Vinod Khosla argues that AI can automate routine physician tasks and let patients manage their own health. Tools like Isabel already match human diagnostic performance, and Kaiser Permanente cut stroke deaths 40% with a data-driven statin protocol. He calls for entrepreneurs to build AI health assistants outside the traditional system.
Chapters
Good morning. I'll keep that iPad down. I thought Twitter feeds are great to watch watch questions. I'm going to, given it's a complex topic, I thought I'd go through it real quickly, what my views on the subject are. What screen are you guys seeing? I'll let you— so let me talk about 2025. And let me talk about whether we need doctors then. This is trial and error. This is speculation. But let me speculate. Let's start with the way the world is. I won't repeat all this. There's a huge misdiagnosis and missed diagnosis rate. There's a lot of patient deaths in U.S. ICUs. Worldwide, it's a much larger number. Diagnostic errors are common, and they result in a lot of costs. Especially insurance costs, which gets in the way of good medicine. There are a lot of other problems. Even when you know there's high-risk diagnosis, and I'll post this on my website and you can look at the references in the notes section, there is data for those, that diagnosis in the patient record that's often not not used. And then you got the other problem of the number of people who don't have enough health literacy to even understand what the doctor is saying, what the healthcare system can offer them. You, you get my point. Here's the surprising part. In a small study Most patients, because they're so intimidated by doctors and the hurried doctors and their busy, overburdened doctors, most patients prefer getting their discharge information from a computer agent than from a doctor. Here's the math most people don't understand. And so I know each one of you has the best doctor in the world, but somebody has the bottom 50%. Human doctors have cognitive limitations, and a huge source of error is cognitive biases, especially our confirmation bias. This was done at the Cleveland Clinic. Pretty significant percentage of the time, if you got a serious second opinion, it significantly changed the treatment course. These are all symptoms that there's something wrong. Now people like me mouth off all the time without knowing a lot about medicine, and I'm the first one to admit I know very little. I'm unburdened by reality. And that's what's needed, because if you have— know too much about what to do and what to expect, you're not going to try different things. Entrepreneurs— and I mostly want to speak here to entrepreneurs who want to come talk to me about how you change the healthcare system not from the inside but from the outside going in. Can explore assumptions, sometimes naive, and iterate rapidly. So let me move to the gray zone of what I call speculations. Patients will be CEOs of their own health. I try to be, but the system fights me all the time. What test I can run, how often I can do it, 80% of what MDs do can be replaced. Let me explain how. Machines are better at integrative medicine. Machines can integrate way more than doctors can. Not just the symptoms, but demeanor. You can see what's going on a patient's micro impressions on their face, and a system can tell that a lot better than a doctor can, especially an overburdened, hurried doctor trying to handle 3,000 patients in a typical year. You can monitor what patients are doing on their phone. We just invested in a company that actually monitors how often you text or email or who you call or how often or if you even leave your bedroom. Because they're going after treating and being an assistant to psychiatrists for mental health patients. The thing to remember is there are some things doctors will do better, and maybe they will be used for that, but a new system has to be much better overall and not be really bad on anything. But it doesn't have to be better than a human doctor, especially that average human doctor, in everything. If it's better overall, significantly better overall, that's the standard I'm talking about. I fundamentally believe one of the problems with medicine is this idea that first do no harm. It's stupid. If you save 10, uh, if you save 1,000 lives, and lose 10, you should do it. But our current drug trial system probably wouldn't allow that. There is a human element of care, and I'm not saying that won't be needed, but you don't need an MD to do that. And some MDs are pretty humane. But let's find the most humane humans to do the human element. Why do they have to be doctors? Now people will laugh at the early systems. Think of those as toddler MDs and digital first aid kits. That's how this transition will start. Remember that phone? That's what a phone was. Today it looks a little different by version 5 or 10. And that's what we'll see. The best MDs will train these systems just like a med student over 10 years. And the systems will pay back in helping these MDs with bionic assists, maybe extra diagnosis, maybe amplify their capabilities initially. They look pretty clumsy in the beginning, point— points of innovation, but then once we have enough data, enough big data, we'll discover surprising things. This little system, iASTI, using the National Breast Cancer Database, in the first 3 hours in the hands of somebody who didn't know about cancer and never done anything, discovered new kinds of breast cancer. And these innocuous innovations will connect up and sound like a wave in a tsunami. I have nothing against MDs. In fact, they've done very well in the practice of medicine. But they have limitations. And there are misalignment of of incentives. No hospital system wants to cut its revenue in half. No pharmaceutical company wants to cut its revenue in half. And we don't use data. Just one thing at Kaiser Permanente, how you use statins, cut the death rate for stroke victims coming into the door by 40%. There's a hundred such things to be discovered. This one study with medical assistants one year after high school, one year of training, was 91% accurate, accurate with the CHAMP system without exams. Another one study with the same system could do 75% or 76% safe triage, eliminate that much load off the patient, off the doctor. The Isabel system is a system developed in the UK, matched expert diagnosis 75% of the time. And remember, these are toddler systems. And the second version got much, much better. So think of Dr. A, my doctor algorithm, as V0 right now, like that big phone that was floor-mounted on your— on the self— on the floorboard of your car. And imagine the system in 2025 looking like the iPhone, with all kinds of attachments and all kinds of things it could do. And Dr. House, plus a lot of other things, will train these systems. They won't learn by themselves. And won't need manners. But they will have better manners, maybe even better than the best doctors. And caregivers will have more time to talk to the patients. The tradition, the tradition of medicine, the practice of medicine, and that's what's wrong with medicine today, it's the practice of medicine, will become the science of medicine. A simple thing like should you reduce fever, and many of you MDs, and so all of you practice this for critically ill children, has never been tested in medical practice. Turns out below 104 degrees, it is the wrong thing to do. The number of deaths in a study at the University of Miami Hospital was 7 times larger in a— it was about our 80-patient study, there were 7 deaths among people whose temperature was treated the way practice of medicine does. Only 1 death among patients where there was no reduction in fever. And then there's the rest of the world we don't think about, where most people don't have access to doctors, where the doctor-patient ratio is 10% of what it is in this country. It is an under-resourced world. And this shift to computerization has happened in other places, in other areas. If you flew here, your pilot didn't fly that plane. It's way too complex for your pilot human being to fly it. Most stock trading was the most human of judgment 15 years ago. You'd get your clock cleaned if you tried to do that without computers. Most of it is automated trading. And Google's driverless cars driving around the streets, 300,000 miles without an accident. Why not MDs? Now, I'll be wrong on most of the specifics, but I think directionally this trend is going to happen. So let me stop here. I know I was going to do questions, but we don't have time for questions. I decided that given my views on this have been misrepresented a lot, that I'd sort of just explain what I mean and not have knee-jerk reactions from much that I've seen in many of the blogs. Thank you all very much.
