Back to the talk
Sources
Every number I put on a slide, where it comes from, and what it does not prove. If you find something wrong here, tell me and I will fix it.
AI in Healthcare: What's Working, What Isn't, and What to Watch
David Marek · El Paso Chamber Healthcare Coalition · August 12, 2026
Slide 3
Blood pressure, monitored at home
A 9.8 point drop in blood pressure after one year of monitored care at home.
- Source
- UC San Diego Health remote patient monitoring program. JMIR Cardio, 2025 (e75170).
- The detail
- Mean reduction of 9.8 mmHg in systolic blood pressure among patients enrolled for hypertension, p < .001.
- Worth knowing
- Vendors quote much larger numbers for the same category of program, often from press releases rather than journals. The peer-reviewed figures run roughly 10 to 17 mmHg. I used the lower published number on purpose.
Slide 4
Ambient AI in the exam room
77% of the visit spent in eye contact, up from 70%. 15% less time on paperwork per visit.
- Source
- Time and motion study at Singapore General Hospital. JMIR Medical Informatics, 2026 (e85580).
- The detail
- Eye contact rose from 69.6% to 77.1% of consultation time (p = .009). Documentation time fell 15.0% (p = .036).
- Worth knowing
- Nine clinicians, one hospital, an in-house tool. This is a real measured result and it points the right way, but it is not a large multi-site trial, and I would not plan a budget around it.
Slide 5
Breast screening read by AI first
29% more breast cancers caught, with no rise in false alarms.
- Source
- MASAI, a randomized trial inside Sweden's national screening program, published across three papers in The Lancet family of journals.
- The detail
- 105,934 women. Detection of 6.4 versus 5.0 cancers per 1,000 screened (p = 0.0021). False positives 1.5% versus 1.4%, which is not a meaningful difference. Radiologist reading workload fell 44.2%.
- Worth knowing
- MASAI also looked at interval cancers, the ones that surface between screens. That result was non-inferior but not statistically significant (ratio 0.88, p = 0.41). It is the number people most want to quote and it is the one the trial did not establish, so it is not in the talk.
Slide 5
Pancreatic cancer on scans ordered for something else
Three years earlier detection of pancreatic cancer, before a tumor is visible.
- Source
- Mayo Clinic's REDMOD model. Gut, April 2026.
- The detail
- A texture signature on routine abdominal CT, readable up to three years before a tumor mass is visible. The model found 73% of pre-diagnostic cancers against 38.9% for specialists reading the same scans, at a median lead time of about 16 months. AUC 0.82.
- Worth knowing
- This is a retrospective validation study, not a trial. It shows the signal was already in scans that had been taken. Mayo's prospective trial, AI-PACED, is underway. The open question is not detection, it is what you do with a patient who has nothing to biopsy yet.
Slide 6
Claims denied in bulk
1.2 seconds of doctor review per automated denial. 300,000 claims denied in two months.
- Source
- ProPublica reporting, March 2023, based on internal Cigna documents describing the PxDx system. Cigna disputes the characterization.
- The detail
- Claims were batched for rapid physician sign-off rather than individually reviewed. Related ERISA class claims were allowed to proceed in part in the Eastern District of California, March 31, 2025.
- Worth knowing
- These are a news organization's figures drawn from documents. They are not a peer-reviewed study and they are not a court's finding of fact. I attribute them to ProPublica every time I say them, and you should too.
Slide 6
Medicare's supervised pilot
Medicare is piloting a supervised version through 2031, and Texas is one of the six pilot states.
- Source
- CMS WISeR model (Wasteful and Inappropriate Service Reduction), running January 2026 through December 2031.
- The detail
- Six states: Arizona, New Jersey, Ohio, Oklahoma, Texas and Washington. Six participating vendors. Roughly 14 to 15 service categories are currently active, out of 17 originally proposed.
- Worth knowing
- This is payer-side review, which is a different thing from the provider-side automation people often describe in the same breath. The model is under active political pressure and faces possible congressional repeal following a GAO determination, so the 2031 end date is not a safe assumption.
Slide 7
A region's health, seen in real time
Public health used to learn about outbreaks weeks late. Now AI reads emergency room visits as they happen. And it finally speaks medical Spanish.
- Source
- There is no statistic on this slide, on purpose. It describes a direction, not a measured result.
- The detail
- Syndromic surveillance has moved from batched paper reporting toward continuous ingestion of emergency department and clinic data. Texas ran statewide surveillance of this kind through the State Medical Operations Center during the 2026 World Cup. Separately, transformer models can now do named entity recognition and normalization directly on Spanish-language clinical notes, which used to be a blind spot in US public health analytics.
- Worth knowing
- The failure mode here is the one this region should care most about. Models trained only on English notes from well resourced academic centers lose information when they meet a bilingual, safety net population. Translation on top of an English model is not the same as a model that was trained to read the notes we actually write.
What we're building next
Through DropDev I help organizations across the region put AI to work, in healthcare and well beyond it. Our healthcare venture, HealthAtlas, works on the payer side of this, which is the part of the system this talk was hardest on. It was named Best in Show at the HIMSS 2026 Emerge Pitch Competition in Payer Systems.
The people in that room are people I would like to know. If you want to put coffee or a longer meeting on my calendar, book a time here. I am mostly interested in hearing what you are actually working on and what is getting in the way of it.
Or reach me directly at david@dropdev.co or 915 234 1444.
Every number above was checked at its primary source before it went on a slide. Photography is licensed stock except the mammography image on slide 5, which is Google's own published image from NHS breast screening.
