About
From 2002 to 2007, I trained in Pediatrics in NYC and Preventive Medicine at Johns…
Articles by Jay
Activity
10K followers
-
Jay Parkinson, MD, MPH reposted thisJay Parkinson, MD, MPH reposted thisWhat should the Hippocratic Oath say in 2026? We asked our How I Doctor podcast guests this question. The biggest takeaway? Doctors want to do good for their patients and they also want to protect themselves. Here's what they said: → Graham Walker, MD: Our patients' needs come first. But that doesn't mean our needs don't matter. → Paulius Mui, MD: Include a part that involves speaking out when you see something that's not right. → Ali Chaudhary M.D.: It's okay to say, "You know what, I need to prioritize myself." → Jay Parkinson, MD, MPH: Your job is to project manage your patients. → Jackie Gerhart: Every single time you are in front of a patient, someone is giving you their time and putting their life, frankly, in your hands. → Dr. Annie Andrews: We've lost our way because we are cogs in a wheel. We are working in a broken system. → Danielle Ofri: May I never see in my patient anything but a fellow person in pain. → Paul Tran, MD: Instead of "Am I doing good enough?" — reframe it. "Am I doing enough good?" → Shoshana Ungerleider, MD: There's always a role for healing, even if we can't cure the underlying issue. → Vineet Arora MD MAPP (Dean at Medical Education & Herb T. Abelson Professor UChicago Medicine): citing Dana Suskind's piece in TIME: We need a Hippocratic Oath for more professionals touching patients, including people working in AI for healthcare. More importantly, we want to hear from you: The Hippocratic Oath was written for a very different era of medicine. If you could add ONE line for doctors practicing today, what would it say? Drop your version in the comments. 👇
-
Jay Parkinson, MD, MPH shared thisIt took me three years of living in the mountains west of Boulder to figure out which weather tools to trust and when. Apple Weather is fine 300 days a year. But the other 65 days can strand you on a dirt road with no power, no water, and no heat. For those days I've learned to rely on a local meteorologist my friends call "The Donk" and a mountain-specific weather service called OpenSnow. Each tool is reliable in specific conditions and useless in others. Knowing which to trust when, and when to override all of them, came from years of paying attention to how my tools fail, not just whether they work. I've started calling this concept AI Failure Mode Literacy. And I think it's one of the most underdeveloped capabilities in professional life right now. Every experienced professional already has it for their traditional tools. A carpenter knows a circular saw binds in wet wood. A pilot knows GPS degrades in certain conditions. But AI is stochastic. Same question, different day, different answer. You can't learn the failure once and be done. You need to understand the patterns. All health professionals using AI today are being asked to trust these tools immediately, on high-stakes decisions, with no failure mode map at all.
-
Jay Parkinson, MD, MPH posted thisSo I’m deep in sales and starting conversations with companies building healthcare AI products. It's interesting. I'm trying to connect with these companies and sell them on our evaluation/validation services. Essentially, I'm saying "your products have weaknesses and you need a standardized, scalable process to evaluate them from a physician user's perspective, understand those weaknesses, and then target those weaknesses with fixes." It makes sense to do that if you're a product builder. Doing this in the pre-AI, deterministic product world became a ridiculously dialed-in science. But it’s pretty immature in the AI world at the moment. And of course it'd be great to have increasing revenue in Automate. Ultimately, what we're trying to do is create a standardized process for clinically validating a healthcare AI product. I'd love for this to be a collaboration between those building AI products and third parties like Automate. We’d help those products get better, define what the product's baseline looks like, and, by doing so, acquire third party clinical validation. We're finding that small internal teams are doing this in some companies, but we can also envision a world where small internal audits aren't the thing that buyers of these products will ultimately trust. How do y'all think this is going to evolve? Are these products going to need third party seal of approvals? Or will buyers trust that companies have done the work?
