Google’s Head of AI Talks Concerning the Way forward for the EHR


This transcript has been edited for readability.

Eric J. Topol, MD: Good day. That is Eric Topol with Medication and the Machine, with my co-host, Abraham Verghese. It is a particular version for us, to talk with one of many main lights of synthetic intelligence (AI) on this planet, Jeff Dean, who heads up Google AI. So, Jeff, welcome to our podcast.

Jeff Dean, PhD: Thanks for having me.

Topol: You will have now been at Google for 22 years. In a latest e-book by Cade Metz (a New York Occasions tech journalist) known as Genius Makers, you’re one of many protagonists.

I did not know this about you, however you grew up throughout the globe. Your mother and father took you from Hawaii, the place you have been born, to Somalia, the place you helped run a refugee camp throughout your center college years. As a highschool senior in Georgia the place your father labored on the CDC, you constructed a software program instrument for them that helped researchers accumulate illness knowledge, and practically 4 a long time later it stays a staple of epidemiology throughout the growing world. I will cease there as a result of I did not know this, and there is this factor known as a pandemic. Are you able to assist us?

Dean: My father was an epidemiologist, so we traveled round, and my mother studied medical anthropology. That mixture of careers, plus a little bit of wanderlust, led to going to new locations, and I used to be an solely baby so I went alongside for the experience. We ended up dwelling in plenty of completely different locations. I did a highschool internship on the CDC, writing some software program for outbreak investigations throughout epidemics. Epidemiologists use some very specialised statistics that almost all statistical packages do not help very properly, and it is vital to supply software program that may be run all around the globe on the comparatively low-end computer systems at the moment. So I began writing software program that would supply the proper of information assortment and evaluation instruments that could possibly be run in a reasonably light-weight means by epidemiologists all around the globe.

This was earlier than the web, so to distribute it, individuals would present up with floppy disks and I might put the software program on it. I’d make copies and educate individuals easy methods to use it, and it propagated from there. And it is nonetheless my seventh or eighth most-cited work.

Pandemics, Then and Now

Abraham Verghese, MD: Jeff, it is such a pleasure to get to speak to you. There’s a variety of cross-pollination between Google and Stanford, and a few of our of us have skilled with you. The epidemiology software program you are speaking about, is that the identical factor that you just labored on for HIV/AIDS for the World Well being Group? Was that an evolution of that software program? That brings me to the query of what you assume of the present pandemic. I do know your energetic thoughts should have turned to that as properly.

Dean: The best way the software program advanced was that I began it in highschool, and each summer season I’d do a brand new model of it. The primary couple of summers I did it on the CDC as type of follow-on internships. Then one among my colleagues, who was working on the CDC, moved to the WHO in Geneva and mentioned, “Do you need to come to Geneva for the summer season as a substitute of Atlanta?” and I assumed That sounds new and thrilling — why do not I do this? So the following few variations of the epidemiologic software program have been carried out there, and I additionally began engaged on some software program to assist predict the longer term trajectory of HIV infections and AIDS instances around the globe, based mostly on early assessments of seroprevalence charges in numerous communities, working with the International Programme on AIDS there. It was an evolution of the work I used to be already doing and a few follow-on extra work to assist do extra forecasting particular to the HIV pandemic.

The present pandemic is clearly tragic in a variety of methods and has dramatically affected the world. HIV could be very completely different from COVID in some methods, however it’s additionally reflective of the general public well being group needing to return collectively and inform individuals what to do to maintain themselves protected. As we be taught extra, the steerage adjustments. However COVID, being an airborne illness, is way more transmissible in some ways than HIV. There are some commonalities but additionally some variations. The response of the general public well being group has been fairly good right here; the deployment of vaccines in such speedy vogue has been fairly outstanding and a credit score to all of the scientists concerned in that, as a result of it helped the world transfer from what may have been even worse to a greater state of affairs. Clearly, we aren’t utterly out of it but.

