Week of June 22, 2026
Clinical guidelines are a foundation for patient care, and, increasingly, the material AI engines draw on to make their decisions. Last week, two of my patients reported complications from appropriate, guideline-supported medications that work partly by lowering blood pressure. One felt profoundly weak, alarmed enough to stop the medication. The other tolerated it for days, but fainted on day 5.
There is a mental screeching halt when patients experience complications from something I prescribe. Two in one week drives a deluge of questions and self-doubt. Did I prescribe the wrong medication or dose? Did I fail to warn them of the early signs of intolerance? Did I miss something in their history or some quirk of their physiology that would have predicted this? In this case, the answers are no. I think I did the right things — followed the guidelines, counseled the patients, and was thoughtful about each patient’s history, and yet, my patients still had complications. I wonder how clinical decision support tools feel about outlier events in highly variable humans.
The Google Notebook LM AI-generated podcast version of this week’s newsletter.
Science and Technology Trends
The FDA recently approved a new kind of cancer drug, one that shreds a harmful protein instead of merely blocking it. The class of medications called PROTACs (proteolysis targeting chimera) was first described in 2001, shown to work in 2008 with a small molecule, and reached patients for the first time this year, treating one type of breast cancer.
Most drugs work like a key in a lock: they bind to a target protein and block its active site, switching the cell's behavior on or off. To work, the drug must stay attached to the cell’s protein. A PROTAC does something different - it is a two-headed molecule, one end grabs the disease-causing protein, the other grabs the cell's own garbage-disposal machinery, holding them together so the targeted protein gets tagged for destruction and broken down. The cell keeps living; it just loses the problem protein.
That mechanism brings three real advantages. A single PROTAC molecule can destroy target after target; it's recycled after each one rather than used up. It's highly specific, so it's less likely to harm healthy cells. And because it doesn't need a receptor to bind, it can reach proteins that other drugs can’t.
This first PROTAC drug isn't a miracle. Clinical trials demonstrate that it slows cancer progression by about 3 months, and only in patients with a specific type of cancer-causing mutation. But it proves this class of medication works.
The work was largely carried out at Yale (with partnerships with private companies), representing a sustained 25-year effort that yielded a new drug class. It's a solid argument for long-horizon funding and basic discovery that takes decades and millions of dollars.
Nature editorial discussing the development of PROTACS
Background on PROTACs
Open Evidence Review - Mechanisms of Chemotherapy (including PROTACs)
AI Analysis
Midjourney, the AI-generated image company, is pivoting (expanding? growing?) into medical imaging. They recently announced medical spas that “allow anyone to step into a shallow pool of (golden lit) water where a ring of underwater sensors sends ultrasonic sound waves through the person’s body from every angle. Then, using AI, the sound waves are reconstructed into a multiplanar, full-body image, like an MRI, but in 60 seconds and without magnets, radiation, or dye. To say the least, I am highly skeptical. My concern is amplified by Midjourney’s statement that they hope people will “[make] this kind of medical imaging commonplace, a casual, everyday, literally ‘whenever you want’ kind of thing.”
- The physics may or may not work. Science communicator Hank Green did a nice job of breaking down the limits of ultrasound data, golden light notwithstanding.
- Casual imaging of your whole body is a great way to find all sorts of false positive findings - i.e. asymptomatic cysts, nodules, stones, and poorly defined findings that are often benign but generate follow-up appointments, imaging, and consternation. Hank mentions this as part of the total-body screening industry. See the Radiological Society of North America’s (RSNA) 2023 commentary on the challenges of direct-to-consumer total-body screening with CTs and MRIs.
I give Midjourney some credit. As AI imaging is increasingly commoditized, expanding its AI tools toward more consumer-driven healthcare imaging makes (some) business sense, even if it makes no scientific sense.
MidJourney announcement
Analysis by Hank Green
RSNA commentary on elective full-body imaging
AI-Assisted analysis of all of the above
Anti-Anti-Science
No matter how much one wishes science away, it is, as they say, what it is. Like peeling paint in an algae-filled reflecting pool, cosmetic fixes tend to hide underlying problems only temporarily.