-
Jay Parkinson, MD, MPH posted thisMost clinical AI teams are shipping updates every two to four weeks. Their clinical eval cycle takes three months, if they even have one. And when they do, it's usually two or three physician advisors doing this on the side. Herding them onto a call, trying to hack together a consistent process, waiting weeks for feedback that may or may not be comparable across reviewers. By the time the results come back, the product those results describe is already gone. So how are decisions getting made in between? Mostly gut check and automated benchmarks that were never designed to catch what actually goes wrong in clinical AI. The answer is clinical feedback at engineering speed. Take 100 real user traces and put them in front of a physician. Ask them to write freely about what they see — no rubric, no forced ranking. What comes back isn't abstract. In one recent evaluation we did, our team of physicians surfaced consistent medication dosing and drug selection errors across four categories that nothing automated had caught. You find the themes, build targeted prompts around them, establish a clinical baseline, and now you can actually measure whether your product is improving. That process runs in days, not months, when clinical evaluation is someone's actual job rather than a side project. I wrote about how we're building this at Automate (link in comments).
-
Jay Parkinson, MD, MPH posted thisWhen radiology became a specialty, radiologists didn't just show up to read films that someone else decided to order. They developed the science of image interpretation, trained the next generation of radiologists, and built the professional standards that made the whole discipline trustworthy. The specialty was co-designed by the people doing the work. Clinical AI evaluation is a new specialty. I'm convinced of that. And it can't be designed by AI engineers alone, or handed off to a credentialing committee, or built from the top down by a company that needs cheap physician-hours to label data. It has to be built with the physicians doing the work. That's what the Automate Faculty is. Six weeks in, we have almost 100 vetted specialist physicians in our Slack community. These doctors get involved in client projects, yes. But they also debate evaluation methodology, push back on our frameworks, and help us figure out what this kind of work should actually look like when it's done well. And we're just getting started. If you're a physician who's been paying attention to what AI is doing to our field and want to be part of shaping how it gets evaluated, vetted, and held accountable, I'd love to have you apply. Link in comments.
-
Jay Parkinson, MD, MPH posted thisI had about 30 conversations at ViVE last week. I asked every healthcare AI company the same question. How are you getting clinical expertise into your development process? The most common answer was "we work with 3 or 4 doctors." These are companies building clinical decision support, diagnostic tools, triage systems. Products where accuracy is, well, important. And the clinical validation process is a handful of advisors giving informal feedback. This is the process because it's the best option available. There's no established playbook for how to systematically evaluate a clinical AI model. Nobody teaches it and there hasn't been a service you can call. So you find a few smart doctors, you ask them to review things, and you do your best. Everyone's winging it. It's sort of like inventing a medication and asking 3 or 4 people if they're feeling better before it hits the markets. The problem is that informal review has real limits. A few doctors catch what they happen to notice and miss what they don't. Nobody's mapped which clinical scenarios the model handles and which ones it doesn't. Nobody's testing for errors of omission, where the research shows 76% of severe clinical AI failures actually live. There's no structured way to know where models excel and where they don't. That's the gap. The healthcare AI industry has mature infrastructure for lots of things. But there's been no structured way to get the right doctors, with the right methodology, embedded in the development process. That infrastructure needs to exist, so we're building it.
-
Jay Parkinson, MD, MPH posted thisI've been talking to a handful of healthcare AI companies lately who are building their clinical validation capabilities in-house with small teams. Three or four doctors, maybe five. They're doing evaluation, writing training data, red teaming, labeling, the full spread of clinical work that goes into making a model actually safe and useful. And I keep coming back to the same concern: you're training your AI to think like three doctors. That might sound fine until you remember the product is going to be used across the full spectrum of wildly different presentations, comorbidities, and clinical contexts. Medicine isn't a field where one expert's judgment scales cleanly to an entire population. The concept of "the best doctor in the world" is a myth. Clinical excellence is distributed across specialties, experiences, practice settings, and honestly, across different ways of thinking about the same problem. Diversity of clinical perspective is a feature, not a nice-to-have. The other pattern I'm seeing is companies asking their clinicians, the ones doing actual patient care, to also handle the AI development work. Reviewing model outputs, writing gold-standard responses, designing evaluation rubrics, all layered on top of a full clinical workload. The context switching alone is brutal, but there's a deeper issue. Being a great clinician doesn't automatically make you great at AI work. These are genuinely different skills. Translating clinical reasoning into structured training data, designing adversarial test cases, building evaluation frameworks that give engineering teams something actionable. That requires a specific kind of thinking that not every physician has or wants to develop. What most companies underestimate is just how many stages of the AI development pipeline benefit from clinical expertise. It's not just labeling data at the beginning and checking outputs at the end. Nearly every meaningful step in between improves when a physician with the right specialty knowledge is involved, and ideally a different physician than the one who worked on the last step. This is exactly the problem we're building Automate Clinic to solve, a managed network of credentialed specialists who do this work because they're genuinely good at it and want to do it. If you're thinking about how clinical expertise fits into your development process, I'd love to talk.