These Google Cats

Topol: The deep studying period, by which you performed a serious function, appeared to get legs with the cat video story. Are you able to inform us a bit bit about that?

Dean: I had been launched to neural networks in 1990, when there was an preliminary wave of pleasure about neural networks and what they might to. At the moment, they might present actually fascinating outcomes on teeny, tiny issues, however they could not scale to something of significance. It appeared like an fascinating means of approaching and fixing some sorts of studying issues, so there was an enormous wave of pleasure however then a trough of disillusionment.

However I used to be a senior in faculty throughout that first wave of pleasure, and I assumed, Oh, these are nice! What if we may simply expand ones? So I did a senior thesis on parallel coaching — utilizing 32 processors to coach neural networks as a substitute of only one — as a result of I assumed perhaps we would have liked only a bit extra laptop after which we may make them do superb issues.

I used to be utterly mistaken. It turned out that we would have liked about 1,000,000 instances as a lot laptop, not 32. However then beginning round 2006, 2008, we began having that a lot laptop energy due to Moore’s Regulation on this planet. Researchers in a number of universities have been beginning to see good outcomes on utilizing neural networks for a broad set of issues — speech, early laptop visible issues. I heard about that, together with others at Google, and we determined to begin a venture to coach very massive neural networks utilizing the computer systems in our knowledge facilities. We put collectively some software program that enabled us to coach utilizing 1000’s of computer systems.

We determined to coach a mannequin utilizing an unsupervised neural community. So we took 10 million randomly chosen frames from YouTube movies, and we skilled a mannequin utilizing an unsupervised studying algorithm. The system discovered to acknowledge a complete bunch of various objects. After all, it discovered to acknowledge cats as a result of YouTube is filled with cats. The actually fascinating factor about that’s that it developed a neuron with the power to acknowledge whether or not a cat face was within the body with out ever being informed what a cat was. So simply from patterns and knowledge and constructing larger and better ranges of abstraction by this unsupervised studying algorithm, the system developed a neuron that would inform whether or not there was a cat face there or not. That was fairly outstanding.

You possibly can then practice it with some supervised knowledge, saying it is a automotive, or a truck, or no matter, and get higher outcomes than if you happen to simply did the unsupervised coaching.

It was eye-opening for us. It confirmed that scaling these sorts of fashions to bigger sizes would produce good outcomes, and it began a complete decade of profitable makes use of, at Google and elsewhere, of neural networks for all kinds of issues — speech recognition, photographs, language understanding, and so forth — in addition to some functions in healthcare and medical diagnostics.

EHR of the Future

Verghese: One of many issues in healthcare that enormously pursuits our viewers is the digital well being document (EHR), which as you recognize has been each a serious boon and a serious supply of frustration with the period of time we spend on these programs. They’ve probably not lived as much as their preliminary promise, however what’s very thrilling is a number of the work that you have carried out on pure language processing.

The place do you see the EHR evolving? Give us a way of hope, if you’ll, for the longer term.

Dean: Docs have a little bit of a love-hate relationship with the EHR and the best way they must work together with it as a result of it takes away time they could possibly be spending with sufferers. But it surely’s additionally a repository of actually invaluable details about selections they’ve made about sufferers and the outcomes.

In collaboration with different organizations, we have carried out some work on utilizing de-identified knowledge in fashions which might be much like how we practice pure language fashions. What you need to have the ability to do in pure language is to take a prefix of a chunk of textual content after which predict the following phrase or sequence of phrases that’s going to happen. In case you are typing an e mail message, that may show you how to by suggesting the way you would possibly full the sentence to save lots of typing.

It seems that very same strategy can be utilized to offer clinicians recommendations about what would possibly occur subsequent within the medical document for a specific affected person. If you concentrate on the medical document as a complete sequence of occasions, and when you have de-identified medical information, you possibly can take a prefix of a medical document and attempt to predict both the person occasions or perhaps some high-level attributes about subsequent occasions, like, “Will this affected person develop diabetes throughout the subsequent 12 months?”

