Graham Walker, MD on LinkedIn: In Anthropic's latest research, they attempt to unravel the mysteries… | 21 comments (2024)

"Ages 45-65 is like pre-school for Medicare" —Eric Bricker, MD on the middle ages being the biggest spend "Primary care has become an administrative specialty without administrative tools" — Paulius Mui, MDGot to spend all day with an AMAZINGLY smart and driven group of people discussing how to improve primary care, what its role should be, and who's to blame 🤣 —cool to see how many people are all trying to improve things from their own perspective. Healthcare is hard. Thanks to Kenneth Qiu, MD Paulius Mui, MD David M. Nichols, MD Gus Roman, MBA/MPH for organizing and supporting the meeting and Health Wildcatters for hosting us in Dallas!

60

9 Comments

I've been increasingly using Consensus and its ChatGPT bot for my quick lit reviews as a substitute for Google Scholar or plain' old Google (POG)... but should I be? We're entering a weird era where:↪️ The Consensus summary often *is* correct —and it gives a decent summary (this is what you'd hopefully suspect since it's hopefully grounded in source material)↪️ But it's often correct *despite* not really using what most of us humans would consider the 𝗯𝗲𝘀𝘁 𝘀𝗼𝘂𝗿𝗰𝗲𝘀. For example in the screenshot below, the most famous trials of carvedilol in heart failure are sources 7 and 8.— Now it's probably still getting the answer right because it's using meta-analysis sources (which include the best sources). But still —if you ask Google (𝘗𝘖𝘎, I'm trying to make it a thing) —it'll immediately give you the trial.↪️ Consensus will sometimes show trials that aren't even relevant to the question at hand, too.It's kind of like your math teacher making you show your work — you might get a point for getting the right answer, but if it's based on the wrong logic —you can't really trust that the student knows what they're doing.

49

24 Comments

Am I the only one that looks at these ads and thinks, "Oh there's ibuprofen in the Advil cream" or "There's acetaminophen in the Tylenol cream?"I can already imagine having the conversation with patients: "Yeah, I know you think you've been using an anti-inflammatory and it's not working. Unfortunately, you haven't." These are essentially BenGay/IcyHot/Salonpas or lidocaine. Product-line extensions: Now we can confuse patients in the over-the-aisle of CVS, and for twice the price!

27

12 Comments

While we were out trying to save lives, flatten the curve, and fight COVID vaccine misinformation, the Pentagon was actively spreading it in The Philippines, one of our oldest Asian allies.Vile. Abhorrent. Reprehensible.

19

13 Comments

Y’know how Wikipedia could be showing you inaccurate information — because the Edit button allows anybody to freely go in and change info on the 𝚃𝚠𝚘-𝚝𝚘𝚎𝚍_𝚜𝚕𝚘𝚝𝚑 page? Well, Wikipedia is kinda how I think about LLMs: they could show false information at any time. Wikipedia has significantly improved its accuracy over the years, and I thought, 𝘐 𝘸𝘰𝘯𝘥𝘦𝘳 𝘪𝘧 𝘵𝘩𝘦𝘪𝘳 𝘵𝘦𝘤𝘩𝘯𝘪𝘲𝘶𝘦𝘴 𝘤𝘰𝘶𝘭𝘥 𝘣𝘦 𝘭𝘦𝘷𝘦𝘳𝘢𝘨𝘦𝘥 𝘵𝘰 𝘦𝘯𝘩𝘢𝘯𝘤𝘦 𝘵𝘩𝘦 𝘳𝘦𝘭𝘪𝘢𝘣𝘪𝘵𝘺 𝘰𝘧 𝘓𝘓𝘔𝘴?🌐 Controlled Data Access and Updates: Just like Wikipedia protects highly visible pages from being freely edited, could LLMs benefit from more curated and vetted data sources? Or give those data sources more weight?🌐 Source Verification and Transparency: Wikipedia relies on citations and references to validate its content. Integrating fact-checking mechanisms and enabling LLMs to generate citations for their outputs can enhance the trustworthiness of generated content.🌐 Traceability and Accountability: Wikipedia's revision history logs all changes, making it easy to trace and revert inaccuracies. Could logs of expected outputs be compared over time to see where the model went astray so you could trace it back?🌐 Collaborative Input and Feedback: Wikipedia uses talk pages for editors to discuss and resolve content disputes. I’d love a platform where validated experts can provide feedback on LLM outputs to improve model accuracy.🌐 Access Control and Expert Oversight: Wikipedia assigns roles to users, restricting editing capabilities based on experience and trust. LLMs could have trusted users/experts’ opinions on output quality matter more when giving feedback.I am nowhere near facile enough in the training and development of LLMs to know if any of these are even practical — but thought it was an interesting comparison as I tend to trust Wikipedia content way more than I did a decade ago.#Wikipedia #LLMs #DataAccuracy #TechInnovation

42

41 Comments

For 🔟 years I wrote a monthly column called 𝙀𝙢𝙚𝙧𝙜𝙚𝙣𝙩𝙤𝙡𝙤𝙜𝙮 in an emergency medicine magazine. I gave it up during COVID as my articles started to become more negative — a sign I was definitely burning out. I've always found writing to be a really great way to organize (and quiet) all the thoughts in my head. Hit 🔟,0️⃣0️⃣0️⃣ followers sometime over the weekend; thank you for reading and sharing my words on how to try to get this whole healthcare thing right!