-
Jay Parkinson, MD, MPH posted thisWe're in the middle of our first project at Automate. Here's what I've learned so far. When you're building a company that sits between AI teams and clinical expertise, the whole thing lives or dies on whether you can actually deliver. Over the last few months, we've built something exceptional from nothing. So proud of our team. Josh Emdur, DO and I met with the client's clinical leadership and co-designed the project over a few weeks. Three clinicians putting our heads together about the most valuable feedback doctors could provide on their model. The nuance was fascinating. Clinical thought filtered through the lens of actionable engineering strategies. We spent most of our time figuring out the right questions to ask our doctors to elicit the most valuable feedback. Then we set up the project in our platform and recruited the right team. And I'm just thrilled with who we put together. These are leaders in this space, and they each have a gift for writing thoughtful clinical feedback. A doctor who has a way with words and knows just the right language to use? That's rarer than you'd think, and it's everything. The deliverable isn't a spreadsheet of pass/fail scores. It's a map of exactly where and how the AI fails clinically, and what to prioritize fixing. There is no playbook for any of this. Every healthcare AI company is figuring out clinical evaluation on their own, reinventing the wheel in isolation. So here's my question: If you had a team of doctors trained in AI evaluation, what would you want them doing that adds the most value to improving your model? I don't think anyone has fully figured that out yet.
-
Jay Parkinson, MD, MPH posted thisWe've been talking to a lot of doctors who've done AI work, including evaluation projects, labeling, RLHF, and clinical review. The feedback is remarkably consistent. They want to do meaningful work shaping these tools, but the arrangement with the docs ultimately means they're disconnected from the meaning of the projects and the results of their efforts. Too often, they're treated as gig workers rather than partners in building something that actually works. They don't want to just rate outputs. They want to be part of the total solution. This is genuinely hard. Nobody has figured out how to systematically translate clinical expertise into artifacts that engineering teams can actually use to improve their models. There's no playbook. The whole eval-to-improvement pipeline in healthcare AI is immature, and that's being generous. Projects lack rigor. Doctors get pulled in, asked to flag errors, but there's no systematic method for discovering why the model fails, no framework for categorizing failure patterns, and no clear path from "this response was wrong" to "here's how you fix it." The clinical intelligence just evaporates. So that's what we're building at Automate. We want to systematically discover model weaknesses, categorize them in ways that actually mean something, and produce deliverables that translate directly into prompt improvements, training data, guardrails, or fine-tuning sets. Clinician-led and engineering-ready. Nobody has cracked this yet. And honestly, that's what makes it so damn interesting. We're geeking out on evaluation methodology every day because if healthcare AI is going to work, someone has to get this right. We're building the bridge between clinicians and engineering teams, and that's the simplest way to describe what we do.
-
Jay Parkinson, MD, MPH liked thisJay Parkinson, MD, MPH liked thisLLM spend is the new AWS bill. Not COGS that operates off revenue (this is coming soon), but an operating expense modeled in our unit economics. At Canvas, our LLM spending is now north of 10% of our payroll, and growing. So optimizing it has become mission critical. Claude has become infrastructure that scales with headcount, and we are learning the same lessons we learned about cloud spend a decade ago. Recently Andrew Hines asked us to audit our token spend for low vs. high ROI activity. Three patterns emerged in the data: 💰 The expensive activity isn't humans typing. The top cost driver last week was a background alert dispatch service, not a person. Developer API keys outpaced our heaviest individual users. The cost leverage is in the agents and pipelines, not the seats. 🧠 Reshaping information is cheap, agents are not. The "rewrite this email" workflow people worry about barely moves the needle. The agent that reads a repo, plans changes, and executes against them autonomously is where the dollars go. 💧Spend without an owner drifts. AWS taught us the playbook e.g. reserved capacity, annual commitments, usage dashboards, optimization investments. LLM providers will build the tools for this because customers will demand it, and there’s a win/win in driving stronger customer ROI from each $ spent on tokens. JP Patil already opened that conversation with Anthropic. tl;dr is we are already starting to budget LLM spend against headcount OpEx, give it ownership and attention, build monitoring and alerting tools, think strategically about our contract, and revisit the data on a regular cadence. We now need both FinOps and DevOps for LLMs. The companies that build it early will build compounding equity while the rest pay rent.