The Wall Street Journal covered the difficulties Mars, the maker of M&Ms, is having in transitioning its candy to more natural dyes. (See my 1-20-2025 issue of What Adam is Reading for a discussion of the knowns and unknowns of dyes, including the limited available data.) While moving to natural dye seems reasonable on the surface, this is a very expensive makeover for an unclear amount of actual benefit. I wonder what the Mars family says behind closed doors.
WSJ Article
https://www.wsj.com/business/mars-is-spending-millions-to-give-m-ms-a-maha-makeover-2fa1bb88?st=QQENXF&reflink=desktopwebshare_permalink
AI Analysis of the WSJ article
https://claude.ai/public/artifacts/68688d3a-9c9e-4e71-bd22-8a040560a306
https://www.wsj.com/business/mars-is-spending-millions-to-give-m-ms-a-maha-makeover-2fa1bb88?st=QQENXF&reflink=desktopwebshare_permalink
AI Analysis of the WSJ article
https://claude.ai/public/artifacts/68688d3a-9c9e-4e71-bd22-8a040560a306
Screwworm, a flesh-eating fly larva that infests wounds in livestock and other warm-blooded animals, is moving north from Mexico into Texas. I hadn't realized the U.S. government had run a sterile-fly program for decades: male flies are sterilized and released by the hundreds of millions, and because females mate only once, those matings produce no offspring, collapsing the population. From the 1960s onward, this pushed the parasite out of the U.S. and, by the 1990s, all the way down to the Darién Gap in Panama, where a permanent barrier of inspection and sterile-fly releases held the line. That line broke with a 2023 breakout in Central America, and the parasite has since marched northward, with the first U.S. case confirmed in a south Texas calf on June 3. Funding and staffing cuts have made monitoring and control far harder, and with surveillance degraded, the true extent is unclear, though confirmed cases now span cattle, a goat, and at least one dog across south Texas. You can't know if you don't measure.
Washington Post Article:
https://wapo.st/4w5DhVU
AI-Assisted Analysis
https://claude.ai/public/artifacts/8cfb3037-1fbd-4fdc-9f4f-3341ae4dce41
https://wapo.st/4w5DhVU
AI-Assisted Analysis
https://claude.ai/public/artifacts/8cfb3037-1fbd-4fdc-9f4f-3341ae4dce41
Measles is still on the upswing, thanks to fewer vaccinated children. You will recall that a population needs about 95% vaccination for measles herd immunity to hold, given how contagious it is. The MedPage Today outbreak map is, unsurprisingly, the vaccination-rate map turned upside down: the cases cluster exactly where coverage is lowest.
MedPage Today:
https://www.medpagetoday.com/infectiousdisease/publichealth/119081
Yale School of Public Health Measles tracking website
https://ysph.yale.edu/research/department-research/epidemiology-of-microbial-diseases/research-themes/disease-areas/measles/#featured-ysph-involvement:-virtual-medical-operations-center-special-reports
Yale School of Public Health Measles tracking website
https://ysph.yale.edu/research/department-research/epidemiology-of-microbial-diseases/research-themes/disease-areas/measles/#featured-ysph-involvement:-virtual-medical-operations-center-special-reports
AI Impact
Eric Topol reviewed two recent studies in Nature examining how well clinical agentic tools (AI) manage patients across primary and emergency care. Both agent systems, at different institutions and set up somewhat differently, were able to beat human physicians in diagnosis, guideline alignment, and medication safety. However, one was a fully sandboxed EHR in the emergency room, and the other was an LLM-based agent working on simulated outpatient visits compared to human reviewers. Either way, it is clear we are getting closer to, if nothing else, co-management of patients with agentic tools running in the background of our electronic health records. There is still considerable variation from state to state in the United States in how this would be managed, and, of course, having agentic systems will change the landscape of medical liability. For now, physician and advanced practitioner medical licenses serve as a backstop against poor decision-making.
