Rohit Vashisht

@kdpsingh.bsky.social focusing on the "build before you buy" mentality. A lot of the AI tools that are most useful to clinicians aren't actually commercially available, so internal teams build tools they need, like their NoteReader which can help automate measurements.
#AMIA2025 #MLSky #MedSky

RNG state management need to be a whole lecture in it self, may be two.

Why include unadjusted estimates in a study when reporting adjusted estimates? Upon reading "The Strengthening the Reporting of Observational Studies in Epidemiology " or STROBE guidelines, one will find advice to (a) Give unadjusted estimates and, if applicable,

“It is simpler to do erroneous things.” - @f2harrell.bsky.social

H/t @phdemetri.bsky.social

stats.stackexchange.com/questions/67...

I remember (pre-COVID) great debates usually in conferences about the value of providing un-adjusted estimates to illustrate likely Simpson paradox (among other bias) and hence the need for adjustments in observational studies. Those discussion never really went anywhere in practice.

👀

One of the fun things about launching @timead.bsky.social is the amazing early career folks. We started a "scientist spotlight" website section. The 3 posted are just our first few - really gratifying to see the work of Master's & PhD students, and post-docs -more spotlights coming!

bit.ly/3VCa84g

Love writing math? Tex is the only way :-)

Try @typst.app if you haven’t allready.

Picture of feet on the edge of a diving board about to jump off backwards, with blue-painted toenails and a tattoo just above the ankle.

I wrote a LinkedIn post on advice for incoming Master's students in public health and epidemiology. tl;dr: do hard things, get help when needed, make friends and thank people, and have fun - it's amazing. #EpiSky #DataSciSky #PublicHealth #GradSchool
Full article: bit.ly/4neVYlE

Tired: Global interpreter lock for data heavy parallel processing in Python

Wired: Reading data in 1GB chunks and processing sequentially

Data fog is a new one!

@caltrain.com station, SF.

We love to see JCHI's @kdpsingh.bsky.social, @ucsdhealth.bsky.social named one of 24 new ACMI Fellows!

His leadership in health AI, model governance & informatics policy is reshaping how AI is used responsibly in care.

Read more⤵️
amia.org/news-publica...

The proper link that would work in here: ai-podcast.nejm.org/e/from-clini...

From Clinician to Chief Health AI Officer: A Conversation with @kdpsingh.bsky.social www.podbean.com/ew/pb-9t5e4-...

For clinicians, Singh’s story is a reminder that the best AI is guided by patient care, deep expertise, and humility about the limits of technology.

Inspiring!

Excellent perspective! Can’t get better than this, and very helpful. Thought our work might be of interest (currently under review) therefore sharing papers.ssrn.com/sol3/papers....

Late night LLM training thoughts:

I’m increasingly appreciating APEs over other position embedding. Their simplicity makes them easy to implement and efficient in distributed training setups. APEs could also provide a stable foundation for enhancing the consistency of LLMs.

qwen is good, like really good!

Statistician: What is zero-shot learning?

ML researcher: It’s when you don’t train your model, you just use it.

Stats: So you just take a shot?

ML: Exactly.

Stats: One shot…

ML: Exactly.

Stats: …

ML: Makes sense tho. Think of it like Python list indexing, where the first shot is the zeroeth.

Tired: Zero shot learning

Wired: No shot learning

Manuscript review requests are like poking one bee and getting the whole dang hive buzzing after you!

Say yes to one, and suddenly you’re swarmed with a zillion more “review this, please!” emails.

Mother: I have twins.

Doctor: Oh, so they were born on the same day?

Mother: No, they were born with the same propensity score.

This is deep, very very deep!

When I started post doc it was ‘patient like me’ then it was ‘patient like mine’ now ‘digital twin’ next ‘Artificial patient’ next to next ‘patient like my AP patient’ … circle of life

Tired: Adjusting for covariates

Wired: “Building a digital twin”

If you understand the nullspace, you understand the matrix!

— Gilbert Strang

“Our model is probably in our view the best model for this task and the fastest and second cheapest but not the best performing but the benchmarks don’t reflect the nuances of real-world use” is a whole lot of words to say that it’s not SOTA.

Looks like they took inspiration from Bilbo Baggins:

"I don't know half of you half as well as I should like, and I like less than half of you half as well as you deserve”

Tired: Long context length

Wired: No context length

The more I learn about @duckdb.org the more amazed I get, and more I realize how much I don’t really know, and a long way to go!

Streamline data app development in Positron with the App Preview feature. 🪟

No more context-switching! Now you can run & debug your #RStats and #Python data apps (Shiny, Streamlit, Dash, etc.) right inside Positron's integrated Viewer pane.

Learn more: positron.posit.co/develop-data...

Thanks Costco!