The missing economics of medical AI
Clinical decision support appears on 94 recent papers and medical image segmentation on 188, among more than 1,300 medical AI papers in the AI journals, but only 65 of all medical papers carry a cost, reimbursement, deployment or workflow keyword, and cost-effectiveness appears on 27 papers in the whole archive. Which medical AI tools pay for themselves, and for whom?
What prompted this#
Since October 2025, 1,607 papers carry a medical, clinical, diagnosis, radiology, healthcare or patient keyword (1,320 in AI journals): "medical image segmentation" on 188, "clinical decision support" on 94, "medical image classification" on 73, "computer-aided diagnosis" on 46. Only 65 of the 1,607 also carry a cost-effectiveness, reimbursement, economic-evaluation, deployment, implementation, adoption, cost-saving, workflow or clinical-practice keyword; "cost-effectiveness" appears on 27 papers in the archive and "clinical workflow" on 7.
Questions#
- For the medical AI tools that have been reimbursed (for example stroke detection and diabetic retinopathy screening), what did the cost per quality-adjusted life year turn out to be in practice?
- Do tools that save clinician time reduce costs, or do clinicians fill the time with more cases and the hospital bills more?
- Which of the hundreds of published segmentation and diagnosis models have reached any clinic, and what separated them?
Data that could help#
Reimbursement decisions and health technology assessments, hospital adoption records, claims data around adoption, and the literature's own deployment reports.
Next step#
Trace the twenty most-cited medical AI models of 2020–2022 to any clinical deployment and record the economic evidence that accompanied each.