I have been an advocate of better transparency in AI reporting for research () and the broader public ().
Really appreciated the opportunity to contribute to the new TRIPOD+AI guidelines to address the “how”:
2023: HBO Launches Max
2024: HBO Launches Min
2025: HBO Launches Mean
2026: HBO Launches Median
2027: HBO Launches Standard Deviation
2028: HBO Launches Interquartile Range
2029: HBO Launches Regression to the Mean
p < 0.001 extremely significant
p < 0.01 highly significant
p < 0.05 significant
p < 0.10 trending to significance
p < 0.20 horizon of significance
p < 0.50 coin flip of significance
p < 1 so you’re saying there’s a chance
Statistician: The p-value is 0.16780.
Clinician: All I want to do is this. I just want you to remove 0.11780, because I know the p-value is less than 0.05.
Statistician: Do you ever use statistics?
ML researcher: Nope. Never.
Statistician: What about when reading a paper?
ML: Nope. Never.
Statistician: Ok. So if you’re reading an ML paper comparing lots of models, how do you know which one is the best?
ML: Bold font.
Computer scientist: The p-value was 0.10, so I tweaked the model until it achieved a p-value of 0.03.
Statistician: But… that’s p-hacking!
Computer scientist: No, no, no! In our field, it’s called “alignment.” I didn’t like the output, so I applied a bit of RLHF until I did.
My lab is moving to
#JuliaLang
, and I’ll be putting together some R => Julia tips for our lab and others who are interested.
Here are a few starter facts. Feel free to tag along!
Julia draws inspiration from a number of languages, but the influence of R on Julia is clear.
If statisticians wrote papers like computer scientists:
“By leveraging logistic regression, we exploited linear algebra to capture a latent log-odds representation of the data, uncovering coefficients with SOTA p-values.”
The presenter puts up a slide showing “random forest variable importance.” You know the one...
The sideways bar plot.
Says “only showing the top 20 variables here...” to highlight the hi-dimensional power of random forests.
The slide is awkwardly wide-screen. Everyone squints.
Some professional news - being on the faculty at
@UMich
, I have learned from great students, colleagues, and mentors.
Next year, I’ll be joining
@UCSanDiego
as the Joan and Irwin Jacobs Chancellor’s Endowed Chair of Digital Health Innovation and Chief AI Officer for
@UCSDHealth
.
Computer scientist: We applied a non-linear sigmoid transformation to the weights in our machine learning model.
Statistician: 👀
Computer scientist: 👀
Statistician: So you did a logistic regression?
Computer scientist: Yes.
The DeepMind team (now “Google Health”) developed a model to “continuously predict” AKI within a 48-hr window with an AUC of 92% in a VA population, published in
@nature
.
Did DeepMind do the impossible? What can we learn from this? A step-by-step guide.
Thank you for folks who have shared or commented on our paper. I know the paper is being used by some to dunk on Epic. Rather than piling on, I want to provide a clear-eyed view of what we found, what it means, and what I would suggest to Epic (& other model devs) going forward.
Study suggests that the Epic Sepsis Model poorly predicts
#sepsis
; its widespread adoption despite poor performance raises fundamental concerns about sepsis management on a national level
A Visual Tour of the Meta-Tidyverse
For years, I’ve been trying out different non-tidyverse implementations of tidyverse. It’s fun seeing folks mold languages to run analysis code inspired by it.
If you like screenshots of code, come along for a visual tour.
Let’s start w/ R.
Clinician: I fit a multivariate regre—
Statistician: NO. It’s not multivariate. It’s multivariable. Anything relating to the explanatory variables ends with “variable” NOT “variate”!
Clinician: By explanatory variables, do you mean the covariables?
Stats: NO, the covari—
Stats 101 syllabus: You know NOTHING. In this class, we will peel a layer from this onion of dark arts known as statistics. Prepare to enter and learn secret knowledge.
Stats 601 syllabus: Everything you learned in Stats 101 was wrong.
Stats 901 syllabus: Everything is wrong.
I recently submitted my R01, which was written and revised entirely* in Google Docs. If you have used Word previously for grants but are considering making the switch, this step-by-step guide is for you.
