AI and Health Care
David Kelnar sent me this deck that he did on the state of AI. It is very good.
This slide got my attention:
It is interesting, and not totally surprising, that the sector that AI-focused entrepreneurs are targeting more than any other is health and wellbeing.
It seems like there is so much opportunity to improve our collective health and wellbeing with data science and machine learning. This is a big part of our thesis around healthcare at USV.
Bits verses atoms, again.We appear to be marching forward with promising (and promised) innovation but as our natural and built environment crumbles around us. More of the world’s accumulated private wealth needs to be unlocked for collective public benefit. We need a new economic model, and soon. The one we have right now works counter to our collective interests.
Doesn’t that new economic model exist in the selling of data already? Shouldn’t we instead pass on some of the revenue to folks whose data is being sold? Sounds like an economic model to me if there is less centralized data ownership and more of a distributed model.
I was thinking in my own mind more about the model for atoms than the model for bits. The bits economy is racing along. Atoms are not.
You’re taking on a lot here, but I suspect that the argument for AI in healthcare would be Bits rather than atoms for the greater ultimate benefit for atoms. RE: “New economic model” I can’t devote the time/energy to debate that, but the inertia of the status quo is going to be a problem for your cause
Great deck indeed. “Developers will have a greater impact on the future of healthcare than doctors”- that’s a pretty loaded statement, but one could argue that doctors also benefit from these AI advancements, which in turn helps them to solve unresolved health research areas and patient care situations.My friend Amar Varma recently donated $1.75M towards creating the first of its kinds AI position at Sick Kids Hospital in Toronto.https://www.theglobeandmail…I’m waiting to see what kinds of vital sign sensors will Apple incorporate in their earpods next year.
Andrew Trask at Open Mined is doing some cool stuff with federated learning that will allow machine learning to unlock value from data that cant go offsite.
The future of healthcare will be found in the Venues of ASCO ACC etc. and the science and bio pathways (MTOR etc) of the human body. The ML AI algorithms are a tool to researchers, nothing more. Sorry but The odds of USV making a dent in a breakthrough in healthcare are between slim to none, and that was slim who left the building.
nowhere did i suggest we will make a breakthrough. we don’t invest in R&D, or even a lot of cutting edge stuff. we are investing in the platforms that will bring all of that to market
You made the bold statement – not me. Platforms to market …. Platforms of markets to do what ? Most people in the US will loose their health due to cancer – hearth disease – vascular disease – diabetes – and aging. Making it easier for kids to buy viagra or women to buy birth control is not an investment in healthcare, its simply a way to make $, it’s an investment in loopholes.With all due respect – in the US alone we we spend 200B on healthcare R&D – AI may improve R&D targets – see the work of Patrick Soon Shiong – but to date – out side of of structured data the AI can’t differentiate an endoscopy from a colonoscopy.
Given the sheer number of illnesses derived from poor diet and lack of activity, such as ALL of the ones you mentioned, it seems possible that a breakthrough could come in the form of a platform that embodies USV’s stated focus on general health and well-being and drives/educates/informs people on the path to taking their health in the own hands. There is rapidly declining faith in the food and healthcare industries so I suspect the market is ripe for change. I’d hesitate to predict that biochemical pathways are the only place the silver bullets lie.
The era of VCs growing unicorns on a vague idea like Twitter / Facebook is in the books. The probabilistic approach to Diet Fitness is well known for delaying disease and is mostly well understood. It not an AI issue. It’s a tradeoff.
Right. But you are missing this point that many of the improvements to health outcomes may come through helping/inspiring/educating people to take advantage of the knowledge we already have to improve their health, rather than producing new knowledge.
Helping/inspiring/educating people about diet and exercise is a business not a VC investment.
There’s a very powerful role for a platform backed by AI to play in shifting human behavior to improve health outcomes. Venture capital has a role to play in scaling candidates. Why all the negativity about this important discussion? Got burned by a VC exit? A VC demand all your equity?
