智慧醫(yī)療:臺灣的機(jī)會(huì)和挑戰(zhàn)李建璋臺大智慧醫(yī)療中心副主任臺大急診部主治醫(yī)師史丹佛大學(xué)

Biodesign教學(xué)研究員哈佛大學(xué)流行病學(xué)博士Automatic

interpretationof

radiological

imagesMonitoringdisease

risksDevelop

newdrugsPredictCOVIDtrendItseemedthatWatsoncould

understandthe

meaning

of

language,

ratherthanjust

patterns

ofwordsWatson’s

Medical

AI

Moon

Shot

ProjectWatson’selectronic

brainDr.

Watson?

Goal:?

Reducediagnosis

errors?

Optimizetreatments?

Alleviatedoctorshortages?

How?

Collectionofsymptomsandcame

upwith

a

list

of

diagnoses

and

treatment?

EachannotatedwithWatson’sconfidence

leveland

links

to

literature?

Confidence?

Memory

banks

held

knowledgeofevery

rarediseaseand

newesttrials?

Processors

notsusceptibletocognitive

bias“Thisproductis

a

piece

ofsh–

.”“Multipleexamplesofunsafeandincorrecttreatmentrecommendations,”It’s

all

about

“DATA”?That

narrativetextaccountsforabout

80

percent

ofa

typical

patient’s

record—and

it’sastewofjargon,shorthand,and

subjective

statements?

Evenfor

MDAnderson,theyonly

have

hundredsof

beds,

could

not

coverthevastvarietyof

human

diseases?Trainedthesoftwareona

small

number

of

hypothetical

cancer

patients,

ratherthan

real

patientdata

leadstodisasterDirty

Medical

Data?

Heterogeneity?

Differentdrug

preferencefordifferent

hospital/physicians?

Differenttreatment

recommendationoverdifferent

periods?

Different

insightfordifferent

levelsofdoctors?

Differentselectionoftesting

orders?

ProblemsspecificforTaiwan?

Usea

lotof

atypical

short-hand?

Physician’s

note,

not

English,

notChinese,but“TaiwaneseChinglish”?

Under-recordingfordoctor’s

note,overrepresentingof

nursecharting

notesImage

applications

account

for>70%Of29

FDA

approved

SamD,臺灣的機(jī)會(huì)和挑戰(zhàn)Let’s

builda

Medical

ImageNetofTaiwanAI

model

is

ethnic/regionsensitiveAI

model

based

on

Westerncountries

may

not

be

able

toapplytoAsian

countries健保資料庫

電子病歷資料庫

人體生物資料庫臺灣三座金山挑戰(zhàn)-

Privacy

IssueOpportunityChallenge資料26Medical

image

needs

lotsofeffortsforannotationCross-disciplinaryTrainingOpportunityChallengeCross-disciplinary

professionals人才AnnotationIST30臺大醫(yī)院建置二套

NVIDIA

DGX-A100目前運(yùn)算效能最快的AI超級電腦醫(yī)學(xué)影像辨識、

基因體分析需要龐大的電腦儲存與算力●

醫(yī)療影像:提升2倍效能●

自然語言:

提升5倍效能●GPU分成七個(gè)執(zhí)行體，提升7倍效能●支援基因高速運(yùn)算，提升2.25倍效能●人類全基因定序僅需20分鐘OpportunityChallengeUnderuseofcloudcomputing算力Strong

IT

industry33AI

paper

hits

record

highNewlyfundedAIstartup

record

lowAI模型需要第三期臨床試驗(yàn)人才轉(zhuǎn)譯資料算力臺大醫(yī)院成立智慧醫(yī)療中心臨床試驗(yàn)場域/推動(dòng)醫(yī)療醫(yī)材軟體試用啟用2臺AI

超級電腦/建置運(yùn)算專區(qū)60個(gè)臺灣大學(xué)人工智慧團(tuán)隊(duì)/華人最優(yōu)質(zhì)的醫(yī)療大數(shù)據(jù)39資訊安全不影響Portal系統(tǒng)效能確效機(jī)制與定期紀(jì)錄病患隱私資料不外洩落地辦法

40臺大醫(yī)院醫(yī)療器材軟體試用申請流程Oppor