AI Diagnostic Tool Outperforms Clinicians

Sequential reasoning with language models could improve diagnostic accuracy while reducing unnecessary testing, according to new research

Category	Detail
Condition	AI-assisted diagnostic test selection and diagnosis in simulated clinical encounters
Key Mechanisms	Diagnostic Orchestrator (DxO), using large language models, asks questions, requests investigations, and updates its diagnosis as evidence becomes available.
Target Population	304 published clinicopathological conference cases and 325 emergency department encounters; real-world patient populations were not evaluated.
Care Setting	Controlled, text-based evaluation using published clinicopathological cases and emergency department encounters.

The SDBench framework included 304 clinicopathological conference cases and 325 emergency department encounters.
Using GPT-5, DxO achieved 83% accuracy on clinicopathological cases versus 36% for physicians with reference materials, and 80% versus 47% on emergency department cases.
For clinicopathological cases, DxO achieved 81% accuracy at an estimated average testing cost of $2,241; physicians achieved 36% accuracy at an estimated cost of $4,725.
DxO generally asked more questions than clinicians before ordering investigations.
The evaluation was text-based and excluded physical examination, pathology slides, imaging data, incomplete records, and multidisciplinary decision-making.

The study reports that DxO asks questions, requests investigations, and updates its diagnosis as evidence becomes available.
The researchers state that the evaluation was conducted in simulated, text-based cases and did not establish performance in routine clinical practice.

The researchers describe potential use of AI to assist clinicians with diagnostic test selection and diagnostic stewardship.
The researchers state that further evaluation is needed to establish effectiveness in real-world practice and effects on patient outcomes.

DxO can request examination findings but cannot perform physical examinations or observe patients directly.
The controlled text-based evaluation excluded pathology slides, imaging data, incomplete records, and multidisciplinary decision-making.

No patient treatment or prescribing population was evaluated; the study used simulated clinicopathological cases and emergency department encounters.

The article reports diagnostic performance and estimated testing costs, not treatment efficacy, medication use, or prescribing outcomes.

The researchers note that accuracy and cost comparisons were obtained in a controlled, text-based evaluation.
The researchers state that further evaluation is needed to establish real-world effectiveness and effects on patient outcomes.
The article reports that DxO can request examination findings but cannot perform physical examinations or observe patients directly.