What's New in Surgical AI: 10/29
What's New in Surgical AI: 10/29
Vol 2: Scribes, (Even) Faster AI, and Best Paper
Welcome back to Ctrl-Alt-Operate, where surgeons dissect innovations at the intersection of artificial intelligence and healthcare.
This week, we focus on the rise of challengers to electronic medical records, highlight a new data science pipeline for brain tumors, and round up the best that Med+AI Twitter had to offer.
Table of Contents
🤖 EMR 1.0 is dead. Could an AI based on a Youtube joke save EMR 2.0?
🏆 Paper of the week: Cracking the genetic code of a brain tumor before surgery
🎺 Best of twitter (pre-Elon edition)
But first, a quick poll to check in:
EMR 1.0 is dead. Could an AI based on a TikTok joke save EMR 2.0?
Last week's letter highlighted the META 0.00%↑ speech matrix, this week we bring tech (large-language-models, speech-to-text, translation) back to the bedside.
Doctors, nurses and patients are begging for disruption of EMR 1.0
Every day, we spend a median of ≥ 1 hour (maximum of 5 hours!!) after the end of a shift “catching up on charting”, causing moral injury, burnout & divorce.
Patients are grieviously harmed from predictably awful UI/UX, and the nurses and doctors face criminal charges for the sins of failed EMR-human interactions.
a little bit of history
If you’re not in medicine, you might not have felt the visceral revulsion doctors and nurses have for the electronic medical record, or EMR. In the old days (ok, boomer Dan), notes and orders were scribbled on a piece of paper at the patient’s bedside. This was problematic but hyper-efficient for doctors and nurses. In 2009 the HITECH act mandated “meaningful use” of electronic records and prescriptions. This fifty-billion dollar industry is dominated by two systems whose minimal advances and shameless attempts to extract payments from the most vulnerable patients are legendary.
I would love to show you how bad this system is, but I could be sued by an EMR manufacturer due to “gag orders” in all major systems. No, I am not joking.
Fortunately, one medical Tik Tok celebrity wants to fix all of this…
Enter the AI “Jonathan”
Randomized controlled trials showed ≥10x increases in satisfaction when physicians pay a human “scribe” to convert their conversations to EMR-gobbledygook in real time, but scribes are expensive and rare (sorry Jonathan).1
AI can effectively transcribe and translate speech from hundreds of spoken languages (last week's letter highlighted META 0.00%↑ speech matrix dataset) and generate automated subtitles in near-real time.
But can an AI speak medicine?
🔗AI Medical Scribe Startups are coming here and here.
Could this actually be the missing piece?
Can we guide the direction of AI scribes to align the promise of the EMR with the humans of medicine?
Or will this be another misguided attempt at “efficiency” (read: extracting revenue from poor patients using uncompensated clinician effort)? Soon we’ll see AI- litigators sueing the AI-scribes (coming right up!).
At this point we are all deeply cynical about MedTech deployments being captured by the horribly broken medical system, but deeply optimistic about innovations in MedTech.
Our inboxes are open for implementations that keep the needs of humans, machines, and the medical system aligned.
🏆 Paper of the week - A pipeline to predict the genetic code of a brain tumor before surgery
Glioblastoma (GBM) is a deadly brain cancer that kills ~200,000 people worldwide annually. New medicines and surgical strategies based on the genetic code of some GBM tumors help some patients survive longer, but we have to drill a hole into the patient’s head to find out if their tumor has the “right” genes.
Matt Pease and the team @MSKCancerCenter @PittNeurosurg built a radiomics pipeline to predict these genetic alterations from MRI scans alone. This could mean faster times to diagnoses, getting the pathways for treatment moving prior to surgery, and at the minimum, allows for another data point of discussion between a clinician and patient + family who just received a life changing diagnosis. Kudos!
Best of Twitter (pre-Elon edition) 🐤
Faster is happening …
10am: Google releases the paper and model code for Flan-T5, an open source language model
1pm (3 hours later) : A hugging-face 🤗 API exists so you can utilize the model in (more or less) a handful of lines of code.
Last week we discussed Replit’s new iOS app allowing surgeons to code on their phones during faculty meetings. To make things even sillier, there’s soon the ability to develop on your apple watch.
An overwhelming number of guides and tweetorials are published everyday in the twitterverse. Here are the ones we actually read and liked:
Getting Started With Stable Diffusion: A Guide For Creators
10 AI websites to automate your life
Feeling inspired? Drop us a line and let us know what you liked.
Like all surgeons, we are always looking to get better. Send us your M&M style roastings or favorable Press-Gainey ratings by email at ctrl.alt.operate@gmail.com
Every single “innovation” in AI + medicine follows this format: “«x» is well-known to make «x» better, but we don’t have enough humans to do it.” Maybe we should just copy-paste, or come up with a catchy name for this like “the replacement dilemma.”