brandonb

The 4th generation heart rate sensor is quite interesting -- Apple Watch used to sample every ~5 minutes, and now it's every 5 seconds (basically continuous). I think that'll enable a lot of interesting health & wellness use cases.

UCSF puts on a "Mini Med School" series of lectures for the public. If you don't live near San Francisco, some of the video are online. One by Dr. Scheinman is here: https://www.youtube.com/watch?v=Fm3SwORdBS4

Good thing is that all of these conditions have screening and treatments, so you can do something about it.

This is a good addition. I'll update the article. While tadalafil is not FDA-approved for this indication, it looks like current American Urological Association guidelines do recommend it as an off-label use.

(Replied to a comment that was since deleted)

IMO 34.6% is a surprisingly high number, especially in that age range. Most men in their 30's would have a CAC of 0, so being able to observe soft plaque through these newer imaging modalities is telling us something genuinely new.

Whether that's good or bad isn't fully known. We'd expect it's bad: soft plaque is more likely to rupture than hard plaque.

But the study itself measured plaque in a cross-section of participants, at a moment in time. To get a definitive answer, we'd need to follow participants for several decades, observe a large number events (heart attacks, strokes, etc) and then see if soft plaque is predictive above and beyond conventional risk scores.

One thing that struck me about this particular study is that they used two different imaging techniques (CCTA and ultrasound) to measure soft plaque. (This is necessary to capture both coronary and peripheral arteries.)

Lepodisiran is a drug in clinical trials which reduces Lp(a), the strongest hereditary risk factor for heart disease.

This study analyzed the trial results and found it also reduced oxidized lipids, namely oxidized apoB. These are LDL cholesterol particles which are believed to be especially harmful.

Related and worth reading: The Third Culture by the former executive editor of Wired (when Wired was good): https://www.science.org/doi/10.1126/science.279.5353.992

This is roughly how GPUs for neural networks got started: after Andrew Ng left Google Brain, he no longer had access to a 10,000-CPU cluster used to train the original DistBelief system. But his Stanford students could buy a GPU...

Agreed that we can't do as much as we'd like. There are some general health things that are believed to prevent cognitive decline, such as managing cholesterol, HbA1c, blood pressure, sleep apnea, etc. One phrase I've heard is that cholesterol at age 50 predicts cognitive health at age 70.

But hopefully we'll start to have more options and wider approvals for things like anti-amyloid drugs.

This Alzheimer's blood test, PrecivityAD2, is based on the p-tau217 biomarker.

In one recent study, people very high p-tau217 had a 38% chance of progressing to cognitive impairment within 5 years vs 12% for those with low levels. The current tests cost about $200-300, so they're not unreasonable as a screening test. PrecivityAD2 looks to be priced around $1,400-$1,500 so at that price, this specific test likely only makes sense for people with established disease.

Curious-is there a particular diagnosis they should be compensating for in these algorithms?

(OP here) I worked on one of the first studies on detecting diabetes from wearables in 2018. It was kind of fun to review how the field has evolved. The basic approach is the same (pretrain a self-supervised foundation model on sensor data, fine-tune for a particular application) but the scale is several orders of magnitude larger.

Uncompensated care attributable from uninsured patients is ~2–3% of operating revenue/costs. So it's there but not massive.

Those are two real effects, but together they wouldn't compensate for the rate cuts.

About 8% of the population is uninsured. The uninsured population skews younger, with less healthcare utilization (Medicare already covers everyone 65 and older).

Another comment in this thread estimated billing overhead at 8.5%. Medicare for All would eliminate some, but not all of this, since Medicare is still a claims-based system. You would remove a lot of overhead around prior auths, which I agree is a good thing, but could be achieved with more focused legislation.

There's a sketch of part of the proof in "the Mellin ansatz." I'd definitely recommend reading the sections before it, though, since they attempt to establish the more foundational connection between sphere packing and harmonic analysis (which is necessary to understand OpenAI's proof).

I'd like to believe this, but the study makes a bunch of really hasty assumptions.

The authors derive the $1T number from $1.3T in total cost savings and $304B in incremental spend (incremental spend is due to insuring more people). The $1.3T in cost savings come from five big buckets: lower pharmaceutical prices, Medicare-level payments to providers, reduced administrative overhead, less fraudulent billing, and fewer avoidable emergency department visits and hospitalizations.

The buckets themselves don't necessarily survive much scrutiny.

Take "Medicare-level payments to providers". Hospitals have an operating margin of 2-5%. Medicare pays 50% less than private insurance. So doing this would require either layoffs, cutting salaries for doctors/nurses/etc, or both. This may well be the right decision for society as a whole--that's a big part of the debate here--but there's no free lunch.

The line item of "fewer avoidable emergency department visits and hospitalizations" assumes greater insurance coverage leads to greater access to primary care. It's true that great primary care prevents hospitalizations, and can be a net cost saving under certain assumptions [1]. But, we're actually in a primary care shortage. Existing insurance payments for primary care are low enough that private practices are going out of business and fewer residents are going into family medicine. Cutting rates (the paragraph above) would make this worse.

