|
|
|
In this report
September moved medical AI from promising demonstrations toward more explicit tests of evidence, regulatory fit, and payment.
- CMS ended the automatic Breakthrough Device route to Medicare add-on payments, while Bayesian Health qualified through the standard clinical-improvement pathway:
- OpenAI connected ChatGPT for Healthcare to Epic data and nine public health sources, while a multi-country randomized trial tested LLM assistance with physicians:
- FDA defined a cardiovascular ML notification category and retained 510(k) oversight for radiology CAD, as CROWN supplied a dedicated cytopathology foundation model:
|
|
Evidence, payment, and deployment
CMS ends automatic Breakthrough Device paymentsDeployment, Safety & Reimbursement of AI in Health Systems
CMS will remove the alternative route to NTAP and Transitional Pass-Through Payments on 1 October, leaving substantial clinical improvement as the central reimbursement test for new applicants.
- Devices with Breakthrough designation by 30 September 2026 and FDA authorization by 1 May 2028 can retain the former route through the FY/CY2029 cycle.
- CMS points to the FDA RAPID pathway, launched in April, as an alternative intended to support national coverage within 60 to 90 days for some authorized devices.
What this means for you For medical AI, FDA designation and Medicare payment are no longer linked by default, so regulatory and clinical-evidence planning now have to account for separate reimbursement scrutiny. |
|
Bayesian Health secures Medicare sepsis paymentDeployment, Safety & Reimbursement of AI in Health Systems
Bayesian Health's FDA-cleared sepsis monitor qualified for NTAP through the standard substantial-clinical-improvement route, giving the reimbursement rule a concrete AI example just as the shortcut closes.
- The payment is up to $61.84 per eligible case across an estimated 739 MS-DRGs, for as many as three years from 1 October.
- A prospective Nature Medicine study reported an 18.2% mortality reduction and antibiotics delivered 1.85 hours earlier when clinicians engaged with alerts; the result is associated with engaged use, not an automatic effect of deployment.
What this means for you Clinical AI reimbursement now has a visible example of the evidence bar CMS can accept without the Breakthrough Device shortcut, but the payment case remains tied to prospective outcomes and real clinician interaction. |
|
LLMs in clinical work
OpenAI connects ChatGPT to Epic workflowsLarge Language Models in Clinical Practice
OpenAI launched governed Epic connectivity for ChatGPT for Healthcare, moving the product from general clinical assistance toward authorized access to patient context inside health-system workflows.
- The integration can bring appointment notes, laboratory results, medications, and specialist documentation into ChatGPT, or embed the assistant directly in the EHR workflow.
- A separate plugin connects to nine official sources, including ClinicalTrials.gov, CMS Coverage, RxNorm, DailyMed, and PubMed; physician evaluations across 27 use cases rated 99.1% of 4,363 responses safe, while another evaluation rated more than 93% of connected-source responses good or better for accuracy.
What this means for you The relevant unit of clinical LLM evaluation is expanding from model output to the governed workflow, source provenance, and permissions surrounding patient data. |
|
Researchers test LLM assistance across three countriesLarge Language Models in Clinical Practice
A randomized trial in npj Digital Medicine evaluated LLM-assisted physician decision-making in Indonesia, Kenya, and the Netherlands, extending clinical evidence beyond the single-site and simulated-case studies that have dominated the field.
- The parallel-group design compared physicians with LLM access against standard resources across settings with different income levels and health-system conditions.
- The abstract does not specify the model, and the full results, effect sizes, and subgroup analyses are needed to judge whether access improved performance and for whom.
What this means for you Clinical LLM evidence is beginning to address external validity across health systems, but the strength of the claim still depends on the trial's detailed comparator, outcomes, and subgroup results. |
|
Regulatory pathways and diagnostic models
FDA funds LLM evaluation for radiology reportsAI in Medical Imaging & Digital Pathology
The FDA awarded Cognita Imaging $1.29 million for an 18-month study of whether multiple LLMs can consistently evaluate AI-generated radiology reports, making evaluation infrastructure itself a regulatory-science subject.
- The proposed jury will assess reports from approximately one million patient exams and compare disagreements with radiologist review.
- Radiologists will determine whether errors came from the generated report, the LLM jury, or the reference radiologist report; the contract is research funding, not FDA endorsement of a Cognita product.
What this means for you Validation of generative radiology systems is widening from reader performance to scalable assessment of report quality and failure attribution, but the framework remains under study. |
|
FDA defines cardiovascular ML notification classRegulation of AI Medical Devices
FDA created a Class II De Novo category with special controls for cardiovascular ML notification software while declining to exempt radiology CAD and triage devices from 510(k) review.
- The cardiovascular classification is effective 11 September and applies retroactively from 3 August 2023, creating a predicate type for later 510(k) submissions.
- The radiology petition covered four CAD device types and was denied on 1 April, with the final order published in September, preserving premarket notification requirements.
What this means for you The regulatory picture is becoming more device-specific rather than uniformly lighter: defined predicates may streamline cardiovascular notifications, while radiology CAD remains under 510(k) oversight. |
|
CROWN builds a foundation model for cytopathologyAI in Medical Imaging & Digital Pathology
Researchers published CROWN in Nature Cancer as a cytopathology-specific visual foundation model, filling a modality gap left by pathology models built mainly around tissue sections.
- CROWN used self-supervised learning on more than 10 million cell-image patches without manual annotations and was evaluated across 202 task-setting combinations.
- It exceeded 95% accuracy in 48 patch-level classification evaluations, reaching 98% in 22, across private institutional cohorts and public datasets; code and weights are released for academic research.
What this means for you Cytopathology now has a dedicated large-scale backbone for research, but the reported benchmark performance does not by itself establish prospective screening benefit or deployment readiness. |
|
Qure.ai earns Class IIb mark for AiraLarge Language Models in Clinical PracticeRegulation of AI Medical Devices
Qure.ai's Aira clinical decision-support system received Class IIb EU MDR certification, placing a primary-care AI platform under a higher level of software-device scrutiny.
- Aira combines documentation, workflow automation, EHR analysis, and clinical decision support and has reached ten pilot sites across Nigeria, Kenya, Mozambique, the Solomon Islands, and Bangladesh.
- Qure.ai reports a 32% reduction in protocol-driven documentation time with 98% completion in a Kenya outreach program, and a 23% reduction in administrative time in Nigerian primary HIV care facilities; these are company-reported deployment measures.
What this means for you EU MDR classification is now covering integrated clinical-support workflows, while the available evidence still separates regulatory status from independently established patient or service outcomes. |
|
How Iris covered your topics this month Every topic you track was searched over the last month. | Topic | Sources found | In this briefing | | AI in Medical Imaging & Digital Pathology | 14 | 2 stories | | Clinical Validation & Trials of Medical AI | 10 | Nothing of notable significance | | Regulation of AI Medical Devices | 21 | 2 stories | | Large Language Models in Clinical Practice | 14 | 3 stories | | Deployment, Safety & Reimbursement of AI in Health Systems | 15 | 2 stories |
|
|
