HN user
Data for good.
Co-Founder at Empirical Health (https://empirical.health). Don't die of heart disease.
Before: Co-Founder @ Cardiogram (ML for heart health)
CTO at Sift Science (YC S11, machine learning to fight fraud)
Data Science @ UCSF Cardiology
HealthCare.gov rescue team
Google (Android speech recognition, search ads ML)
brandonb.cc
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.
Likely uses pulse transit time and the shape of the pulse wave to infer changes in blood pressure. I wrote a bit about how this works (in the context of Apple Watch's hypertension notifications) here: https://www.empirical.health/blog/apple-watch-blood-pressure...
GLP-1s will cost $50 / month for those eligible.
To be eligible, you need either a BMI of >=35, or a BMI of >=27 and a set of specific health conditions (uncontrolled hypertension, chronic kidney disease, pre-diabetes, etc). You also can't be qualified for GLP-1s under Medicare's previous Part D coverage.
In practice these criteria are narrower than they look. Of the 13 million Medicare beneficiaries with overweight or obesity, about 4 million actually qualify, because most are excluded for already having a diagnosis, like type 2 diabetes or sleep apnea, that covers a GLP-1 another way.
(OP here) Happy to answer questions!
Very sad news. :(
Right now lots of good health companies hiring, including my own (Empirical). https://www.workatastartup.com/ lets you filter YC companies to see health startups, by stage and location (or remote).
IMO, there's an interesting opportunity for AI to make healthcare deflationary.
For example, Medicare is launching a new program in July that pays a fixed rate for achieving defined outcomes, like lowering blood pressure or cholesterol. Medicare's explicit goal here is to create incentives to automate the repetitive parts of care delivery with software. (Much of preventive cardiology is surprisingly algorithmic and guideline-driven, so this is more plausible than it seems.)
This reverses the incentives of the current system, where CPT codes incentivize doing more "stuff" (but not necesarily delivering the most effective care efficiently).
If you're a software engineer who cares about health, and have been sitting on the sidelines till now, I think the next few years are a really interesting time to make a contribution.
It's pretty well-established science now that vitamin D is a hormone, not a true vitamin. Vitamin D binds a nuclear receptor that regulates roughly 1,000 to 2,000 genes (5-10% of the human genome).
The "Vitamin D" moniker has just stuck around since it was named in 1922.
TARGET-D is an in-progress study that supplements vitamin D based on blood levels (your idea).
Another more recent trial (TARGET-D) is showing a 52% reduction in heart attack risk: https://www.empirical.health/blog/vitamin-d-heart/
That trial used a dynamically-adjusted dosage of a vitamin D3 supplement, where dosing was set as to keep blood levels within a target range of 40–80 ng/mL. IMO part of the reason this trial is showing better results than the previous clinical trials of vitamin D supplementation quoted in the above article is that vitamin D has bad effects if too low and too high. Adjusting the dose dynamically to achieve an optimal range gets you the benefits without some of the negative effects.
Biggest changes in the new guidelines from the American Heart Association / American College of Cardiology:
- Universal Lp(a) testing recommended for everyone - Lp(a) is the strongest hereditary risk factor for heart disease.
- Risk Equations switched to PREVENT, which predicts both 10-year and 30-year risk.
- Treatment is now recommended for younger adults, based on these 30-year risk scores.
- CAC scans recommended in more cases (for intermediate risk).
- Specific LDL targets are back, after being removed in the 2013 guideline
FYI - when I view the page (from New York), the language shows up in French.
This paper is being published at ICLR 2026 (top AI conference), and was selected as one of three outstanding papers.
(OP) The science behind eggs being healthy, or at least not harmful for heart health, has been pretty settled for decades.
Unfortunately, official medical guidelines take a while to catch up. It's only in 2026 that the American Heart Association put out updated dietary guidance which makes it official that most people shouldn't limit dietary cholesterol. Fiber and saturated fat are more important drivers of blood cholesterol, which is still recommended as a major risk factor (alongside blood pressure, inflammation, HbA1c, and so on).
The post also tries to explain why there was a limitation on eggs and cholesterol in the first place, starting from the 1968 guidelines.
The graph y axis is showing fractions (i.e., 0.21 is 21%). Sorry for the slightly confusing label.
For most of these metrics, zero is not a logically possible data point. For example, somebody with an HbA1c of 0 would be dead.
The existing CPT codes (roughly) pay proportionately to physician time (RVUs). So I wouldn't say there's an an incentive toward delivering care efficiently, but rather hospital management wants to maximize billable hours.
Medicare is a government-run insurance program, so this is one of the few cases where a private insurance company wouldn't receive data.
(There is such a thing as Medicare advantage, where a patient can choose to put their Medicare dollars toward private insurance, but it's not part of the initial launch of this program.)
Isn't the first em dash taken from an interview that the writer did with the subject over Zoom? I think using an em dash to punctuate a broken or partial sentence like that is pretty standard journalistic practice when you don't want to modify the original quotation (e.g, denote a paraphrase with brackets), and definitely not an AI tell.
The other uses are honestly pretty standard rhetorical patterns; they do not seem especially AI-flavored to me.
I run a YC startup that was accepted to Medicare ACCESS.
Historically, insurance has paid for activity: time spent in visits, RVUs generated, and minutes logged. This was a reasonable starting point, but the flaw is that there's no strong incentives to be efficient.
ACCESS is explicitly a "deflationary" approach. Medicare has set the payment rates high enough to be viable for startups, but low enough that you have to use software (including AI) to deliver a large part of your program.
So Medicare has basically created economic incentives to reward software without prescribing the exact shape of the programs. I thought it was a really interesting approach and builds on 15 years of lessons from CMMI (Medicare's innovation group).
Lipoprotein(a) is the strongest hereditary risk factor for heart disease. Each Lp(a) particle is basically an LDL cholesterol particle with an extra wrapper protein.
This statement is from lipidologists (basically, people who study cholesterol), but the American Heart Association released similar guidelines in Mar 2026 which recommend Lp(a) testing for everybody.
Only 1 in 400 people test Lp(a) today, although that's up 22x in the last decade: https://www.empirical.health/blog/lpa-testing/
Four major drugs are in development that target Lp(a) specifically; one of them lowers it by 94%.
Previous generations might have said the same thing about Ableton itself, vs playing a physical instrument. In that regard, AI might become just another power tool for creative expression.
This study tested a relatively new drug, PCSK9 inhibitors, which lower LDL cholesterol/ApoB above and beyond what's possible with statins and ezetimibe.
The patient population was 3600 people with high-risk diabetes, but not atherosclerosis. So they're at elevated risk compared to the average person.
PCSK9 inhibitors are still expensive (about $1,800 per year), with pretty limited insurance coverage. But this will likely change as the evidence builds.
OpenAI's spinning up in deep RL is free and pretty good: https://spinningup.openai.com/en/latest/
It includes both mathematical formulas and PyTorch code.
I found it a bit more practical than the Sutton & Barto book, which is a classic but doesn't cover some of the more modern methods used in deep reinforcement learning.
The basic principal behind this shift is that exposure to bad cholesterol acts like radiation dose. Both the intensity (how high LDL/ApoB is) and the time exposed (how many years) matter.
These newest guidelines from the AHA/ACC, released yesterday, operationalize this partly by using a new set of equations that predict cardiovascular risk over a 30-year period (rather than 10 years).
Archive link: https://archive.is/KXksL