brandonb

Part of the gap is that men develop heart disease 7 years earlier: https://www.empirical.health/blog/men-vs-women-heart-disease...

I think your second sentence is the strongest -- it's specific about what you actually help with.

Right now, the title and preview text on mobile probably look like "Quick scheduling question" / "Hi [name], I've followed [company] for...". This looks like a lot of other emails in my inbox that are requesting a meeting time. So I think a lot of people who quickly scan their inbox would pass this by and not even open the full email.

I'd also drop the "re: " part unless it's genuinely a reply. It breaks trust.

Every now and then, I think it's useful to look at what great work looked like in its first form.

In retrospect, this result was the first step toward the LLMs that are now the dominant approach in AI.

In 2017, this was a very modest result and probably didn't seem very promising. It took real courage and tenacity on OpenAI's part to keep at it for the next 5 years.

This is the first new paper from Alec Radford since leaving OpenAI. Token-level data filtering is kind of a simple idea, but so are many effective ideas in LLMs.

One advantage is that this type of safety guardrail can't be undone by an adversary in post-training, so it's a good fit for open source models.

The experiments are all done in preventing models from acquiring medical capabilities, while preserving related capabilities like e.g., biology.

> My doctor said she’s seeing a huge number of patients coming back with excessively high Vitamin D levels after taking supplement doses recommended by influencers. It happened to me, too, with what I though was a conservative dose of Vitamin D (5K IU, not even taken every day)

IMO that’s part of what’s interesting about this study design — they tested vitamin D blood levels and adjusted the supplement dose based on that. This seems like a much better approach than taking a high dose blindly.

I think the headline is accurate. The 52% number is from the experiment arm (participants who received a vitamin D supplement, with the quantity guided by blood testing). While it’s technically possible for the supplement dosage to be calculated as zero, 85% of participants were deficient at baseline, so this isn’t the main effect.

The 365º view of your heart health was supposed to be a joke. Although not necessarily a good one. :)

The study was presented at the AHA scientific sessions; full manuscript isn't out yet. It's in the caveats section in the article.

Cool! Is this the one? https://pubmed.ncbi.nlm.nih.gov/37302737/

Can you paste an example title and email here? I'm a founder and get 100's of emails per day. I often reply to cold emails, but honestly, with 90% of them, I can't really tell if they will solve one of my top 3 problems.

This was a write-up of a new study (TARGET-D) that used vitamin D supplements -- with the supplement amount guided by blood testing -- to reduce heart attack risk.

I've been working in heart health in 10 years and I was surprised at the magnitude of the effect here.

I hope it holds up as they move toward the final publication. Vitamin D supplementation is cheap and this could have a huge benefit.

NASA got its support in much the same way during the space race. Spreading the jobs widely is a good way to get political support.

One source: https://www.escardio.org/communities/councils/cardiology-pra...

Definitely shows the comparative power of medications. Statins, ezetimibe, and PCSK9 inhibitors can reduce LDL or ApoB by 85-95%.

The lack of self-consistency does seem like a sign of a deeper issue with reliability. In most fields of machine learning robustness to noise is something you need to "bake in" (often through data augmentation using knowledge of the domain) rather than get for free in training.

FWIW, Apple has published validation data showing the Apple Watch's estimate is within 1.2 ml/kg/min of a lab-measured Vo2Max.

Behind the scenes, it's using a pretty cool algorithm that combines deep learning with physiological ODEs: https://www.empirical.health/blog/how-apple-watch-cardio-fit...

We trained a foundation model specifically for wearable data: https://www.empirical.health/blog/wearable-foundation-model-...

The basic idea was to adapt JEPA (Yann LeCun's Joint-Embedding Predictive Architecture) to multivariate time series, in order to learn a latent space of human health from purely unlabeled data. Then, we tested the model using supervised fine tuning and evaluation on on a bunch of downstream tasks, such as predicting a diagnosis of hypertension (~87% accuracy). In theory, this model could be also aligned to the latent space of an LLM--similar to how CLIP aligns a vision model to an LLM.

IMO, this shows that accuracy in consumer health will require specialized models alongside standard LLMs.

Archive link: https://archive.is/mJSEh

Pretty extensive discussion here: https://news.ycombinator.com/item?id=46707572

"Fibermaxxing" is admittedly a silly term, but not only does a high fiber diet reduce cardiovascular mortality by 26%, it also reduces risk of cancer by 22%. Your grandparents were right!

Long Covid patients often face post-exertional malaise (PEM).