You’ll be able to turn out to be fairly good at predicting a variety of issues that clinicians would possibly care about in fascinated with easy methods to deal with a specific affected person or a specific situation of a affected person they’re seeing. You’ll be able to, for instance, recommend 5 diagnoses which may make sense, given the affected person’s present signs plus their previous medical historical past, based mostly on studying from different de-identified medical information.

Clinicians go to medical college, get skilled, and see about 20,000 sufferers throughout their careers. That is extraordinarily helpful, however it additionally means they could have pretty restricted expertise, particularly with uncommon issues they could by no means see. An aspirational aim, which is absolutely laborious to attain for some time bunch of causes, is with the ability to use each previous medical determination to assist inform each future medical determination. That might be nice, as a result of we might be studying from the collective knowledge of what labored and what did not work on billions of individuals, with a view to present higher care for everybody sooner or later. That is difficult, however it’s a very good north star about what we would be capable of obtain if we put our minds to it.

Topol: I do assume it is achievable — the entire concept of a digital infrastructure with twinning and nearest-neighbor evaluation — if we may get the information. Kai-Fu Lee and I wrote about this type of massive considering in “It Takes a Planet.” Sometime that might be extraordinary.

Well being and the Human Eye

Topol: You and I have been at a convention (and by the best way, it is nice that there are such a lot of publications from Google on AI in healthcare), however one space that was hanging to cement the ability of coaching neural networks was the retina. And you may recall — not as a result of that is the best way we need to decide whether or not an individual is a male or feminine — however by placing a photograph of the retina by a neural community, as a substitute of fifty% accuracy from famous retina professional Pearse Keane, you possibly can get 97% accuracy on whether or not it was a male or feminine. That means that human eyes are not so good as skilled machine eyes. Can we are saying, firmly proper now, that the mix could also be even higher?

Dean: The fascinating factor about that work is {that a} neural community that had been skilled on a bunch of retinal photographs may be taught to foretell organic intercourse from the retinal picture alone. And it may additionally predict different issues related to cardiovascular threat that ophthalmologists and even skilled retinal specialists could not essentially decide up on. That tells me that lurking in well being info are refined indicators that, with the correct framing of the issue and machine studying, would possibly enable us to choose up issues which might be complementary to what human clinicians already decide up on. The mixture of those computing programs plus skilled professionals will end in higher outcomes than both can obtain alone.

Topol: It is actually extraordinary to see what has already been constructed out, not only for the potential of monitoring diabetes, however glucose regulation, blood strain regulation, hepatobiliary illness, and the calcium rating of the coronary arteries, and even Alzheimer’s illness, probably utilizing the retina as a window to neurodegenerative ailments. The record retains getting longer.

Google has carried out a variety of work on diabetic retinopathy, for which half of all these with diabetes are by no means screened. You have labored with India’s Aravind Eye Hospital. The place does that work stand? Is that the way forward for screening for diabetic retinopathy?

Dean: Diabetic retinopathy is among the earliest issues in medical imaging diagnostics that we began taking a look at, as a result of there’s an enormous want for extra screening capability around the globe. In lots of elements of the world, there simply aren’t sufficient clinicians, who want a good quantity of particular coaching to evaluate a retinal picture for indicators of diabetic retinopathy on a five-point scale. If you happen to catch it in time, you possibly can forestall blindness or partial lack of imaginative and prescient. It is really very treatable, so the power to display screen extra individuals will really forestall blindness.