Ah, there’s more! Since I used to work in an AI project at IBM’s Watson lab, I regard AI as largely incompetent and something between a colonoscopy and an unanesthetized root canal procedure, really nearly all hype, for anything having to do with medicine, irresponsible hype, e.g., from describing largely heuristic computations with wildly inappropriate anthropometric terminology.It’s an old scam: Computer history can show that back in the days of vacuum tube computers commonly the machines were described as “giant electronic human brains”.This AI scam has gone from spring of hope, summer of hype, fall of failure, and a cold winter, more than once. Maybe the next time the fad, hype bubble bursts, AI, especially for medicine, will get buried with a stake through its heart.
Thanks Fred.A16 podcasts have a deep tech series on medical innovation that is really good.Listened to one the other day while traveling that dealt with the economics of custom gene editing for rare diseases that broke all the hard edges of my thinking.
Exciting stuff around AI and health care happening here at Mount Sinai: https://www.mountsinai.org/…
Okay, okay, okay, okay, want “AI” for health?Okay, one problem is, “Here is Joe and all the data we have on Joe. Is Joe sick or well?”Here we want a known and adjustable false alarm rate. And for whatever false alarm rate we select we want the best detection rate (of course if we are willing to accept a higher false alarm rate, then we get a higher detection rate, lower rate of missed detections, of real sickness), e.g., as in my especially general proof of the Neyman-Pearson result from the Hahn decomposition from the Radon-Nikodym result. Well, we don’t have data enough for Neyman-Pearson, but there is still a weaker sense in which, for whatever false alarm rate we select, we get the highest detection rate possible — darned good to know and a nice companion to the Neyman-Pearson result.Using some advanced pure math prerequisites, I created new applied math theorems and proofs for new methods for just that detection problem. On some real data, the work looked fine. On some really challenging synthetic data, it looked terrific.So, yes, the work is in the framework of mathematical statistics and, in particular, hypothesis testing. But the work is novel in that it is both multi-dimensional (uses data, say, essentially jointly on several variables) and distribution-free (makes no assumptions about probability distributions), features crucial for the real problem, and may be the first test with these features. Moreover, for any alarm, we can be given the lowest false alarm rate at which the data is still a detection. In this way we get a measure of seriousness of a detection.Moreover, taking an expectation of a conditional expectation, the Radon-Nikodym result again, shows that the false alarm rates are unbiased, intuitively, in the long run, exact. A bit amazing.I published the basics of the work in Information Sciences.History: I was in an AI project at IBM’s Watson lab where the project was monitoring server farms and networks for sick or well. For something much better than our AI approaches, I did the work for the methods just described.I’ve invented some algorithms, not in the paper, to make the computations fast. Current solid state disks should be terrific for these algorithms.My work is not based on (i) a statistical model as in what Leo Breiman did when he pushed old regression analysis to more for such work on clinical medical data as inLeo Breiman, Jerome H. Friedman, Richard A. Olshen, Charles J. Stone, “Classification and Regression Trees”, ISBN 0-534-98054-6, Wadsworth & Brooks/Cole, Pacific Grove, California, 1984. better known as CART, andLeo Breiman, “Statistical Modeling: The Two Cultures (with comments and a rejoinder by the author),” Statistical Science, Institute of Mathematical Statistics (USA), Volume 16, Issue 3 (2001), 199-231. or (ii) just empirical curve fitting as in the machine learning work with neural networks.Disclosure: Breiman is one of my favorite authors, not because of these two works above but for his super clear treatments inLeo Breiman, Probability, ISBN 0-89871-296-3, SIAM, Philadelphia, 1992. of the strong law of large numbers, the martingale convergence theorem, measurable selection, and diffusion processes! His result on measurable selection was just what I needed in my Ph.D. dissertation on stochastic optimal control!!!! Fun days!!Comparisons:(1) It appears that mostly current AI, etc., for this problem is based on statistical models based on extending regression analysis as Breiman described or on just empirical curve fitting as in neural networks.(2) From my reading of Breiman, it will not be so easy with his methods to do so well on false alarm rates and detection rates.(3) For neural networks, my guess would be that doing well on these rates would be still more difficult.(4) My methods can do well with large amounts of data, say, as needed for Breiman’s approaches or neural networks, which tend to need much more data. Generally needing huge amounts of data pushes the work into less bias but also less specificity.(5) My methods also do well with small amounts of data and, in particular, do not encounter the severe problem of model over-fitting.(6) It may be that some resampling techniques could help Breiman’s work and neural networks to do something on the rates of false alarms and missed detections.Ah, it’s fun to see some math with theorems and proofs and actual powerful, valuable results in practice beat, make look totally brain-dead silly, blow out of the water with the doors blown off AI, machine learning, and data science, along with their buddies alchemy, flogiston, dowsing rods, genuine, 100% snake oil, and witchcraft!Uh, guys, if want to see farther than others, then stand on the shoulders of giants, e.g., A. Kolmogorov, E. Hopf, J. Doob, J. Neveu, P. Halmos, K. Ito, L. Breiman, etc.
