|
3, 2, 1: Health AI Brief
Every Friday
August 14, 2026
|
|
AI is reshaping healthcare fast. Below are 3 key AI developments, 2 studies, and 1 takeaway for this week to help you better lead with AI. Target read time: 5 minutes. |
|
3
Market Signals
CMS issued the formal notice on Aug. 7 for its RAPID coverage pathway: a proposed national coverage determination the same day FDA authorizes a participating device, with a goal of a final decision roughly 60 days later for Class II devices (90 for Class III). The conventional coverage process runs 9 to 12 months once it starts, and a 2023 study put the gap between authorization and Medicare coverage at a median of 5.7 years for novel devices needing a new coverage pathway. My Take
This 60‑90 day clock proposal is great, but only if that's enough time for CMS to truly figure out what's actually worth covering in the first place. The OpenAI Foundation announced $100 million on Aug. 13 for the Breakthroughs to Follow‑Through initiative with the Common Health Coalition: funding plus technical assistance for states to move health AI from pilot to deployment. Alabama, Illinois, Louisiana, and Massachusetts go first, with 4 more jurisdictions expected within 6 months. The opening focus is hepatitis C, using AI tools to find patients who have fallen out of care, with a stated goal of at least doubling cure rates in participating states within 2 years. My Take
OpenAI's foundation has moved from giving public health free AI to paying for AI deployment. And doubling hepatitis C cure rates is a valuable, tangible target with a tight 2‑year clock. We need more of this. New Orleans‑based LCMC Health will deploy Qualified Health's GenAI platform across care delivery, operations, and revenue cycle for 15,000 clinicians and staff at its 8 hospitals. The announcement leads with the controls: secure access, patient‑consent safeguards, enforceable governance, and real‑time performance monitoring. Qualified Health raised a $125 million Series B in March. My Take
Calling out consent safeguards and real‑time monitoring in the press release is progress. This ultimately creates a procurement template that other vendors will increasingly need to satisfy, too. |
|
2
Research Studies
Machine‑learning models trained on 1.93 million periodic health assessments from 668,684 US Army soldiers (2015–2019) predicted 24‑month outcomes with an AUROC of 0.81 for nonfatal suicide attempts but 0.72 for suicide deaths (0.5 is chance; 1.0 is perfect). The highest‑risk 10% of assessments captured 46.5% of attempts; for deaths, only the top 5% showed meaningful risk elevation, and that group captured just 18.6% of them. The two risk scores were only modestly correlated: 11.9% of assessments flagged elevated risk for at least one of the two outcomes, but only 3.1% for both. My Take
This study showed strong prediction for attempts, weaker predictions for deaths, and clarity that addressing each of these needs requires separate tools and models. Researchers ran 100 gastrointestinal oncology cases with validated multidisciplinary tumor board decisions through GPT‑5 under 5 zero‑shot prompting frameworks (a simulated tumor board, multi‑expert deliberation, and 3 specialist personas) plus a majority‑vote ensemble. Concordance with the tumor board ranged from 78% to 87%, a spread that was not statistically significant across the evaluated frameworks. Specialty‑specific language appeared in 97–100% of outputs regardless of accuracy. The authors position the model as decision support, with humans making the final call. My Take
78‑87% concordance regardless of prompting means the performance ceiling lives in the model. Prompt engineering (i.e., invoking the AI to be a specialist) only polished the words on the screen. The decisions didn't change. |
|
1
Key Insight
Medicare wants to pay for AI faster, and per use.
On Aug. 7, CMS issued the formal notice for its RAPID pathway: Medicare coverage roughly 60‑90 days after FDA authorization for participating breakthrough devices, against a gap that a 2023 study put at a median of 5.7 years. Six days later, STAT detailed how existing new‑technology add‑on payments already pay hospitals for eligible inpatient cases where a qualifying AI device is used. Takeaway
Fee‑for‑service built modern healthcare's volume (sans quality) problem. Paying per use is rebuilding this problem, just at the speed of AI. We need outcome‑based guardrails here, too. |
|
Know someone who'd find this useful? Share |
CMS wants a 60-day clock for AI coverage
3 signals, 2 studies, 1 insight for health leaders
Aug 15, 2026
•
3 min read