-
Jay Parkinson, MD, MPH liked thisJay Parkinson, MD, MPH liked thisToday, we're dropping the top voices shaping AI in healthcare part 2! 🔥 At Offcall, we believe deeply that physicians belong at the center of healthcare AI. We're fully committed to lifting up the AI leaders who are making medicine safer, smarter, and more human, which is why we're so excited to celebrate every person recognized on this list. Thank you. To our community: follow these voices, lift them up, and engage with their ideas because the conversation about how to shape AI in medicine safely, with physicians at the center, needs to be louder. If there's anyone else who should be on the list, tag them in the comments! We'll update and include your suggestions. Here are even more leaders shaping AI in medicine for the better 👇 🔷 Dr. John Lee 🔷 Dr. Eric Poon, MD MPH FACMI 🔷 Dr. Sara Murray, MD, MAS 🔷 Dr. Anthony Chang, MD, MBA, MPH, MS 🔷 Daphne Koller 🔷 Dr. Isaac Kohane 🔷 Dr. Rajiv Narula, MD 🔷 Dr. Jennifer Joe, MD 🔷 Dr. Dr. Keith J. Dreyer 🔷 Monica Agrawal 🔷 Dr. Adam Rodman 🔷 Peter Lee 🔷 Dr. Toby Cosgrove 🔷 Dr. Karen DeSalvo 🔷 Dr. Amy Abernethy 🔷 Dr. Jesse Ehrenfeld MD MPH 🔷 Dr. Andrew Mellin 🔷 Dr. Lynne Nowak, MD 🔷 Dr. Christopher Longhurst, MD, MS 🔷 Dr. Michael Blum, MD 🔷 Dr. Alistair Erskine MD MBA 🔷 Dr. Rasu Shrestha MD MBA 🔷 Dr. Daniel Kraft, MD 🔷 Tom Lawry 🔷 Dr. Michael Howell, MD MPH 🔷 Andrew Ng 🔷 Fei-Fei Li 🔷 Dr. Kalie Dove-Maguire, MD 🔷 Dr. Jay Parkinson, MD, MPH 🔷 Yann LeCun 🔷 Dr. Byron Crowe #AIinMedicine #HealthcareAI #PhysicianLeadership #Offcall #HealthTech #MedicalInnovation
-
Jay Parkinson, MD, MPH liked thisJay Parkinson, MD, MPH liked thisA few years ago, calling yourself a consumer health company in front of VCs was a virtual death sentence... but now, not only is consumer health being taken seriously, there are also some very serious consumer health businesses being built in the wild. This, plus other hot takes from Holly Maloney, Alison Ryu, Kurt Seidensticker and I at the inaugural Stanford Consumer Health Summit: ⚡ It's certainly possible now to build a large business purely based on cash pay + DTC acquisition. But playing nice with "The System" (e.g. B2B2C distribution, referrals from trad providers, taking reimbursement risk) still has its merits in helping one achieve even more durable scale and unit economics. ⚡ Sequencing of GTM motions matters: if you go "B2C first", you have the gifts of being able to fully control your product roadmap and swiftly acquire users - but it will be more expensive and potentially take longer to acquire them on your own. Going "B2B2C first" means you'll get access to large chunks of users in one fell swoop, but you'll have to pay a "product tax" to appease your B2B partner's requirements, not to mention needing to survive through long enterprise sales cycles. This sequencing choice informs your capital raising strategy in the early days. ⚡ A year ago, everyone was scared to say "AI Doctor", and now everyone is claiming to be building an "AI Doctor"! ⚡ To that end, will healthcare be dominated by an AI Doctor SuperApp as the front door, or will we continue to have fragmentation of apps by use case / condition / demographic? A large portion of consumers who don't have a PCP will find one in the form of an AI Doctor, and the best AI Doctors WILL be SuperApps that connect into a network of specialists and IRL clinical services through a single front door. ⚡ This necessitates a SuperData layer that allows for context sharing across all apps - and not just trad EHR data, but also wearables, genomics, non-traditional biomarker tests, information gleaned through ongoing engagement with AIs, patient-reported outcomes, health plan benefit design data info, etc. The buzz in the room felt a lot like the early versions of the Health 2.0 conference in the late 2000's, BUT with companies that are really working and scaling this time around. Congrats to Zach Teiger and his organizing team for bringing this community together - it's time to build in consumer health!