We have adopted a reasonably detailed development from early analysis initiatives to this risk — whether or not we are able to really practice a machine studying mannequin to evaluate whether or not a retinal picture reveals diabetic retinopathy. The earliest research we did confirmed that it was on par or maybe barely higher than board-certified ophthalmologists in the USA. With some extra refinement of how we labeled the coaching knowledge and the way we skilled the mannequin, we’re in a position to get it to be on par with retinal specialists who’ve extra coaching on this, which is the gold commonplace of care on this space. After which we have been working with our companions in numerous locations, together with India, Thailand, Germany, and France, to deploy this in actual scientific settings and do screening. We have simply reached a milestone of fifty,000 sufferers screened with this strategy, and the numbers are persevering with to go up. It is good to see issues that aren’t simply analysis papers however are additionally really deployed on this planet.

We have been taking a look at different modalities of medical imaging. After we noticed the early success with diabetic retinopathy, we realized that this was a usually repeatable sample with different medical imaging modalities and we may in all probability get good outcomes. There’s cautious consideration to plenty of issues, like gathering the proper of coaching knowledge, coaching the proper of mannequin, evaluating it appropriately, and so forth.

Now That Computer systems Can See

Verghese: One of many issues that your group has carried out so properly is to create open-source packages or methods of consolidating the method of performing AI. Studying about your work, it appears to me that there are limitless functions. You possibly can have a look at nearly each sector of life and each business. How do you prioritize issues? And in healthcare, the place do you see a very powerful challenges that stay that you’ll concentrate on?

Dean: Our group realized that there have been a variety of potential functions of machine studying on this planet, in just about each sector of human endeavor. A part of the reason being that between 2010 and 2013-2014, computer systems successfully developed the power to see. That is fairly transformative. If you happen to assume again to evolutionary instances when animals developed eyes, I believe that was an enormous deal. We reached that very same level in computing a number of years in the past. That has meant that there is a broad set of makes use of for the newfound capabilities of with the ability to see.

We wished to make software program packages and instruments accessible to those that have been making an attempt to use machine studying to completely different settings. That is why we developed a system known as TensorFlow and launched it as an open-source package deal for individuals to make use of in no matter means they wished. We had an Apache 2 free software program license. I am not a lawyer, however it means you are able to do no matter you need with it. That permits individuals to make use of it in plenty of completely different settings, and it has been downloaded 100 million instances, I believe, which is sort of outstanding for a reasonably obscure program or particular instrument.

Within the medical setting, persons are seeing the potential of machine studying, each for medical document knowledge and for medical imaging modalities. So we need to present instruments that make it straightforward to deploy these programs. Within the medical setting, there are a variety of difficult regulatory and privateness points that, for good purpose, impede the rollout of those programs in ways in which do not exist in much less regulated settings. But it surely’s additionally good to understand that there’s actual potential right here, and we have to get these items out in methods which might be optimistic for the world and the medical outcomes that we predict might be improved.

What’s This on My Pores and skin, Google?

Topol: One other physique of your work is on pores and skin lesions and issues. You developed Derm Help, which I perceive is being launched in Europe, having gotten regulatory approval (however not in the USA, which has completely different regulatory hurdles). This can assist individuals determine no matter pores and skin difficulty they’re involved about, by taking a photograph and getting an automatic preliminary prognosis. Are you able to inform us about Derm Help? As a result of that is some of the widespread the explanation why an individual goes to see a physician.

Dean: Dermatologists are one of many hardest specialists to get an appointment with as a result of they’re oversubscribed, and so generally a normal clinician will attempt to make a prognosis and maybe confer with a dermatologist. You do not want specialised tools essentially to seize the correct info to have the ability to look and assess it. We adopted an analogous path to the one we utilized in our diabetic retinopathy screening work, the place we first wished to collect knowledge on all kinds of various pores and skin tones and pores and skin circumstances, to see if we had one thing that was of high-enough high quality to place by a regulatory course of. We have now carried out a number of iterations to enhance the set of information that now we have to make it extra consultant, so we are able to roll it out to be used in assessing pores and skin circumstances. You probably have a rash in your arm, you possibly can take an image of it and it’ll type of present a number of options of what the rash is likely to be. Then you possibly can determine what to do provided that preliminary evaluation of what it is likely to be.

