Artificial Intelligence In Healthcare: Separating The .

Artificial Intelligence inHealthcare: Separating theReality from the HypeRobert M. Wachter, MDProfessor and Chair, Department of MedicineUniversity of California, San FranciscoAuthor, The Digital Doctor: Hope, Hype & Harm at theDawn of Medicine’s Computer Age@bob Wachter / robert.wachter@ucsf.edu

ccuityMedicalBoard ofDirectorsDoctorsCompanyBoard ofDirectorsTeledocAdvisory BoardAmino.comAdvisory BoardPatientSafeSolutionsAdvisory BoardEarlySenseAdvisory t PatientsInvestorHonoraria forNuance, GE,Health Catalyst, speakingAvaCare

Why I Decided to Explore Health IT

Road Mapu Whatthe EHR years taught us abouthealthcare and digitalu Whatyou (really) need to know about AIu Somespecific areas in which AI is likely tomake a differenceu Speedbumps and unanticipatedconsequences to anticipateu Bottomline

EHRs in U.S. Hospitals, 2008-17 30billion

(Re) Enter the Digital Giants .

7-year-old Girl’s Recollection of her Visit to the DoctorToll E. The cost of technology. JAMA 2012

Advertisement For Arizona ER JobArizona General Hospital will be coming to The GrandCanyon State later this year!! Located in a suburb ofPhoenix, Arizona General Hospital is a 40,000 squarefoot boutique general hospital.Services offered include: Emergency Room Radiology Suite inc. CT, X-Ray, and Fluoroscopy Two State-Of-The-Art Operating Rooms Outpatient Surgery 16 Inpatient Rooms NO ELECTRONIC MEDICAL RECORD

From the EHR era of Health IT

The Four Stages of Health IT1.Digitizing the record2.Connecting the partsa. PCPs to Hospitals, Hospitals toHospitals, etc.b. Patient-facing systems to enterprisesystems, and to one another3.Gleaning meaningful insightsfrom the data4.Converting these insights intoaction that improves value

Health IT: The Mother ofall Adaptive Problems“ problems that require people themselves to change. In adaptiveproblems, the people are the problem and the people are thesolution. And leadership then is about mobilizing and engaging thepeople with the problem rather than trying to anesthetize them sothat you can just go off and solve it on your own.”– Ronald Heifetz, Kennedy School of GovernmentBottom line re: Health IT: Much harder than it looks .

AI Without that Nasty Math

Examples of AI Applications inHealthcareFrom Topol E. High performance medicine: the convergenceof human and artificial intelligence Nature Medicine 2019

Early Efforts at ComputerizedDiagnosis (circa 1970s-80s)

Healthcare questions that will be answeredor tasks addressed via AI & Big DatauDiagnose a conditionu TeachuPredict risk for developing a disease or outcomeu E.g.ucomputer to identify cases of lung cancer from path slidesPredict sepsis or sepsis mortality based on EHR dataPredict treatment responseu E.g.Predicting response to treatment of hypertension or cancerbased on genetic data (“precision medicine”)uImprove practiceu Predictu Helpclinic demand or no-showsMD complete EHR note based on prior examples

It’s very unusual to see pt w/diagnosis of GI bleeding still onanticoagulant. You’re eithervery innovative or very wrongPatients like this areusually afebrile by now.Rethink the diagnosis!

Let’s assume AI gets everythingright (diagnosis, prediction,outliers, EHR note) Then what?Sundar Pinchai,CEO, Google

The Traditional Concerns About AIuBlack boxuuGarbage In-Garbage Out (GIGO)uuBMJ study found that 4am CBC (even if normal) was top predictor of mortalityPromoting healthcare disparitiesuuMight appear that African-American patients with fractures require less pain meds, andbuild that into algorithm; melanoma detection less accurate in darker skin peopleWacky and confounded resultsuuE.g., much EHR data fed into AI generators is from billing data of questionable accuracyBiasuuWill clinicians follow predictions/recommendations whose derivation they don’tunderstand or that have no biologic plausibility?If much of AI is served up via iPhone, there will be haves and have-notsPrivacy and security issuesuRequires huge data to be amalgamated & often combined; can algorithms be hacked?