*2 collaborators used Word and final version was edited in Word
ML researcher: To really master neural networks, you need to understand a variety of tensor operations with sophisticated terminology.
Statistician: Can you give me an example?
ML researcher: Like unsqueeze.
ML researcher: So why do you want to predict mortality?
Clinician: Because we try so hard to save lives, but some ppl still die. Is there a problem?
ML: Yes, the outcome is imbalanced. Not enough ppl are dying.
Clinician: So what are we supposed to do?
ML: Isn’t it obvious?
datar is a relatively new Python package with dplyr/tidyr syntax supporting numpy, polars, and arrow backends.
This looks like a really promising implementation of the tidyverse, including support for factors, rowwise, nesting, and more!
I’m excited to join
@UCSanDiego
@UCSDHealth
@InnovationUCSDH
!
I pledge to work with colleagues and patients to improve the experience of receiving and delivering care, and to help advance the science of using AI towards better, faster, and more accessible care.
Very excited to share that Dr. Karandeep Singh will serve as the inaugural 𝐂𝐡𝐢𝐞𝐟 𝐇𝐞𝐚𝐥𝐭𝐡 𝐀𝐈 𝐎𝐟𝐟𝐢𝐜𝐞𝐫 at UC San Diego Health and 𝐉𝐚𝐜𝐨𝐛𝐬 𝐄𝐧𝐝𝐨𝐰𝐞𝐝 𝐂𝐡𝐚𝐢𝐫 𝐨𝐟 𝐃𝐢𝐠𝐢𝐭𝐚𝐥 𝐇𝐞𝐚𝐥𝐭𝐡 at UC San Diego School of Medicine!
👉
When ppl ask me about my priorities as a Chief Health AI Officer, I share the Health AI Paradox.
Researched models aren't implemented. Implemented models aren't researched.
My priorities are simple. Help researched models get implemented, and implemented models get researched.
Regression to regression: the phenomenon when a data scientist, after having tried every possible configuration in XGBoost, returns to using a regression model.
Statistician: How come you didn’t report the 95% CI in your paper?
ML researcher: So sorry I left it out. Didn’t realize you cared about that. The 95% CI is 3.8.
Statistician: … ?
ML: Well I’m using 4 GitHub Actions for CI. And 95% of that is…
Statistician: Get out.
ML researcher: One of the biggest limitations of linear models is that you can *only* use them to learn linear relationships.
Statistician: I have the same problem with linear algebra.
ML researcher: 👀 🤔
Statistician: 👁 👁
ML: I’m wrong, aren’t I…
Stats: Yes.
Undergrad: Will anyone notice if I change the page margins to 1.5”? How about 14 point font?
Faculty: Will anyone notice if I change the page margins to 0.45”? How about 10.5 point font?
In case you coincidentally happen to be looking for alternative ways to learn
#rstats
and tidyverse because your current learning platform is busy ... *checks notes * ... suing your interactive development environment, may I suggest this series of 73 YouTube lecture videos:
Introducing rjs: R in JavaScript
1. Add <script src=''></script> to your HTML file
2. Add class = "r-code" to any HTML element.
That's it.
GitHub:
(
#rstats
rendered on the cloud, powered by
@opencpu
)
@OscarBaruffa
As data scientists get older, they slowly, molecule by molecule, begin to merge with all the datasets they’ve ever worked with until one day, they leave behind the mere shadow of an inner join.
👋 🧹📊 TidierPlots.jl for
#JuliaLang
A 100% Julia implementation of
#rstats
ggplot2. Powered by AlgebraOfGraphics.jl, Makie.jl, and Julia’s meta-programming capabilities, TidierPlots.jl is an R user’s love letter to data visualization in Julia.
Introducing Tidier.jl for
#JuliaLang
:
A 100% Julia implementation of the
#rstats
{tidyverse}. Powered by the DataFrames.jl package and Julia’s meta-programming capabilities.
Still a work in progress.
Here's a quick tour of the highlights.
One myth about supervised machine learning (or predictive modeling) is that “different algorithms work best for different (clinical) datasets” and thus you must test lots of different algos and pick the best one. Let me present the counterexample: let’s separate blue from red.
