The authors ran it both ways. Predicting who had microplastics: smoking won and everything else dropped out. Predicting who had obstructive coronary disease: microplastics held up and smoking didn't. With correlations this strong, you can really tell either story.

Where does it come from? Smokers had far more than nonsmokers. So did people breathing dirtier air. Put both together (a smoker in a high-pollution area) and detection hit 100%.

And it wasn't just detection. Concentrations were markedly higher in the STEMI group. Within those patients, the plastic load was elevated everywhere they looked, not only at the blocked artery. These people were simply awash in more plastic.

Microplastic studies have a contamination problem (plastic is everywhere, including the lab). These authors went hard: fume hoods, cotton coats, glass vials, blank saline runs, mass spec plus a laser method. The numbers look real.

61 patients at two Italian hospitals, all getting cardiac catheterization. 19 with a STEMI (a major heart attack), 20 with chronic coronary disease, 22 with completely clean arteries. Blood was pulled from the artery itself as well as a peripheral vein.

We've known micro- and nanoplastics get into the body. Showing they actually do harm has been much harder. This new study in the European Heart Journal is one of the more provocative attempts. buff.ly/1sEeCVq

They sampled blood straight from the coronary arteries of people mid-heart-attack. 84% had detectable microplastics. In people with clean coronaries, it was 32%. 🧵

So either we enforce the rules and close the compounding loopholes, or we admit that for some of these drugs, if everyone can get a prescription, maybe no one should need one.

Almost all of it was compounded, not branded. Compounding a copy of an available drug generally isn't allowed, so 60% of the prescriptions added a supplement (usually B12 or glycine). That's one way to argue you're not simply copying Ozempic.

Other signs the supervision was thin: the same provider writing scripts across multiple sites, prescriptions issued despite missing required photos, and a couple written in under five minutes. Only two sites noticed he'd ordered the same drug somewhere else.

Turns out, not much. Every site had a questionnaire. Only 55% asked about eating disorders. 13 of the 49 required a video visit and 3 required a phone call. The rest was forms.

The persona was built to be an easy yes: a 26-year-old man with a BMI of 35, hyperlipidemia, hypertension, prediabetes, sleep apnea, and no contraindications. So the question was never whether he'd get a prescription. It was how little anyone would ask first.

FWIW, these were early-2026 models (GPT-5, Gemini 3 Pro, Opus 4.6), and in this field, if you don't like a model's performance you just wait a couple weeks. They also ran them at temperature zero, which can quietly cost accuracy.

Run 100 discharges through the best model. About 48 trigger a pop-up. 9 are real catches. 39 are false alarms (you'll start tuning them out fast). And 2 patients bounce back that the AI never flagged at all.

But sensitivity is a double-edged sword. The model that catches the most real misses also flags the most patients who are perfectly fine. Most sensitive, least specific, and vice versa. Every model landed somewhere on that tradeoff.

Then they handed the same charts to the models and asked, simply, is something being missed here? Sensitivity varied wildly. Claude Sonnet 4 caught the most real misses (about 86% in the discharge cohort). GPT-5 mini caught the fewest (43%).

Not every bounce-back is a miss. Back pain that later turns out to be an epidural abscess (no fever, normal labs at first) isn't one. A DKA hiding behind an unaddressed anion gap is. Of 288 reviewed cases, 39 (13.5%) were genuine misses.

How do you even know a diagnosis was missed? They took two ED cohorts (patients discharged then readmitted within 72h, and floor patients bumped to the ICU within 24h) and had two physicians review each chart to decide if something was genuinely missable.

The catch is alert fatigue. Beep at a doctor enough and they tune the beep out, even when it's right. An AI safety net has to be accurate without firing on every patient. Six LLMs put to the test in JAMA Network Open: buff.ly/fWd94A9

It isn't airtight. Vaccinated people might avoid COVID in other ways (shot looks better than it is), or be frailer to start with (shot looks worse). So the authors adjusted for age, sex, race, region, and calendar time. The numbers barely moved.

In the data (179 hospitals, 7 states), 5% of the COVID-positive patients had gotten this year's shot. Among the test-negatives it was 12%. That gap is the whole ballgame: the vaccine cut medically attended COVID by about 50%, and hospitalizations by 55%.

The alternative they actually used works like this. Take people who show up sick at the ER or urgent care and get swabbed. Test positive for COVID, you're a case. Test negative, you're a control. Then compare vaccination rates between the two groups.

And even setting ethics aside, an RCT takes years. COVID and flu vaccines get retargeted to new strains every season. By the time a trial reads out, you're three variants and two formulations down the road. The question it answered is already obsolete.

The study is a test-negative analysis of how well this season's COVID vaccine actually worked, now out in JAMA Network Open. The official reason for spiking it: the method was supposedly inadequate to tell whether a vaccine works. buff.ly/drEs1QY

In April, the CDC blocked one of its own studies from publication. It had cleared internal scientific and editorial review and was slated for the March 19 MMWR. Then the acting director pulled it. This week it came out anyway (just not in a CDC journal). 🧵

The biggest surprise was diabetes itself. If metformin were going to win anywhere, surely it'd be here. Yet 60% of the lifestyle group developed type 2 diabetes over follow-up, vs 69% on placebo and 71% on metformin. The diabetes drug had the most diabetes.

And the gap widened the deeper they looked. At 3 or more conditions, lifestyle clearly pulled ahead. Overall that worked out to 10% fewer chronic conditions, and 43% lower risk of the costliest disease pairs. Metformin moved none of them.

First cut: almost everyone got there. 85% developed multimorbidity, and the raw rates barely differed by arm. But after adjustment, the lifestyle group had a real edge (HR 0.79). Metformin came out no different from placebo.

The new question wasn't any single disease. It was multimorbidity (2 or more chronic conditions), the slow pileup that actually defines aging. They read it straight out of Medicare claims, decades after the trial ended.

The original DPP enrolled 3,234 adults with prediabetes and randomized them three ways: intensive lifestyle change (7% weight loss, 150 min/wk of activity), metformin at 850 mg twice daily, or a placebo pill.