For "less fraudulent billing," a lot of people in the industry believe that Medicare has a large amount of undetected fraud. That's unfortunately the flip-side of reduced administrative overhead. The authors assume an 8% savings here, but the 2003 paper they cite uses the word "fraud" only twice and doesn't give a number.

Healthcare reform is hard.

[1] Reasonable breakdown on the economics of advanced primary care models: https://olearykm.medium.com/the-cost-equation-for-new-primar...

(OP) The first of ten problems OpenAI solved was an improvement to the Cohn-Elkies bound. I briefly worked with Henry Cohn on sphere packing in undergrad. This topic is actually surprisingly accessible to the amateur mathematician so I thought I'd write up a primer and accessible description of the actual result.

Can't rule out exercise.

Previous studies have shown that heart health at 50 predicts brain health at 75.

This study is interesting in that it quantifies the benefit as 12.5 extra dementia-free years. People with 0 risk factors had 30 years of dementia-free survival, vs 17.5 years for people with all 3 risk factors measured.

The risk factors in this specific study are blood pressure, diabetes (HbA1c), and smoking. You could reasonably also include cholesterol/ApoB, inflammation (hs-CRP), and perhaps Lp(a) as potential risk factors.

Blood pressure is a major input into the risk equations for heart attacks, strokes, kidney disease, and so on. For the most part we've used a one-time measurement, assuming it's an average. But if we had a time-series of blood pressure, I could see blood pressure variability actually becoming an important metric.

We've known for a while that inflammation is an independent risk factor for heart disease, alongside LDL cholesterol (or ApoB), Lp(a), and so on. Inflammation is most commonly measured by by hs-CRP or IL-6.

The ZEUS trial tested an anti-inflammatory injection (ziltivekimab). These phase III results showed a hazard ratio of 0.99, i.e., no effect.

This is evidence for a lipid-first model -- i.e., lipids drive oxidation, which then triggers inflammation.

(OP) The first oral macrocyclic peptide was approved by the FDA two weeks ago. It's a cholesterol-lowering drug, the first oral PCSK9 inhibitor -- current PCSK9 inhibitors are injected.

I found the underlying chemistry and scientific process really interesting.

I tried to answer two main questions in this specific article: 1. What was the actual process from 2013-2026 that led to enlicitide? 2. If we were restarting this work in 2026, would AI have helped? If so, at what specific stages?

Happy to answer questions if people have 'em.

FWIW, several YC startups have also published ML research--there were several of us at the last NeurIPS. So perhaps this is a niche that startups can occupy if the big labs don't.

This is cool -- I like the metaphor of AI as an sharp editor.

This is a cool new study from UCSD research group. 11–19% mARD is interesting, although not quite accurate enough to displace CGMs (the FDA-cleared ones have ~8% mARD).

But the sweat angle is could add signal alongside other techniques like Raman or MIR spectroscopy, and maybe a combination of these and an ML system would be accurate enough to use in practice.

Definitely agreed. I'm rooting for Apple, Samsung, Oura, etc to finally crack this problem and some of the techniques in the article (MIR, Raman spectroscopy, wearable foundation models) seem like they're leading in the right direction.

(OP) Updated the title. This was unfortunately a mis-fire of some of HN's automatic title formatting.

The lead author ran a study of personalized trigger mapping for Afib, using an n-of-1 study design: https://jamanetwork.com/journals/jamacardiology/fullarticle/...

Unfortunately this particular intervention wasn't successful (not much difference between treatment and control groups). But conceivably, afib triggers is something that varies from person to person and perhaps a future design would tease this out.

This is the first oral PCSK9 inhibitor; it cuts LDL cholesterol (or ApoB) levels by 50-60%. It uses a different mechanism than a statin so you can layer them to get an 80% or so reduction overall.

It also reduces Lp(a), the strongest hereditary risk factor for heart disease, by 28%.

Previous PCSK9 inhibitors like Repatha were injectables (similar to GLP-1s). Only about 1% of people eligible for injectable PCSK9 inhibitors use them, so having a convenient daily pill is a potentially huge win for prevention.

The speed of the pressure wave is one signal that correlates with blood pressure. It's a bit like a string being pulled taut -- waves travel faster with higher pressure. The shape of the wave also gives clues. For example, the rise time of the wave tells you something about the resistance encountered, which is a function of blood pressure.

Rather than hand-engineering these features, most modern systems are built on a wearable foundation model that's been trained reconstruct the signal (similar to how an LLM is trained to predict the next word). Those foundation models are picking up on these signals and likely others.

You're right that calibration with a cuff is required of all systems currently on the market.

The Signal Ring folks' claim they can do a blood pressure number without calibration, which is quite novel and seems to be their "secret sauce". They did run a clinical study as well, so presumably more details will come out whenever that's published.