Over-optimism for AI comes partly from the coincidence that the people developing it (coders) happen to perform the economic task it is the most adept at (coding). This creates a very strong self-reinforcing loop of hype and (imo) naive over-extrapolation of capabilities to other domains

I find 'alpha spending' to be an incredibly awkward concept. How many hypotheses can I test? At what point does it stop being kosher? How much do I need to correct for multiple testing? How many tests can I run and how many I should just never test again? Is pre-specification the only answer? #stats

One of the best pieces of advice I got as a doctor in training was to share my uncertainty and transparently lay out my thought process with patients. It goes against conventional wisdom that one may appear incompetent, unconfident, or indecisive. I've yet to meet a patient who didn't like it #meded

If you have a statistician, ML engineer, or some similar methods expert on your project, give them commensurately high author line credit. They usually contribute far more and consequentially to studies than most other authors, yet often end up low middle-author or sometimes left off entirely #stats

#ThePitt has been widely appraised for uncannily accurate dramatizations of emergency medicine. Pleased to report they've also done a great job with transfusion! www.trasci.com/article/S147... #medsky

Among adults undergoing cardiac surgery at high risk of bleeding and platelet transfusion, cryopreserved platelets did not meet noninferiority for controlling postoperative bleeding and were less effective than liquid-stored platelets. #CCRdownunder ja.ma/4rEC06z

The way the exponential form generalizes so many common, seemingly unrelated distributions (and all the properties you get for free, like sufficiency) has to be one of the most elegant things I’ve learned in #stats, and in general Good reader: brunomaga.github.io/Exponential-...

In the era of evidence-based medicine, why do the similar patient face different treatment strategies at hospitals in the same healthcare system? We analyzed 800k+ admissions at 24 hospitals to examine transfusion practice at teaching and non-teaching hospitals link.springer.com/article/10.1...

For those wrestling with applying MINT trial and new AABB RBC MI guidelines results with nuance, I think this figure strikes a good balance for navigating the evidence base at the bedside doi.org/10.1016/j.cj... #transfusion #medsky #cardiology

A lot of people in healthcare are asking, "I want to use AI for [my field]" Higher yield: Let’s make a list of problems worth solving, then discuss if any are best addressed with AI (most won't be) (e.g. studies aren't typically motivated by, "I want to use retrospective cohort design")

I think most people in the ML/AI space understand that, for most real-world uses, LLMs are merely the gateway drug to entice authority figures to invest in resources that will enable more reliable analytical ML systems. Kind of like how the WWF uses pandas to fund saving other endangered species

How are you diagnosing and treating mixed AIHA? pmc.ncbi.nlm.nih.gov/articles/PMC...

Profiting from selling unproven therapies to desperate patients is as old as time. Why does it persist? Co-authors discuss antecedents, consequences, and contemporary examples of medical grifting from #transfusion, #hematology, and biotherapies onlinelibrary.wiley.com/doi/10.1111/...

I was a programmer before I was a data analyst, and so have a toxic trait of jumping to nested loops for problems where dataframe transformation would be far more elegant. I can't possibly be alone? #stats #ML

Classic teaching is that no post-1 hr increment points to immune refractoriness. One important exception is device-related failures, e.g. plts going into a clotted line, a faulty dialysis/apheresis circuit, an ECMO machine, etc! #hemesky #transfusion

What is the current state of AI/ML research in transfusion medicine? Co-authors Na Li, @ruchika_goel1, et al, have compiled a critical summary of research applying AI/ML in transfusion, including some glossary tables lay readers will helpful Read here: www.sciencedirect.com/science/arti...

My progression as a reader of scientific literature can be summarized by the first sections I read when I pick up an article High school: Abstract → Conclusion Undergrad: Abstract → Introduction Medical school: Abstract → Discussion Postgraduate: Abstract → Results Now: Abstract → Methods #medsky

Can we improve the way we transfuse plasma? Yes. Big gaps. E.g the ineffectual practice of 1-2 plasma unit dosing is widespread, accounting for 4/5ths of all orders Study from our group now out in @aabbupdates.bsky.social journal onlinelibrary.wiley.com/doi/10.1111/...

Fancy health AI tool/models have to produce *some* real-world healthcare benefit on one of 5 core health outcomes: 1. Death 2. Disease = occurrence/severity/complications 3. Discomfort = phys/emotional 4. Disability = functional impact 5. Destitution (cost to individual/system) #ai #mlsky #medsky

I believe when working with big data for causal inference, you HAVE to pre-specify your analyses (esp within a project). Do it for yourself, not for reviewers. As Feynman said, “[in science] the first principle is that you must not fool yourself and you are the easiest person to fool” #stats

🧪 ICARUS-ED: Early albumin in sepsis? 📉 Minor gains in numbers, no real outcome benefit 💰 30x the cost vs crystalloids ⚠️ Outcomes > optics 🔍 Full breakdown: https://wp.me/pdrP8b-5oc ❤️ Like 🔁 Repost 🔔 Follow for more #FOAMed #MedX #MedTwitter #MedEd