-
Jay Parkinson, MD, MPH liked thisJay Parkinson, MD, MPH liked thisjoked with my old cofounder about using the next 90 days to vibe replicate all the products and data tools that took us 3 years to build at our startup... I think i'll be done in 7. next up: AI agents to help hospital IT teams get thru every task that sits ahead of my ticket to grant me access to their Epic scheduling APIs...
-
Jay Parkinson, MD, MPH liked thisJay Parkinson, MD, MPH liked thisBig news in mental health AI safety today from Google and the Gemmini team. Perhaps the biggest to date for the entire space? This announcement matters because it reflects a broader shift: AI is no longer just an information tool, and the teams with the largest models are also investing real resources in making mental health AI safer. That is a powerful combination and forces us to consider whether 'general' AI systems will have more safety features and networks of support than any mental health-specific one can. We already got a glimpse of the challenges smaller mental health AI systems face around safety in https://lnkd.in/eUB8ngyV, but granted that is a cross-sectional sample (but at least neutral). Anyway, the answer to the impact of all this will come from how these new safety features and networks are rolled out and what happens next, but either way, this seems to be a win for mental health. -30 million dollar Google.org funding to crisis hotlines globally -4 million dollar expansion partnership with ReflexAI for organizations to scale mental health support services -New one-touch access to crisis support -Model improvements to better recognize and respond to acute mental health situations More details at: https://lnkd.in/eUPVVHjY https://lnkd.in/e9fdRH4p
-
Jay Parkinson, MD, MPH liked thisJay Parkinson, MD, MPH liked thisFirst day at Bridge today! Bridge is on a mission to simplify access to health insurance, a space long overdue for better and more scalable solutions. I’ve joined the team as the Founding Demand Gen Leader, supporting our kickass sales team as we grow. Grateful for this exciting opportunity and the journey ahead! 🥳
-
Jay Parkinson, MD, MPH liked thisJay Parkinson, MD, MPH liked thisThe NYT just published a glowing profile of a $1.8B company. I want to throw my laptop through a wall. Matthew built Medvi with two people, $20K, and a stack of AI tools. Compounded weight loss drugs, telehealth wrapper, insane growth. Sam Altman wants to meet him. LinkedIn is losing its mind. Cool story, but I'm not celebrating (at all). Because there are serious concerns and allegations circulating that NYT didn’t mention: (1) There are allegations that the company (or the partners they work with) built 800+ fake doctor Facebook accounts to run flash sale ads for their drugs. (2) They are named in a lawsuit alleging a nationwide scheme to manufacture and sell a fraudulent, unapproved oral tirzepatide pill. (3) Examples have been shared of onboarding flows accepting clearly unrealistic inputs (e.g Feb 31 as birthday or telling them they had a 94% chance of hitting their goal weight of 200lbs starting from 7'11" and 350lbs). (4) Plenty of 1-star reviews claim undisclosed charges, no refunds, product never delivered, and impossible to cancel. You can decide for yourself what to make of that. If it smells like a fish, looks like a fish, it's probably...... 🐟 🐡 🐠 We watched the Adderall crisis unfold. We watched what happens when profit motive runs ahead of clinical guardrails in prescription drugs. And now we're celebrating the guy who figured out how to do it faster with AI. What we should be celebrating are companies that use AI to make clinical care more accessible and safer. Those two things are not in conflict. But they do require more than two employees and a vibe. tldr; We shouldn't celebrate velocity without asking what corners were cut. In healthcare, those corners are usually patient safety. H/t to Colin Morelli, Brendan Keeler & Sheel Mohnot. PS: Sharing my perspective based on publicly available information, including allegations and user-reported experiences. Open to corrections.
Experience & Education
Recommendations received
1 person has recommended Jay
Join now to viewOther similar profiles
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content