One of the main treatments for Long Covid is graded exercise therapy. This works for some subset of patients. But for other patients, it actually makes things worse. Right now we have no objective test that tells who is in what category.

I like this study since it identifies a specific mechanism of why PEM happens. And maybe that will lead to a more objective test for different sub-types of long covid, so that we can actually prescribe the right treatments to people.

Congrats on the launch. I always love to see smart ML founders applying their talents to health and bio.

What were the biggest challenges in getting major pharma companies onboard? How do you think it was the same or different compared to previous generations of YC companies (like Benchling)?

The thing that took the most time was normalizing biomarker names and units across labs. Even for the same lab chain (say, Quest), you'll get the same biomarker with slightly different names (e.g., Lp(a) vs Lipoprotein(a) vs Lipoprotein a) or units (e.g., cells/uL vs 10^9/L).

Lepodisiran has reduced Lp(a) by 94% in clinical trails. After it gets approved, IMO the author should re-write this section.

I started a company that does exactly that (except we also have doctors who can prescribe the medications, not just LLMs). So I don't think the approach you describe is naive, but others might. :)

I think there's a a bit of a paradox here: cardiovascular disease is solved biomedically, yet still remains the #1 cause of death worldwide.

From a biomedical standpoint, we have highly accurate biomarkers (e.g., ApoB, Lp(a), hs-CRP), long-term risk prediction models, knowledge of nutritional biochemstry, and next generation drugs like PCSK9 inhibitors and lepodisiran that can lower ApoB and Lp(a) by 90%. So there's no fundamental reason why cardiovascular disease has to be in even the top 10 causes of death.

Practically speaking, providing guideline-recommended preventive care would require ~27 hours per doctor per day. And the incentives are misaligned: health systems profit when hospital beds are full, so they lack the business model to actually invest in prevention.

So it's a clear illustration of a systematic gap between research and care delivery.

There are multiple independent risk factors for heart disease. The major ones are:

  - LDL / ApoB
  - Blood pressure
  - inflammation (hs-CRP)
  - Insulin resistance (HbA1c)
  - Lp(a): strongest hereditary risk factor.
  - eGFR: a measure of kidney function

Non-statin drugs like PCSK9 inhibitors have been shown to reduce heart attacks, strokes, and other cardiovascular events on top of statin therapy. One randomized control trial was FOURIER in 2017: https://www.nejm.org/doi/full/10.1056/NEJMoa1615664

I think the title is deliberately provocative, but they're not wrong.

Heart disease is largely solvable from a biomedical standpoint: we have accurate biomarkers (e.g., ApoB, Lp(a), hs-CRP), long-term risk prediction models, precision nutrition, and highly effective next-generation drugs (PCSK9 inhibitors, lepodisiran, etc).

But practically speaking, heart disease remains the #1 cause of death due to bottlenecks in care delivery: e.g., 46% U.S. counties have no cardiologists, providing guideline-recommended preventive care would require ~27 hours per doctor per day, and incentives are misaligned (health systems profit when hospital beds are full, not from prevention).

The main side effects of statins are muscle pain and brain fog (from some statins -- others cross the blood-brain barrier much less).

The benefit of statins is to not only lower LDL cholesterol, but also inflammation, which is now actually a stronger risk factor for cardiovascular disease than cholesterol: https://www.empirical.health/blog/inflammation-and-heart-hea...

Layering PCSK9 inhibitors, ezetimibe, and statins can lower ApoB/LDL cholesterol by 85–90%, which would have been unheard of until recently.

On the horizon, drugs in clinical trials lower Lp(a) (the strongest hereditary risk factor for heart disease) by 94%. Currently, there are four RNA-based drugs in trials that effectively silence the gene that makes Lp(a) in liver cells: lepodisiran, olpasiran, pelacarsen, and zerlasiran.

This is cool! It'd be interesting to correlate this with Vitamin D too.

This is a little different since the Apps SDK lets developers create specialized tool calls to their servers, and create specialized in-chat UI components. It's an evolution of the same concept as the GPT store, but a very different take on the idea.

HbA1c, or just diabetes as a binary variable, has been one of the main inputs into predicting heart attack risk for a long time.

The main marker of kidney function, eGFR, was added with the AHA/ACC's PREVENT equations in 2023.

I wrote a bit about the science behind heart risk calculators, and their various inputs like cholesterol, blood pressure, A1c, eGFR, and so on here: https://www.empirical.health/blog/heart-attack-risk-calculat...