Eric Topol’s analysis:
AMIE, a multi-LLM-based agent system optimized to care for virtual primary outpatients over time:
MIRA - a sandboxed agentic emergency room tool caring for patients across an ER encounter:
AI-assisted Analysis of all of the above:
Nature also published a related news piece about deskilling. It walks through two studies (not in Nature) suggesting that AI support degrades professional skill. In one study, gastroenterologists who had grown accustomed to AI-assisted colonoscopy detected fewer precancerous adenomas once the tool was removed, with their unassisted detection rate falling from 28.4 percent to 22.4 percent. In the other, IT workers who leaned on a coding assistant scored worse on a comprehension quiz than those who coded by hand, 50 percent versus 67 percent.
As you dig deeper, though, more nuance surfaces, at least among the coders. Anthropic, which studied the coders, found that developers who used AI as a thought partner, asking it to explain concepts and verifying its output, retained their learning and performed comparably to the no-AI group. Skill degradation occurred in those who simply had AI write the code for them.
Of note, the colonoscopy study is observational rather than randomized with respect to the deskilling question, so confounding remains possible. The coding study is a non-peer-reviewed preprint, with a small sample size.
Also of note, my Claude analysis carried a caveat about bias, since Claude was writing about a study by the company that built Claude (for the second time in two weeks in the newsletter).
AI Analysis (from Claude)
NEW The J&E Random Kidney Facts of the Week (Better know your kidneys!)
Two loyal readers suggested I salt the newsletter with random kidney facts for the nephrologic edification of the readers.
Your kidneys are a key part of the body's blood pressure sensor and feedback loop. When they sense a drop in blood pressure, they release an enzyme called renin, which triggers a cascade that produces the hormone angiotensin II. Angiotensin II does two things: it constricts your blood vessels, and it tells your adrenal glands (which sit on top of your kidneys) to release aldosterone, a hormone that makes the kidneys hold onto salt, and with it water. More constriction, more fluid, higher intravascular pressure. Meanwhile, your brain pitches in with ADH, which retains even more water and further constricts blood vessels.
These signals are always on, and there are countervailing ones telling your body to dump salt and water, too. It's less a switch than a room full of people shouting, where the loudest voice wins. Many of our most effective blood pressure medications work by dialing down the volume of the “constrict blood vessels and hold onto salt” voices, by blocking Angiotensin II, aldosterone, or the cellular pumps that take up salt or water. (And, by the way, occasionally, patients on these medications have exaggerated responses that can cause fainting or feeling poorly.)
Things I learned this week
Thanks to Scotland's participation in the World Cup, the 20,000 members of the Tartan Army (a nickname for Scotland men's national football team fans) descended upon Boston to watch the game. There were numerous reports of bars in the area being "completely running low" on beer. Thanks to some quick AI questions, it appears that there are about 1,200 bars in Boston, each with an average of about 245 gallons of beer on hand. This means that each member of the Tartan Army would have to drink between 14 and 15 gallons of beer to completely empty every pub. This would equal about 117 pints per person during a World Cup game.
I dove into the breadth and depth of technology- and employment-related fear this week. In 1961, with the rise of IBM mainframe computers automating intellectual business tasks, Time Magazine wrote a compelling article on how automation would leave many previous workers without jobs, how retraining and unemployment would continue to be increasing issues, and how an entire subclass of the permanently unemployed would emerge. There are even hints of a universal basic income discussed. Fast forward, we have been fearing machines for more than a century now:
- 1961 article — https://time.com/archive/6624989/business-the-automation-jobless/
- 1965 — TIME's The Cybernated Generation cover (mainframe-era "mass leisure class" / 20-hour-week forecast)
-
1980 — TIME's The Robot Revolution cover (industrial robotics; the tentacled-robot-strangling-workers image)
- 1982 — NYT's A Robot Is After Your Job (the white-collar turn)
- 1983 — TIME's Machine of the Year: The Computer Moves In (the PC, where the framing flips from threat to celebration)
Overall review of thoughts on the the impact of computers from the Pessimists Archive (1920 onward):
Once this intellectual horse left the barn, I kept riding down the road.
