AI for Earlier and Safer
Medicine
Yu-Chuan (Jack) Li, M.D., Ph.D., FACMI
Professor in Biomedical Informatics,
Dean, College of Medical Science and Technology
Taipei Medical University
A bit about myself
• Professor in Biomedical Informatics
• Board-certified Dermatologist
• Elected Fellow, ACMI (American College of Medical
Informatics) and IAHIS (International Academy of Health
Information Science)
• Fellow, ACHI (Australian College of Health Informatics)
• Editor-in-Chief, Computer Methods and Programs in
Biomedicine (IF 2.7)
• Editor-in-Chief, International Journal for Quality in
Healthcare (IF 2.6)
http://Jackli.cc
Computer Methods and Programs in
Biomedicine
International Journal for Quality in
Health Care
Editor-in-Chief
ISQua / OUPElsevier
2000 submissions, 360 paper published / year
3
Defining AI
Artificial intelligence (AI) is intelligence exhibited by machines.
…
Colloquially, the term "artificial intelligence" is applied
when a machine mimics "cognitive" functions that
humans associate with other human minds, such as
"learning" and "problem solving".
…
"We don't need Artificial Intelligence if we don't have Natural Stupidity!"
- Professor Allan T. Pryor
Author of the book -
“Health Care
Future with AI”
7
Evolution of AI
• 1960 Age of Reasoning
• Logic-based
• heuristic search
• 1990 Age of Representation
• Rule-based
• Knowledge engineering
• Expert system
• 2015~ Age of Machine Learning
• Big Data-driven
• Autonomous learning
• 2045 Age of Superintelligence?
8
Why AI in HC
• Taiwan has a strong ICT industry/academia
• Taiwan has one of the most“high
performance”healthcare system in the world
• Very high outpatient visit – 15 visits/pers/yr
• Diagnoses/Drugs coded by physicians, NOT
coders
• Accurate e-prescription – $$$ by NHI x 200
• 23 million people for 23 years EHR
• Very standard coding and data schema
Dimensions of AI in Health Care
• Stakeholders
• Locality
• Urgency
• Business/sustainability model
• AI Technology (supervised, unsupervised learning…etc.)
• Data source/quality
• ELSI (Ethical, Legal and Social Implications)
Three-Dimension Model
• 4 Stakeholders
– Patient/Consumer, Clinician, Administrator,
Payer/Insurer
• 4 Locality
– GP office, Hospital, Home, Other institution
• 3 Urgency types
– Preventive care, Acute care, Long-term care
10
Preventive
Care
AcuteCare
Long-term
Care
Preventive
Care
AcuteCare
Long-term
Care
Preventive
Care
AcuteCare
Long-term
Care
Preventive
Care
AcuteCare
Long-term
Care
GP Office 22 6 5 23 7 6 10 3 4 17 6 7
Hospital 15 28 5 22 29 9 17 20 7 15 15 4
Home 30 7 17 17 3 9 6 2 5 14 2 8
Other Institutions 7 4 4 4 1 3 2 2 2 3 0 5
Locality
Stakeholders
Urgency
BeforeDiscussion
Patient /Consumer Physician Administrator Payer /Insurer
Three-Dimension Table (4x4x3)
Key Issues in Current Health Care
•Medical Errors 醫療錯誤
•Poor/Inconsistent Quality 品質不佳
•One-size-fits-all Approach 以偏概全
•Prevention ignored 輕忽預防
13
Prevention is Hard
• No visible target
• Repetitive & Slow
• No pain
• People don’t understand probability
• Science can’t produce useful predictions
 Low market value
14
Predictions are Useless
Unless they are:
•Accurate
•Timely
•Individualized
•Actionable
The Levels of Prevention
AI-based Earlier
Detection
AI-based Earlier
Intervention
AI-based Earlier
Risk Reduction
MoleMe Project
• To determine whether the risk of a
mole is high enough to justify a visit
to the doctors
• 3,000 patients; 4 dermatologists
• Images + 5 simple variables
• Machine learning, supervised
• ResNet -> AUROC of 90%
Earlier detection of moles
MoleMe
MoleMe
AESOP Project
• AI-Enhanced Safety of Prescription
• 700M prescriptions; 1.3B Diagnoses;
2.5B medications
• 16,000 variables; 400 doctors
• Machine learning, unsupervised
• 75% to 90% agreement by experts
