The Morning Nap as a Risk Factor for Death Dr F Perry Wilson comments on a study that linked napping habits with mortality for older adults.

If a patient or loved one is napping while the clock still says AM, probably better not to sleep on it. Full commentary on Medscape: buff.ly/vLpQyPT

Bottom line: a post-lunch siesta tracks with a normal circadian dip and is probably fine. The signal is in the AM naps, when energy should still be high.

Still, I doubt a true causal link. When the authors excluded participants with mild cognitive impairment or dementia, the association largely disappeared. Napping may be flagging subclinical disease, not driving it.

Couldn't this just be sicker people napping more? The authors adjusted for nighttime sleep, BMI, depression, physical activity, chronic conditions, medication use, and disability. The signal persisted.

Timing was where it got interesting. Compared with people who napped in the early afternoon, morning nappers had a 30% higher mortality rate. Equivalent to being 2.5 years older.

The topline: each extra hour of daily napping was associated with 13% higher mortality. Each additional nap, 7%. Roughly equivalent to being 1.1 and 0.6 years older, respectively.

This wasn't self-reported sleep (which we know is unreliable). Participants wore wrist actigraphy for about 10 days, letting the authors quantify nap duration, frequency, variability, and timing objectively.

New in JAMA Network Open: in 1,338 older adults followed for up to 19 years, more daytime napping was linked to higher all-cause mortality. The biggest signal was naps in the morning...

Ready or Not, LLMs Are Coming for Medicine Current chatbots might show deficiencies in clinical reasoning for now, but don’t hold your breath.

11/ I break all of this down in my latest Medscape column.
buff.ly/VIYea2A

Large Language Model Performance and Clinical Reasoning Tasks This cross-sectional study evaluates the end-to-end clinical reasoning ability of off-the-shelf large language models using standardized clinical vignettes and introduces a multidimensional comprehens...

10/ We will look back at papers like this and wonder how we missed the signs.

Full study (open access):
buff.ly/P5brAFC

9/ Then we'll get AI triage lines --> AI taking initial histories in the ER waiting room --> an RCT showing non-inferiority to human docs --> FDA approval of the first AI agent doctor --> state authorizes AI agents to prescribe meds. Timeline? Probably sooner than you think.

8/ The authors say AI will "augment" human physicians. Sure. At first...

7/ The biggest gap: no human comparison data. These MSD vignettes are publicly available. I searched PubMed and every study using them was testing chatbots. Has anyone tried to see how doctors do on them? Is 95% final diagnosis accuracy good? It seems good.

6/ What bugs me: models with reasoning capabilities had reasoning turned off. Models that could search the internet were blocked from doing so. This was to "level the playing field" but like - don't we WANT our AI docs to search the internet and stuff?

5/ The authors created a new composite metric (PrIME-LLM) capturing balanced performance across 5 clinical reasoning domains. Grok 4, Gemini 3 topped the charts. Newer versions beat older ones across the board.

4/ But to get full credit, models had to flag EVERY correct option and exclude every incorrect one from a long list. That is a high bar for anyone.
Here's the list to choose from for a case of dyspnea... Getting it perfect is... tricky.

3/ The bad news, according to the authors: failure rates on differential diagnosis were 90 to 100%. Sounds terrible.

2/ LLMs nailed the final diagnosis more than 90% of the time. DeepSeek edged out others slightly but they were all bunched together. That is honestly better than I think most of us would do on these cases.

New in JAMA Network Open: 21 frontier LLMs tested on 29 clinical vignettes that simulate real diagnostic workflows. Not just "what's the diagnosis" but the whole process. Differential, testing, final dx, management.
The party line? They're not ready for primetime. But I'm not so sure...

LFG!!

We're all aging. I want to slow it down as much as you do. But this trial shows small, likely spurious effects on imprecise surrogate markers with no connection to clinical outcomes. Get your micronutrients from food.

Full write-up: buff.ly/n9fUjcV

The study, in Nature Medicine:
[LINK: buff.ly/dzmvQ8e]

The authors test this. They ask: does the cognitive benefit from multivitamins in COSMOS-Mind work through epigenetic clock changes? Answer: no. No significant mediation. If vitamins help your thinking, this isn't the pathway.

Even if real: is changing DNA methylation the same as "slowing aging"? Blood pressure predicts stroke and treating it helps. Grip strength predicts frailty but squeezing a stress ball doesn't. Which kind of surrogate are epigenetic clocks?

The authors say they didn't correct for multiple comparisons because the clocks are correlated. But they're not THAT correlated. And even with correlated outcomes, you need to control false discovery. We have ways to do this. You can't just test everything and pick the winners.

The two "hits": multivitamin slightly reduced PhenoAge and GrimAge increases at year 2. The PhenoAge difference was ~0.4 years over 2 years. That's 0.08 standard deviations. (0.2 SD is a minor effect typically).

The study tested 2 supplements x 5 clocks x 2 time points = 20 hypothesis tests. No single primary outcome. Of 20 tests, 2 crossed p<0.05. With 20 tests you'd expect 1+ false positive by chance. You'd see 2+ about 26% of the time. These are the real results.

We only have one calendar age. But we have LOTS of epigenetic clocks, all trained differently, and they don't agree with each other nearly as well as you'd hope. The scatterplots of one clock vs another look like shotgun blasts.

Quick primer: epigenetic clocks measure methyl groups that accumulate on DNA over time, like dust on a mantel. Algorithms use these patterns to estimate your "biological age." The gap between biological and chronological age has been linked to disease and death.

For this sub-analysis, 958 participants had DNA methylation measured at baseline, year 1, and year 2. The question: does a multivitamin or cocoa extract slow "biological aging" as captured by epigenetic clocks?