-
1895 and 1913 — https://quoteinvestigator.com/2021/07/17/auto-fad/
Edison tells the New York World, "the horse is doomed," predicting a cheap horseless vehicle within a decade. And A Country Gentleman piece argues that for short-haul, many-stop delivery, the gas wagon "can never displace the horse," even doubting enough gasoline exists. - 1899-12 — https://pmc.ncbi.nlm.nih.gov/articles/PMC9335811/
"The Passing of the Automobile," a veterinary trade journal scolds the media for hyping the "horseless age" and defends the horse as "a legitimate field of investment." - 1902-08 — https://pmc.ncbi.nlm.nih.gov/articles/PMC9336273/
"Uncle Henry on Passing of the Horse," a satirical poem, recounts how the trolley, bicycle, and Edison's batteries each "killed" the horse, each refuted by survival. - 1915 — https://www.saturdayeveningpost.com/2017/01/get-horse-americas-skepticism-toward-first-automobiles/
Harness dealer Ed Klein's "Horse vs. Automobile" newspaper ad runs the cost comparison and lands on depreciation: your horse holds its value while a car is "already sliding toward obsolescence." - 1890–1920 — https://blogs.microsoft.com/today-in-tech/day-horse-lost-job/
A modern retrospective quantifying the collapse of the horse economy: U.S. carriage builders fell from 13,800 firms to 90, taking the farriers, harness-makers, and feed merchants with them.
AI Analysis of fear-mongering of lost jobs due to automobiles through time:
All of this is to say, I would be skeptical of doom and gloom scenarios. Work evolves, jobs change, and humans find plenty to do. Horses are still around too, I note.
AI art of the week
A visual mashup of topics from the newsletter, and an exercise to see how various LLMs interpret the prompt. I use an LLM to summarize the newsletter, suggest prompts, and generate images with different LLMs.
A visual mashup of topics from the newsletter, and an exercise to see how various LLMs interpret the prompt. I use an LLM to summarize the newsletter, suggest prompts, and generate images with different LLMs.
A vintage 1950s pulp science-fiction magazine cover, in the lurid style of Frank R. Paul and golden-age Amazing Stories / Galaxy covers. Aged, slightly foxed newsprint paper with off-register color printing, visible halftone dots, saturated cyan-magenta-yellow palette, dramatic airbrushed glow. Painterly gouache/airbrush texture, NOT digital-smooth. Portrait 2:3 aspect, like a magazine on a newsstand.
MASTHEAD across the top in towering chrome-and-red retro-futurist letters: "AMAZING MEDICINE." A smaller banner beneath: "What Adam Is Reading." A circular "25¢" price burst in one corner and "VOL. XXVI · No. 6."
CENTRAL SCENE: A colossal, chrome-and-tentacled robot, a tiny "A.I." badge on its chest, looms over a gleaming medical spa where a swooning patient in a hospital gown floats in a shallow pool of golden, glowing water, ringed by humming ultrasonic sensors firing golden rays through her body. In the immediate foreground, a second careful-looking patient is mid-faint, one leg in a cast, a spilled bottle of pills rolling away. Tiny white-coated doctors flee in panic, dropping their charts.
BACKGROUND: a cracked Texas plain where a monstrous segmented screwworm larva erupts from the soil, jaws gaping; a small herd of cattle and a single startled horse stampede past an early vintage automobile. Off to one side, a small green two-headed molecule (a PROTAC) drags a crumpled protein toward a chrome garbage-disposal hatch.
PULP COVER BLURBS scattered in jagged yellow starbursts and red banners:
- "THE DESKILLING! When the Machines Forget How to Teach Us!"
- "INVASION OF THE FLESH-EATING FLY — From Mexico... to YOUR Ranch!"
- "The Golden Pool That Sees ALL — Miracle or Menace?"
- "PLAGUE RETURNS! The Map They Tried to Flip!"
- "20,000 Thirsty Scotsmen DRAIN a City Dry!"
Composition: dense and dynamic, diagonal action lines, deep shadow with a single hot light source, campy melodrama, exclamation-point energy.
Clean hands and sharp minds,
Adam
Comments
Post a Comment