reviewers
733.4 Millions
Prescriptions
80 Million Dx-Med and 2.25 Million Med-Med Associations Explored
Medication codes
Mapped to 1,500 unique WHO
codes
Diagnoses ICD codes
20,000 unique ICD codes
Machine Learning
Learn from Doctors‘ Behavior
1.34B 2.53B
Prevent Medication Errors at the Earliest
20
Preliminary Results
A Medical Center in Taiwan
• Patients:72,378
Reminders:2140 (3%)
Agreed:1038 (48%)
• High risk medications
• Patients :17,793人
Reminders :114
Agreed :62 (54%)
A Healthcare System in the US
• Patients : 31,728
Reminders : 2,723
(8.6%)
*Estimated
Temporal Cancer Prediction
• To determine the occurrence of
cancer in the next 12 months based
on the previous 36 months of PHR
• 80K liver cancer patients; 320K
control; 2,700 variables
• Machine learning, supervised
• Turn Time-matrix into images
• AUROC of 94%
23
“Time Matrix” CNN transformation
本團隊榮獲肝病防治基金會
研究獎助金 (2018.08.11)
94%
Accuracy
(AUROC)
Accurate Prediction of Liver Cancer
24
Patient
Profile
Diagnosis
/Problem
ProceduresMedication
Lab/Exam
7
11
Age, sex, allergy, weight,
height, blood type, body
temperature, …etc.
YC (Jack) Li et. al., 2004
Current and/or chronic
dz, DM, H/T,
Pregnancy…etc.
Surgery, transfusion,
endoscopy,
angiogram, PTCA,
rehabilitation…etc.
Propanolol vs
theophylline,
Cipro vs aminophylline,
Acetaminophen vs
Phenytoin…etc.
CBC, D/C, Chem-
20, hCG, PT,
APTT, INR…etc.
e.g. Coumadin vs
INR
e.g. Wafarin vs
angiogram
e.g. Penicillin vs
PCN allergy
e.g. Retinoids vs
pregnancy
Data Interaction Model for Adverse Event detection
2x
2x
2x
2x
1x
Input Variables for AIHC with the Temporal Dimension
Patient
Profile
Diagnosis
/Problem
ProceduresMedication
Lab/Exam
12
9
7
8
5
11
10
4
3
2
Birth
YC (Jack) Li et. al., 2016
Phenotype
(Environmental)
Output Variables of AIHC
Death
YC (Jack) Li et. al., 2016
Treatment Rehabilitation
Prognosis
Management
Diagnosis
Prediction Early
Detection
Suggestion/Recommendation
Conclusion
• AI and Healthcare should go hand-in-
hand
• AI is opening a whole new page of
preventive & earlier medicine
• AI has to change the future of medicine
(or we may not have one)
•  because we deserve it!
Thank you for your attention

[台灣人工智慧學校] 開創台灣產業智慧轉型的新契機

  • 1.
    AI for Earlierand Safer Medicine Yu-Chuan (Jack) Li, M.D., Ph.D., FACMI Professor in Biomedical Informatics, Dean, College of Medical Science and Technology Taipei Medical University
  • 2.
    A bit aboutmyself • Professor in Biomedical Informatics • Board-certified Dermatologist • Elected Fellow, ACMI (American College of Medical Informatics) and IAHIS (International Academy of Health Information Science) • Fellow, ACHI (Australian College of Health Informatics) • Editor-in-Chief, Computer Methods and Programs in Biomedicine (IF 2.7) • Editor-in-Chief, International Journal for Quality in Healthcare (IF 2.6) http://Jackli.cc
  • 3.
    Computer Methods andPrograms in Biomedicine International Journal for Quality in Health Care Editor-in-Chief ISQua / OUPElsevier 2000 submissions, 360 paper published / year 3
  • 4.
    Defining AI Artificial intelligence(AI) is intelligence exhibited by machines. … Colloquially, the term "artificial intelligence" is applied when a machine mimics "cognitive" functions that humans associate with other human minds, such as "learning" and "problem solving". … "We don't need Artificial Intelligence if we don't have Natural Stupidity!" - Professor Allan T. Pryor
  • 5.
    Author of thebook - “Health Care Future with AI”
  • 7.
    7 Evolution of AI •1960 Age of Reasoning • Logic-based • heuristic search • 1990 Age of Representation • Rule-based • Knowledge engineering • Expert system • 2015~ Age of Machine Learning • Big Data-driven • Autonomous learning • 2045 Age of Superintelligence?