“This trial adds to the growing body of evidence suggesting no clinically meaningful benefit to albumin administration in the resuscitation of septic patients”

We need trials of AI showing better patient important outcomes (I'm not aware of any) #AI #ML #medsky

Seeing a lot of PCC use in liver disease lately. What why how #hemesky #medsky

Why do most hospitals have vending machines full of junk food but not a gym? #medsky

You can tell if the person in your machine learning course has a math, stats, or compsci background by the quirky way they write code #ML #stats

If you still use 5% IV albumin outside of PLEX, why? #medsky

Canadian Blood Services data show rising demand for blood, reflected in increasing per-admission blood use in Canadian hospital databases. This is concerning as baby boomers are aging out of the donor pool, and zoomers are not replacing their numbers. onlinelibrary.wiley.com/doi/10.1111/...

June 19th is World Sickle Cell Awareness Day. Below is a link to a good review article on the subject, and in this thread are my two wishes as a physician who treats the condition: www.healthaffairs.org/content/fore...

Blood is shared across legal borders to save lives. Equitable and efficient movement of blood is guided by science and justice and must be insulated from ideologically motivated practices, argue co-authors @jwjacobs12.bsky.social @aabbupdates.bsky.social onlinelibrary.wiley.com/doi/10.1111/...

I was today years old when I finally saw this masterpiece, which is an embarrassing confession as a movie buff. Absolutely blown away by the writing and pacing #filmsky

The ALBICS trial is a humbling reminder that no matter how elegant physiologic reasoning may be, we need real-world experiments to know whether transfusion will benefit patients jamanetwork.com/journals/jam... #nephrology #transfusion

In this Nature Medicine commentary @drgsbooth1.bsky.social and I discuss how the current political environment in the USA, including the weaponization of immigration enforcement, is impacting global scientific research, and the implications for science. www.nature.com/articles/s41...

I know there are many more sophisticated uses for causal inference, but DAGs are just straight up awesome for brainstorming study questions and thinking through models #stats #causalinference #statsky

Mixed Autoimmune Hemolytic Anemia is a fascinating and challenging disease of blood breakdown via multiple immune pathways Our systematic review summarizes what is currently known about features and outcomes of patients with this condition onlinelibrary.wiley.com/doi/10.1002/... #hemesky #medsky

People who are MD + PhDs What would make you advise someone to do PhD or MD first? #medsky

🩸Would you want “unvaccinated” blood? Some patients now demand it—and lawmakers are listening. Our new Annals piece warns why this shift could endanger blood safety, amplify misinformation, and erode public trust. Read: www.acpjournals.org/doi/10.7326/... #BloodSafety #Bioethics #Misinformation

Is directed blood donation (donation earmarked for a specific recipient) ethically or medically justified? In @annalsofim.bsky.social, my co-authors argue that in the vast majority of clinical situations, the answer is no www.acpjournals.org/doi/10.7326/...

Proud to have contributed to OpenAI's #Healthbench, alongside physicians from around the world. This was a unique opportunity to evaluate how AI performs on real health challenges and help shape how we measure progress. Learn more: openai.com/index/health...

Does anyone know a good educational resource on federated learning? Can be lectures, books, etc (no research papers... I am not there yet) #stats

Can we rationalize the way we give diuretic prophylaxis before transfusion? Colleagues @uhntoronto.bsky.social incl talented co-fellow Dr. Rotin and @jacobpendergrast.bsky.social have made a substantial stride in dose modelling to target urine output onlinelibrary.wiley.com/doi/pdf/10.1... #medsky

#NATA2025 The side for liberal hemoglobin transfusion thresholds make their case for MI and neurological injury #medsky

Viscoelastic Testing in Post-Partum Hemorrhage Potentially useful, understudied HYPOcoagulability is important as well described HYPERcoagulability is also important, poorly studied #NATA25 #medsky #transfusion #obgyn

Intravenous Iron at Home Dr. Elvira Bibs, reports on the Spanish experience Better for older patients, saves hospital beds, gives a high need treatment. Takeaway: select the right patients. Mind the logistics "What is lacking? Courage to implement it" #NATA25 #hemesky #medsky #transfusion

Great talk to start off #NATA2025 by Dr. Paul Kalra, the investigator on IRONMAN trial for iron replacement in CHF → Patients feel better, and do better, with greater benefit in TSat >20% → No signal for increased infection related death No brainer! #medsky #hemesky #cardiology

One of my favorite medical conferences. Excited to connect with friends and experts on all things patient blood management!

Neat abstract (and accompanying arXiv pre-print) on LLM agents for medical tasks at #neurips neurips.cc/virtual/2024... #ml #medsky

You don’t appreciate how much GIM people know until you become a sub-specialist and realize how much you’ve forgotten #medsky

The more I understand cross-validation, the less I understand why it isn't standard practice in all retrospective analyses Maybe #statsky can explain..