Students: We are happy to move to Julia, but can you put together some resources for us to learn the tidyverse equivalent in Julia?
Me: Give me a month to prepare some “resources.”
***frantically re-creating tidyverse in Julia***
My lab is moving to
#JuliaLang
, and I’ll be putting together some R => Julia tips for our lab and others who are interested.
Here are a few starter facts. Feel free to tag along!
Julia draws inspiration from a number of languages, but the influence of R on Julia is clear.
Why does a proprietary sepsis model “work” at some hospitals but not others?
Is it generalizability? Measurement? Intervention? Patient population? Margin for improvement? Resource constraints?
Working with a team led by
@_plyons
, we looked at a 9-hospital network.
A story.
Factors Associated With Variability in the Performance of a Proprietary Sepsis Prediction Model Across 9 Networked Hospitals in the US | Critical Care Medicine | JAMA Internal Medicine | JAMA Network
@WashUi2db
@OHSUPulmCCM
@umichDLHS
Tidier.jl 1.0.0 is now on the
#JuliaLang
registry.
It’s 𝘧𝘪𝘯𝘢𝘭𝘭𝘺 a meta-package. It re-exports TidierData.jl, TidierPlots.jl, TidierCats.jl, TidierDates.jl, and TidierStrings.jl.
In Punjabi, the phrase “corona” (karo-na) means “Don’t do it.” The name of the virus is an apt description of the prevention strategy.
Wanna go to Broadway? Don’t do it.
Wanna go to Disneyworld? Don’t do it.
Wanna go to school? Don’t do it.
The name pretty much sums it up.
A prediction model paradox in health: while many are developed, few are recommended.
But when models *are* recommended, they often come from tertiary care hospitals.
Is this a problem? We tested 3 prostate ca models in regional/ national data.
Paper:
🧵 In 2019, DeepMind published a paper in Nature describing models to predict AKI in US Veterans, finding AUCs of 92% for RNN & 89% for GBDT, better than prior work.
Inspired, we built & tested the generalizability of similar models in & out of VA.
Link:
ML researcher: Then I trained this model using an ML dataset and—
Statistician: A what?
ML researcher: An ML dataset. You know, a dataset for doing ML. It’s how we advance the field.
Statistician: …
ML researcher: Don’t you people have statistics datasets?
Statistician: …
I’ll be giving a talk on implementing predictive models at
@HDAA_Official
on Oct 23 in Ann Arbor. Here’s the Twitter version.
Model developers have been taught to carefully think thru development/validation/calibration. This talk is not about that. It’s about what comes after...
.
@elonmusk
Hospital recruitment
“An MD is definitely not required. All that matters is a deep understanding of biology, biochemistry, pharmacology, physiology, pathophysiology, genetics, social dynamics, life and death, and an ability to implement this in clinical practice.”
Tesla AI recruitment 🔥
“A PhD is definitely not required. All that matters is a deep understanding of AI and ability to implement NNs in a way that is actually useful (latter point is what’s truly hard). Don’t care if you finished high school.”
@elonmusk
Why do seemingly useful models fail to improve clinical outcomes when implemented? Resource constraints.
In this paper, we describe constraints, how they affect net benefit, and how they apply to other measures.
Paper:
R pkg:
Leading scientists have raised concerns about the autocratic tendencies during the “reign” of the p-value, including misinformation being spread by the p-value on social media.
Statistician: Elon Musk thinks we’re living in a computer simulation with parallel universes. That guy is just so weird.
ML researcher: Agree! BTW, what is a 95% confidence interval again?
Stats: Assume we’re living in a computer simulation with multiple parallel universes...
8 months later and still haven’t turned back.
We still have existing projects in R and Python, but the syntax, tooling, and developer experience in Julia are so smooth that it’s hard to go back.
Language interoperability is great, and it’s nice to have a fast glue language.
My lab is moving to
#JuliaLang
, and I’ll be putting together some R => Julia tips for our lab and others who are interested.
Here are a few starter facts. Feel free to tag along!
Julia draws inspiration from a number of languages, but the influence of R on Julia is clear.