Riffs on 4 Other Areas Likely to BeAssociated with SpeedbumpsuHow is this actually going to work?uIntegration and interoperabilityuThe need for a new layer that connects patients to the systemuWill better predictions actually change behavioruThe AI version of alert fatigueuDeskillinguHow good is good enough (AKA, the Tesla Problem)uThe politics of AI

Health IT Needs Its Golden Spike

Ida Sim, NEJM 2019

Patient 42 has irregularHR and is SOB. Let’s doa televisit ASAPPatient 13’s weightis up and O2 sat isworse. I’ll lock thesalt shaker and thefridgePatient 112’s sugar is highagain: the algorithm bumpedthe insulin but let’s get thecoach involvedThe Care Traffic Controllers

One More Problem: Alert FatigueuOne month in UCSF ICUs (70 beds)u 2,558,760alertsu Audibleu Whatualert every 7 minuteswould get a nurse scared?Experience with other alerts is similarly underwhelmingDrew B. Plos One 2014

The Digital Squeeze on Questioning ofExpertise

Deskilling in theFace of Automation

Referring to the Lion Air copilot .

The Phenomenon of “Deskilling”“How do you measure the expense of anerosion of effort and engagement, or awaning of agency and autonomy, or a subtledeterioration of skill? You can’t. Those arethe kinds of shadowy, intangible things thatwe rarely appreciate until after they’re gone.”-- Nicholas Carr

The political-economics of healthcare AIPressure for value, competitive market will push care down to tech-enabledsolutions: Specialists Generalists Non-MD Clinicians Coaches Families Ptsu How much people will be willing to pay for higher priced, more personalentities will be determined empirically (See: concierge medicine, travelagents, tax preparers)uuThis includes the value that people place on ”human touch” and empathyMay be some tasks that incumbents are happy to lose (menial, rote, poorlypaid), esp. if there is more interesting & equally (or more) lucrativereplacement worku When there isn’t, expect a dogfight from incumbentsuuuuuMDs are a better guild than taxi-drivers (nurses are too, and they’re unionized)Much depends on who is at risk for costs (& quality?)The liability landscape will also have a sayIf machine is ultimately better and cheaper (or cheaper and notdemonstrably worse), it will win as standard of practiceuWith ability to buy up to “business class”

Research to date often supports the Centauridea. But when you hear the warm, fuzzy answer(“man machine is better than either alone”) .

“In theory there is no differencebetween theory and practice. Inpractice there is.”- Yogi Berra (maybe)

Additional Slides If Needed

AI reverses traditional healthcarepredicting/reasoninguuThe traditional deductive process (predictors - test)uBased on biological plausibility and prior research, hypothesize (for example)predictors for readmission, sepsis, stroke, death uTest, through a variety of clinical research methods, whether these predictorsare in fact associated with the outcome we care aboutThe new inductive process (test - predictors)uFeed the machine data on thousands/millions of patients with all sorts ofpossible predictor datauComputer looks at who had the outcome (eg, sepsis, lung cancer,readmission), works backwards to find the variables that predicted theoutcome, and then creates a model that optimizes the predictionsThanks to Ziad Obermeyer,UC Berkeley

The New Digital TriadRemoteSensing81% of NorthAmericans own asmartphoneConsumerfacingPersonalTechSmartphones can sensemotion & position,record vital signs & falls,record surveyresponses ArtificialIntelligenceSim I. Mobile devices andhealth. NEJM 2019

Topol, Nature Medicine 2019

“The Challenge That WillDominate Your Career ”

Two Transformational TrendsPressure todeliver highvalue careThe digitizationof the U.S.healthcaresystemThe Dominant Issue Prediction: The DominantTodayIssue in 2025

Reality from the Hype Robert M. Wachter, MD Professor and Chair, Department of Medicine University of California, San Francisco Author, The Digital Doctor: Hope, Hype & Harm at the Dawn of Medicine’s Computer Age @bob_Wachter / robert.wachter@ucsf.edu