sad that Agriculture is last… not for long!
Hi, David here (wrote the deck). Let me know if you’ve any questions I can address!
Hello David, Have you published additional material on the AI vendor landscape? It is a confusing terrain with mixed semantics and more spin vs. actual value delivery. Any research you can share will be incredibly valuable.
Hi there. Check out mmcventures.com/research for materials we’ve published. The State of AI report includes a market map of all early stage UK AI vendors – might be useful? It also includes quantitative analysis of all the early stage AI vendors in Europe. Hope this helps and happy to chat further – best wishes, DK
time for VCs to spend some time here: https://medicine.illinois.edu/
Healthcare is the single largest market AFAIK, so this makes sense.
Some of the AI efforts have made impressive progress, but the MMC materials claim too much for AI and neglect too much about terrific results from the past.A lot of what MMC claims is good AI is essentially good old wine with new AI labels.Also too much of the MMC materials is at least uninformed. E.g., the MMC materials claim“Fraud detection algorithms enhanced with AI can identify fraudulent transactions, while reducing false positives, more effectively than traditional approaches.” This remark is wildly uninformed and even incompetent: My post earlier today on anomaly detection can be used for fraud detection and offers false alarm rates settable and knowable in advance. As in that post, false alarm rates can be as low as one pleases.AI is a way for software to perform difficult tasks more effectively, by learning through practice instead of following rules. Not following “rules” as in old expert systems is wise but avoids just one way, of millions, of ways to make a mess.This “learning through practice” by itself means nothing; in some particular cases, with some probabilistic assumptions, might be able to say something. E.g., there was some classic statistics that did really well detecting a cheating roulette wheel.AI is important because, for the first time, traditionally human capabilities can be undertaken in software efficiently, inexpensively and at scale. That was true for an abacus, the first telescope, slide rules, mechanical desk calculators, software playing a perfect game of Nim, with the fast Fourier transform detecting audio signals much better than humans, etc. But none of these, and so far NOTHING has anything like the intelligence of humans, apes, monkeys, orcas, dolphins, dogs, cats, raccoons, ravens, etc.AI capability has reached an inflection point. After seven false dawns since the 1950s, AI technology has come of age. It’s still awash in hype and not at all intelligent. AI has numerous, tangible use cases. We describe 31 across eight industries and highlight why some industries will be affected more than others. So far, AI is WAY behind the accomplishments of pure and applied math. If want good progress in practice, then “stand on the shoulders of giants” and use the best tools for processing available data — pure and applied math.Advances in AI technology are creating new possibilities. Custom silicon is enabling a new generation of AI hardware. Emerging software techniques, including reinforcement learning and transfer learning, are delivering breakthroughs in multiple domains and freeing system design from the constraints of human experience. New, generative AI will reshape media and society. For decades and even centuries, essentially all serious work in physical science and computing have been thoroughly freed “from the constraints of human experience.”Demand for AI talent has doubled in 24 months. That sounds like the ads for workers for the orchards of California during the Great Depression. Two times 0 is still 0.Europe is home to 1,600 AI startups. I have a startup: It makes powerful, valuable progress on meaning in real human intelligence. But calling my work AI would be a grand insult.Healthcare is a focal point for AI entrepreneurship. Activity is thriving given new, transformational opportunities for process automation through AI, and stakeholder engagement. As I described in this thread