Jacob Rosenthal

Dehumanization of Patient Care—Are Computers the Problem or the Solution? Edward H. Shortliffe, MD, PhD; Dehumanization of Patient Care—Are Computers the Problem or the Solution?, Journal of the American Medical Informatics Assoc

This piece from more than 30 years ago is incredible - replace "computers" with "AI" and it could have been written yesterday. I predict that it will still be relevant 30 years in the future for whatever the next big technology leap will be

academic.oup.com/jamia/articl...

Perspectives on Artificial Intelligence–Generated Responses to Patient Messages This cross-sectional study of patient queries in US electronic health records examines laypersons’ satisfaction with answers generated with artificial intelligence (AI) compared with clinician respons...

When patients have questions about their health, they prefer answers from AI over answers from human physicians. Helps explain why people are increasingly turning to chatbots for medical advice!
jamanetwork.com/journals/jam...

There is a lot of literature discussing the problems of "black box" algorithms in medical AI applications. But the world isn't black and white, and this framing ignores important differences between algorithms and user interfaces. Better to view AI interpretability/explainability as shades of grey.

There should be more overlap between coders and surgeons. The OR is the only environment I’ve been in during med school so far where a deep focus flow state is common. Hours fly by working on detail-oriented, tedious tasks. Pursuit of this focus also drives many 10x programmers.

Regulation can boost innovation by building trust! In absence of regulations, more likely to have bad events that end up delaying whole fields of research. Eg after Jesse Gelsinger’s death, all gene therapy trials were paused in the US. Tragic for his family and net negative for scientific progress!

Snails are macro-scale example of the chirality arms race. Mirror mutants have fitness advantage because the snakes that eat them evolved to have more teeth one side of their jaw.

Scary but not surprising that the chirality arms race goes down to the molecular level!

www.nature.com/articles/sre...

Have been reading about Test-Time Training after seeing it in the ARC prize technical report. Very interesting ideas and tied with some things I have been thinking about recently re: AI in medicine. AI models deployed in Clinical trials should continue learning at test-time (i.e. in the real world)!

What benchmarks should we be using to evaluate medical AI models? Answer: all benchmarks have limitations. Need to instead use real-world metrics that reflect what actually matters to patients. Huge opportunity for AI models fully integrated into EHR.

Ayers et al, JAMA 2024

Tech Debt in ML Systems (Sculley et al, NeurIPS 2015)

Takeaways:
1. SYSTEMS are what matter most, not individual models
2. Training new models is not the only way to conduct research and/or have impact
3. Models are useless w/o infrastructure for data ingestion, serving predictions, and monitoring

Papers that have changed how I think about AI in Medicine: 🧵

Artificial Intelligence and Radiologist Burnout This cross-sectional study investigates whether there is an association between artificial intelligence (AI) use in radiology and radiologist burnout in China.

New study: AI _increases_ burnout in Radiologists

Highlighted difference between screening and clinical settings. AI can actually increase workload in clinical settings when an abnormality is found due to time spent in differential diagnosis.

jamanetwork.com/journals/jam...

Great list. That’s the first AI paper that I have seen quote the Talmud!

One commit at a time.
One flashcard at a time.
One mile at a time.

Consistent incremental progress over time is so powerful. Crush that project/exam/marathon!

There aren’t many everyday scenarios where we have access to such precise probabilities to base decisions on. Usually it is more coarse: my mental model is closer to a categorical (eg certain, likely, unlikely, etc.) than a continuous probability. In which case 70% and 80% both fall under “likely”

Who is the person with the API keys to access the data? That is who you should collaborate with!

Medical knowledge isn’t as special as we might think! In my experimenting, I have never run into problems with GPT-4 lacking domain knowledge

A difficult but important skill for research is being able to appraise a new software/package/model and decide whether it is worth the investment of time and effort to learn how to use it

There are tasks that humans do today, and tasks that we would never expect a human to be able to do. More interesting are the tasks in the middle: what can humans not do today, but could do tomorrow with the right tools/knowledge?

Don’t link to a GitHub repo if it isn’t public. If you want to publish code alongside a paper, then both should be released at the same time. This concludes my TED Talk.

*whoosh* The sound of your pun going over my head

Although to be fair, I don’t think TI-basic even has classes!

Why not both? I first learned how to code on my graphing calculator while bored in class in 8th grade

Are there any datasets that are truly unbiased?

A more pragmatic challenge is how we can best assess and mitigate biases in data/models, knowing that there will always be *some* biases (in the statistical sense)

He applied to medical school twice but didn’t get in. I wonder what he would think about the arc of technology in medicine

Isaac Asimov was 20 years old when he started writing the Foundation series

Good points!

One reason for Python’s success is that its syntax is relatively close to English prose. But I imagine this would have ~zero benefit for non-English speakers.

Are there programming languages that are more popular & accessible in other parts of the world, built on languages other than English?

I love this city 🗽

People overestimate how much AI will change medicine in the next year, but underestimate its impact in the next 10 years

Copilot has completely changed how I write code. Even though it is not perfect, the gains from 10x speed-up of iteration cycle far outweigh time spent fixing mistakes. I am very excited for how LLMs will similarly transform clinical note writing!