, rather than a standalone diagnostic tool. For each case, it processes de-identified patient data packets that include standardized Human Phenotype Ontology (HPO) terms for clinical presentation, filtered variant tables capturing variant rarity, predicted protein effect, and ClinVar classification, plus family metadata. Unlike tools that output uncontextualized gene rankings, the model generates evidence-linked justifications for each candidate diagnosis to support specialist decision-making OpenAI OpenAI.
Before testing on unsolved cases, OpenAI validated the workflow on 51 cases with established rare disease diagnoses. In duplicate runs, it recovered the correct gene and variant in 48 instances, a 94.1% success rate for confirmed cases. It also returned correct diagnoses in 45 of 57 neuromuscular cases, a 78.9% accuracy rate for that specific cohort OpenAI.
For 15 long-read genome test cases, the workflow identified the correct gene in all 15 runs, a 100% success rate for that subset. The model’s self-reported confidence scores correlated strongly with accuracy: mean confidence for correct calls was 85.6, compared to 42.1 for incorrect or unknown results OpenAI.
When deployed on the 376 previously unsolved cases, the workflow’s outputs underwent expert specialist review, additional confirmatory testing, and CLIA-certified laboratory validation before being counted as confirmed diagnoses. This process yielded 18 new confirmed rare disease diagnoses that had not been identified in prior review rounds, translating to the reported 4.8% additional diagnostic yield for the cohort OpenAI.
Alongside the health intelligence update, OpenAI released a separate ChatGPT Enterprise spend controls feature for business users. The tool gives administrators granular visibility into credit usage across individual users, supported models, and ChatGPT products, addressing cost tracking needs for enterprise deployments OpenAI.
Frequently Asked Questions
Is the ChatGPT rare disease workflow a replacement for clinical diagnosis?
No. OpenAI explicitly frames all workflow outputs as evidence-linked diagnostic leads for specialist review, with hard guardrails preventing use as a substitute for clinical adjudication. All outputs require formal specialist review before being used in any patient care context, per OpenAI’s official documentation OpenAI.
What patient data does the workflow process?
The workflow processes de-identified patient data packets that include standardized Human Phenotype Ontology terms for clinical presentation, filtered variant tables capturing variant rarity, predicted protein effect, and ClinVar classification, plus family metadata. It outputs evidence-linked justifications for candidate diagnoses rather than uncontextualized gene rankings OpenAI OpenAI.
How was the workflow validated before real-world testing?
The workflow was first validated on 51 cases with established rare disease diagnoses, recovering the correct gene and variant in 48 duplicate runs, a 94.1% success rate. It also achieved a 78.9% correct diagnosis rate (45 of 57) on neuromuscular cases and 100% correct gene identification (15 of 15) on long-read genome test cases OpenAI.
What was the real-world diagnostic yield for unsolved cases?
When tested on 376 previously unsolved rare disease cases reviewed by multiple commercial and institutional genomic pipelines and multidisciplinary clinical teams, the workflow delivered an additional 4.8% diagnostic yield. Following expert review, additional testing, and CLIA-certified laboratory confirmation, this translated to 18 new confirmed diagnoses OpenAI.
Bottom line: Clinical teams working to resolve unsolved rare disease cases can use OpenAI’s clinician-gated ChatGPT workflow as a validated support tool that delivered 18 new confirmed diagnoses across 376 previously unsolved cases already reviewed by multiple genomic and clinical teams, provided all outputs undergo formal specialist review and are never used as a substitute for clinical adjudication.