Verghese: One of many pushbacks on AI early on was that it may generally enlarge the inequities in society, and nobody would bear in mind that it was really doing that until you took a cautious look. Have we gotten higher? Have we constructed in additional safeguards? Is that this a difficulty that we are able to fear about rather less or is it a unbroken difficulty?

Dean: Any time you are fascinated with deploying a machine studying mannequin or an AI system on this planet, it is advisable be aware and conscious of many alternative features of that course of and the way they will introduce bias or equity points or ways in which the system interacts with individuals. You need interpretable fashions so that folks can perceive that if the mannequin is saying one thing, why is it saying that? That is all through the entire course of — trying on the dataset, on the sorts of algorithms you are doing and the way you consider it. You need to watch out to not simply consider it and report a single quantity. You need to have a look at completely different subsets of the inhabitants that is likely to be affected and at the way it performs on — for instance, with dermatology-related work — completely different ethnicities or pores and skin tones.

These are issues we take note of, and what we predict everybody deploying machine studying fashions — notably these making consequential selections about individuals’s lives — needs to be being attentive to. In 2018, we put out a set of ideas for considering, in a structured means, about each use of machine studying at Google, what we needs to be fascinated with. These fashions ought to keep away from creating unfair bias, they need to be interpretable by people, and so forth.

These are vital points and they’re going to proceed to require everybody’s consideration to make it possible for we do not do what you instructed: take knowledge from the world as it’s, after which automate some course of in order that we are able to take one thing that isn’t actually the world as it’s and enlarge or speed up the consequences of improper selections. We have to make it possible for we take the world as we wish it to be, and make it possible for our machine studying fashions are as near that aim as we are able to get.

Topol: Going again only for a second to Abraham’s level in regards to the EHRs. As you recognize, Jeff, there are burgeoning efforts for keyboard liberation, for taking the dialog and making an artificial observe based mostly on that, which could possibly be edited by each the affected person and machine that has been skilled by the physician or clinician’s earlier notes and edits. Do you see that as being imminent? Eliminating the keyboard could possibly be the favourite AI advance within the historical past of medication.

Dean: Our group has carried out a bit bit of labor utilizing an audio recording of a patient-doctor dialog, to create a draft of a medical observe {that a} clinician can simply edit a bit bit versus having to kind up a prolonged observe. Everyone knows that usually clinicians copy and paste the newest observe and do not actually edit it appropriately. That is partly as a result of it’s totally cumbersome and unwieldy to work together with a few of these programs, and speech and voice are a extra pure means of making notes. Creating summarized notes from conversations can also be maybe a very good assistive instrument that may scale back the scientific burden on clinicians but additionally maybe create higher-quality info within the medical document itself, which might be unbelievable.

Verghese: Do you are concerned, as you make open-source machine studying software program accessible to everybody, about individuals placing it to nefarious use? And conversely, do you are concerned about extra authorities regulation coming in to observe it to some extent the place it begins to snuff out the aptitude to do analysis and advance the trigger?

Dean: Like a variety of applied sciences, machine studying might be put to each optimistic and unfavorable makes use of on this planet. By and of itself, it’s sort of a impartial factor, however once you apply it to one thing, that is the place you make aware selections about what the system goes to do and what and who it’s going to impression — this individual, this populous, this group or the world. Pc imaginative and prescient has superb makes use of in medical diagnostics, however it additionally could possibly be used to create autonomous weapons. Society, governments, and different concerned constituencies must make selections about what sorts of makes use of we need to enable and what issues we do not need to enable.

When it comes to regulation, it is way more about specific makes use of which might be vital to take a look at. In lots of areas, together with medical gadgets or prescribed drugs, there are present regulatory frameworks for approval, so with some adaptation to what machine studying is ready to do in that area, you have already got a reasonably properly advanced regulatory framework that’s fairly serviceable. For different kinds of issues, corresponding to autonomous automobiles, there’s much less of a regulatory framework, and that is an space the place you may need governments to weigh in on what is acceptable and what’s not in a brand new regulatory framework that must be basically created from nothing.

