Publications
Peer-reviewed research from our team on what clinical AI actually costs, what it orders, and where accuracy stops being the right question.
Twenty-four AI systems wrote plans for 200 real outpatient and emergency visits, priced at 2026 Medicare rates. Their care cost $314 per visit against $97 for the treating physician — 3.2× — driven mostly by testing at the 66% of visits where the physician ordered none. Cost-aware prompting narrowed the gap but never closed it.
Link on publicationA DECIDE-AI stage 1 evaluation across 1,138 emergency department patients over four weeks. Outputs were rated clinically appropriate in 99 of 100 sampled cases — yet clinician adoption fell from 68% to 30%. Sustained engagement, not algorithmic accuracy, was the binding constraint.
Read the paper →Fifteen models, 198 US malpractice cases, 3,072 simulated consultations. The models that best matched court-determined standards of care were also the models that ordered the most — 9.3 procedures at $1,118 versus 1.3 at $221 (ρ = 0.95). The coupling tightens with each model generation, and replicates on UK cases.
Read the paper →A dynamic simulation in which AI systems work up 216 peer-reviewed cases by ordering real tests, imaging and procedures, each priced by CPT code. Three systems indistinguishable on accuracy (93.5–94.0%) differed 1.75-fold in cost and 2.1-fold in physician effort — differences no accuracy benchmark can see.
Read the paper →Ten models against 305 neurology board certification questions, each graded on three axes of complexity. Splitting the reasoning across specialised agents — analysis, retrieval, synthesis, validation — carried LLaMA 3.3-70B from 69.5% to 89.2%, and evened out the subspecialty gaps that retrieval alone never closed.
Read the paper →One hundred acute stroke presentations, each read by GPT-4 and scored against a stroke specialist and the care actually delivered. Agreement on thrombectomy reached AUC 0.94, and its 90-day mortality ranking (AUC 0.89) outperformed the supervised risk models in clinical use.
Read the paper →