earlier today, I’ve had work terrific for detection of medical illnesses for 20+ years with absolutely no practical interest at all. While AI is still floundering around with that problem, I have a terrific solution close to best possible from any means of processing the data. I see nothing “thriving”.Competition for talent, the limited availability of training data, and the difficulty of creating production-ready technology are entrepreneurs’ key challenges. For my work, as above in this thread, and my (very different) startup, I’ve got all that well handled, and that fact means NOTHING.Get in touchAt MMC Ventures, AI is a core area of research, conviction and investment. In the last 24 months we’ve made 20 investments, comprising 50% of the capital we have invested, into many of the UK’s most promising AI companies. If you’re an early stage AI company, get in touch to see how we can accelerate your journey. AI is too weak to be of interest. From the magnificent history of pure/applied math, it is MUCH more powerful than AI; so is my work described earlier in this thread and my startup.I doubt that there is anyone on Sand Hill Road able even to direct a competent review and evaluation of my work.‘AI’ is a general term that refers to hardware or software that exhibit behaviour which appears intelligent. Then AI is nearly the empty set. It is so far in practice nearly always at best just some computer programming for some narrow work in some narrow fields using some weak techniques.Machine learning enables programs to learn through training, instead of being programmed with rules. By processing training data, machine learning systems provide results that improve with experience. Without some assumptions, this statement has counterexamples.Machine learning can be applied to a wide variety of prediction and optimization challenges, from determining the probability of a credit card transaction being fraudulent to predicting when an industrial asset is likely to fail. AI for optimization, a well developed field, is JUNK. NO WAY is AI able to compete with optimization — linear programming, Lagrange multipliers, Kuhn-Tucker conditions, quadratic programming, integer programming, multi-objective programming, capacitated network flows, deterministic optimal control, stochastic optimal control, etc. NOT even a weak little hollow hint of a tiny chance.Deep learning emulates the way animals’ brains learn subtle tasks – it models the brain, not the world. The use of “deep” is a deliberate and deceptive misuse of words: What is deep is just the number of stages in a neural network.Deep learning emulates the way animals’ brains learn subtle tasks – it models the brain, not the world. Absolutely 100% total outrageous incompetent nonsense. There is no evidence of any connections with, say, mammalian biological brains, in particular because so far no one has even a weak little hollow hint about anything significant in such brains, learning, or intelligence.For all those claims, there is a much more powerful path — pure/applied math. Such math has terrific methodology, and AI and computer science are fatally short. General analytical tasks, including finding patterns in data, that have been performed by software for many years can also be performed more effectively using AI. Counterexample: AI can’t do at all well solving significant differential equations, Maxwell’s equations, the Navier-Stokes equations, etc.Some years ago I took a problem in 0-1 integer linear programming with 40,000 constraints and 600,000 variables, did some derivations in non-linear duality theory, wrote some code to call the IBM OSL (optimization subroutine library), and in 500 primal-dual iterations found a feasible solution, from some bounding, guaranteed to be within 0.025% of optimality. There is NOTHING in current AI able to do any such thing.
Related: I wrote this fictional story after using a “Lego Recycling Machine” as an AI/ML teaching example for my kids. https://medium.com/@brucewa…
In case you are interested in the FDA status of some of these AI based algos:https://www.linkedin.com/fe…
The knowledge is already there to greatly improve everyone’s health.