Microchips in Minutes

Topol: Talking of issues which might be a bit bit faraway from medication, like autonomous automobiles, you and your colleagues at Google have been concerned with what I’d contemplate fairly outstanding breakthroughs in AI very not too long ago. Each have been printed in Nature, which is sort of an fascinating place to have these papers. I might prefer to get your touch upon these.

One is the work that was carried out for AI to design microchips in order that what would usually take months, if not years, to be created as a substitute took minutes or hours. And the opposite is the work on protein construction, such that any sequence of amino acids may predict in 3-D all of the folding for basically a overwhelming majority of proteins within the human proteome, after which ultimately maybe for all proteins. Clearly, that has implications for drug discovery and understanding the biology in man and elsewhere. The medical group might not be conscious of this but, although.

Dean: The chip design work is a bit additional afield from the medical group, however it’s fascinating. One of many phases in designing a pc chip is that you’ve the precise logic and transistors that you just need to placed on the floor of the chip, however you do not know the place to place them. So that you usually have human bodily design engineers sit there and play this difficult recreation of “the place ought to we put this factor” with a view to decrease fairly numerous completely different constraints: the size of wire between completely different items, and the quantity of energy and the realm that the chip will eat. You need to make it as small as doable and eat as little energy as doable and decrease wire size.

We used a reinforcement studying algorithm, the place you’re taking a bunch of actions after which on the finish you get a reward. That is much like how our DeepMind colleagues mastered the sport of Go, the place you possibly can basically play a bunch of strikes and go. On the finish you get a reward sign: Did you win or lose? Right here you may get a reward sign that is a little more multidimensional, just like the precise measurement of the chip and the ability that it consumes and so forth. However you possibly can basically play the sport of inserting the elements of the chip and see how properly you probably did. Then you possibly can strive once more, given the suggestions you bought the earlier time. It seems that it is a a lot higher-dimensional downside area than Go, however you possibly can be taught to do that efficiently in an automatic means as a substitute of it taking a group of 10 individuals many weeks or months to undergo this course of. That is sort of thrilling. And it leads to smaller chips created extra rapidly and rolled out into the world, which is good.

When it comes to the protein folding, that is nice work from our colleagues at DeepMind. The implications of which might be nonetheless not recognized as a result of the work was simply launched up to now few weeks. However it’s actually a testomony to the truth that we are able to now sort out a few of these primary issues in science with machine studying and develop new capabilities that then allow us to do issues we could not do earlier than and can have an effect on plenty of downstream fields and makes use of as properly.

The potential of machine studying in medication is in its infancy, and the power for us to assist in giving everybody on this planet higher medical care, higher details about their very own medical situation, and provides clinicians higher recommendation in regards to the issues they need to be fascinated with or selections they make is outstanding, however nonetheless largely not rolled out on this planet. How will we transfer from the place we’re to the place we could possibly be?

Topol: Nicely, full circle out of your work on pandemic preparedness 40 years in the past to your work now. I can be following you intently. Thanks a lot for becoming a member of us.

Eric J. Topol, MD, is among the prime 10 most cited researchers in medication and ceaselessly writes about know-how in healthcare, together with in his newest e-book, Deep Medication: How Synthetic Intelligence Can Make Healthcare Human Once more.

Abraham Verghese, MD, is a critically acclaimed best-selling writer and a doctor with a world fame for his concentrate on therapeutic in an period when know-how usually overwhelms the human aspect of medication.

Jeff Dean, PhD, has lived on 4 completely different continents (North America, Africa, Asia, and Europe). He says one among his private targets is to play soccer and basketball on all seven however acknowledges that Antarctica is likely to be robust.

Comply with Medscape on Fb, Twitter, Instagram, and YouTube


Supply hyperlink