166

20 Comments

𝟳 𝗙𝗲𝗮𝘁𝘂𝗿𝗲𝘀𝗘𝗽𝗶𝗰 𝗰𝗼𝘂𝗹𝗱 𝗹𝗲𝗮𝗿𝗻 𝗳𝗿𝗼𝗺 𝗔𝗽𝗽𝗹𝗲 𝗜𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝗰𝗲In the era of AI hype, I've learned to demand proof before trust, but Apple's AI product is grounded in the right *concepts* — and I thought, "Wait, why can't we have that in healthcare?" Here's what resonated:1️⃣ Personal context: I'm an ER doctor; my day to day work could not be MORE DIFFERENT from a pathologist. Yet we're both treated similarly in the EHR. Apple understands that personal context matters deeply — and AI should acknowledge that. Certain things about a patient are very high priority to me and are much lower priority to someone else.2️⃣ Privacy: Duh; processing the least data possible on the cloud is better for privacy reasons. (I'll again shame Apple here; they frequently talk about how much they want to impact healthcare: why not make their devices, iMessage, and AI HIPAA-compliant?) 3️⃣ Context-aware Summary: I send a lot of long consult messages. Apple will summarize long text messages AND highlight a text message that's high priority; why couldn't Epic or other vendors do the same for my consults?4️⃣ Taking action: You'll be able to control your Mac or iPhone with Siri AI; why couldn't we do this within the EHR? "Show me a timeline of the antibiotics this patient has been on" or "Order a CT for appendicitis and insert the clinical history for the radiologist"5️⃣ Support/Help: Siri will be able to guide you through steps to "Change my Mac screensaver," why can't we have this for the EHR? "How do I place a referral to outpatient Orthopedics?" or "How do I filter for pathology results?" 6️⃣ Summary: While I'm not sure Apple's summarization tools would work perfectly in healthcare, we could build custom LLM models that would. "Summarize what happened in this ER visit and send a referral to ENT" could again be context-aware and include what's relevant to a head and neck surgeon, and also save me a lot of time.7️⃣ Call the Cloud for Compute: If the AI determines that your question can't be answered by your on-device LLM, Apple can send the question to its private cloud LLM, or can dial out to ChatGPT (or other models in the future) to answer questions. This reduces energy usage and data transfer needs.

95

39 Comments

Sunday fun: Dancing to Chappell Roan can give you a decent 🧠neuro exam: 1️⃣ upper extremity weakness 2️⃣ pronator drift 3️⃣ coordination 4️⃣ facial droop (+ 5️⃣/6️⃣ aphasia and dysarthria if you sing along)! #JunePride🌈

34

5 Comments

Welp, this wasn’t on my 2024 medical AI bingo card. ✅𝚀: What do you get when you cross an interventional cardiologist with a subtle heart attack EKG?𝙰: The human v. AI debate that I didn't think we'd see for a few more years.1️⃣ So a cardiologist posts an EKG on Twitter, and gets 𝘃𝗲𝗿𝘆 split opinions on how fellow cardiologists would handle the situation and diagnosis.2️⃣ This is a subtle EKG finding, no doubt —I've only seen it once in my career, when I was an intern in 2009. (I've read thousands of EKGs.)3️⃣ Robert Herman, MD himself a cardiology fellow and founder of Powerful Medical (an EKG AI company), runs the EKG through his algorithm called 👸🫀Queen of Hearts and it says "Hey, I think this is might be a heart attack!"4️⃣ A really fascinating debate takes place in the comments surrounding medical AI tools and their accuracy compared to humans:𝑊ℎ𝑎𝑡 𝑠ℎ𝑜𝑢𝑙𝑑 𝑏𝑒 𝑡ℎ𝑒 𝑠𝑡𝑎𝑛𝑑𝑎𝑟𝑑? 𝐼𝑠 𝑎𝑛 𝐴𝐼 𝑡𝑜𝑜𝑙 *𝑠𝑜 𝑔𝑜𝑜𝑑* 𝑡ℎ𝑎𝑡 𝑡ℎ𝑒 ℎ𝑢𝑚𝑎𝑛𝑠 𝑛𝑒𝑒𝑑 𝑡𝑜 𝑗𝑢𝑠𝑡 𝑏𝑙𝑖𝑛𝑑𝑙𝑦 𝑙𝑖𝑠𝑡𝑒𝑛 𝑡𝑜 𝑡ℎ𝑒 𝐴𝐼 𝑡𝑜𝑜𝑙 𝑖𝑓 𝑖𝑡 𝑠𝑎𝑦𝑠 "ℎ𝑒𝑎𝑟𝑡 𝑎𝑡𝑡𝑎𝑐𝑘?" 𝐼𝑓 𝑛𝑜𝑡, 𝑤𝑖𝑙𝑙 𝑖𝑡 𝑒𝑣𝑒𝑛𝑡𝑢𝑎𝑙𝑙𝑦 𝑏𝑒? 𝑊ℎ𝑒𝑟𝑒 𝑠ℎ𝑜𝑢𝑙𝑑 𝑐𝑙𝑖𝑛𝑖𝑐𝑎𝑙 𝑜𝑝𝑖𝑛𝑖𝑜𝑛 𝑓𝑖𝑡 𝑖𝑛 ℎ𝑒𝑟𝑒 — 𝑎𝑛𝑑 ℎ𝑜𝑤?Powerful Medical isn't out here pushing their tool on anyone, which I applaud. They are actively seeking research opportunities to test out their tool (and I'd encourage anyone interested to reach out —I have no financial stake, I just like what they're doing and have been following their work for many years).My take: we kinda did this with CT scans. We used to rarely get CT scans and trusted the surgeon's physical exam to guide appendicitis diagnosis. Now, 97% of the time if the CT scan shows appendicitis, the patient is probably getting an appendectomy. (Yes, yes, I recognize that an AI model is somewhat different than a CT.)

82

18 Comments