  • 8.
    8 Why AI inHC • Taiwan has a strong ICT industry/academia • Taiwan has one of the most“high performance”healthcare system in the world • Very high outpatient visit – 15 visits/pers/yr • Diagnoses/Drugs coded by physicians, NOT coders • Accurate e-prescription – $$$ by NHI x 200 • 23 million people for 23 years EHR • Very standard coding and data schema
  • 9.
    Dimensions of AIin Health Care • Stakeholders • Locality • Urgency • Business/sustainability model • AI Technology (supervised, unsupervised learning…etc.) • Data source/quality • ELSI (Ethical, Legal and Social Implications)
  • 10.
    Three-Dimension Model • 4Stakeholders – Patient/Consumer, Clinician, Administrator, Payer/Insurer • 4 Locality – GP office, Hospital, Home, Other institution • 3 Urgency types – Preventive care, Acute care, Long-term care 10
  • 11.
    Preventive Care AcuteCare Long-term Care Preventive Care AcuteCare Long-term Care Preventive Care AcuteCare Long-term Care Preventive Care AcuteCare Long-term Care GP Office 226 5 23 7 6 10 3 4 17 6 7 Hospital 15 28 5 22 29 9 17 20 7 15 15 4 Home 30 7 17 17 3 9 6 2 5 14 2 8 Other Institutions 7 4 4 4 1 3 2 2 2 3 0 5 Locality Stakeholders Urgency BeforeDiscussion Patient /Consumer Physician Administrator Payer /Insurer Three-Dimension Table (4x4x3)
  • 12.
    Key Issues inCurrent Health Care •Medical Errors 醫療錯誤 •Poor/Inconsistent Quality 品質不佳 •One-size-fits-all Approach 以偏概全 •Prevention ignored 輕忽預防
  • 13.
    13 Prevention is Hard •No visible target • Repetitive & Slow • No pain • People don’t understand probability • Science can’t produce useful predictions  Low market value
  • 14.
    14 Predictions are Useless Unlessthey are: •Accurate •Timely •Individualized •Actionable
  • 15.
    The Levels ofPrevention AI-based Earlier Detection AI-based Earlier Intervention AI-based Earlier Risk Reduction
  • 16.
    MoleMe Project • Todetermine whether the risk of a mole is high enough to justify a visit to the doctors • 3,000 patients; 4 dermatologists • Images + 5 simple variables • Machine learning, supervised • ResNet -> AUROC of 90%
  • 17.
    Earlier detection ofmoles MoleMe MoleMe
  • 18.
    AESOP Project • AI-EnhancedSafety of Prescription • 700M prescriptions; 1.3B Diagnoses; 2.5B medications • 16,000 variables; 400 doctors • Machine learning, unsupervised • 75% to 90% agreement by experts reviewers
  • 19.
    733.4 Millions Prescriptions 80 MillionDx-Med and 2.25 Million Med-Med Associations Explored Medication codes Mapped to 1,500 unique WHO codes Diagnoses ICD codes 20,000 unique ICD codes Machine Learning Learn from Doctors‘ Behavior 1.34B 2.53B Prevent Medication Errors at the Earliest
  • 20.
  • 21.
    Preliminary Results A MedicalCenter in Taiwan • Patients:72,378 Reminders:2140 (3%) Agreed:1038 (48%) • High risk medications • Patients :17,793人 Reminders :114 Agreed :62 (54%) A Healthcare System in the US • Patients : 31,728 Reminders : 2,723 (8.6%) *Estimated
  • 22.
    Temporal Cancer Prediction •To determine the occurrence of cancer in the next 12 months based on the previous 36 months of PHR • 80K liver cancer patients; 320K control; 2,700 variables • Machine learning, supervised • Turn Time-matrix into images • AUROC of 94%
  • 23.
  • 24.
  • 25.
    Patient Profile Diagnosis /Problem ProceduresMedication Lab/Exam 7 11 Age, sex, allergy,weight, height, blood type, body temperature, …etc. YC (Jack) Li et. al., 2004 Current and/or chronic dz, DM, H/T, Pregnancy…etc. Surgery, transfusion, endoscopy, angiogram, PTCA, rehabilitation…etc. Propanolol vs theophylline, Cipro vs aminophylline, Acetaminophen vs Phenytoin…etc. CBC, D/C, Chem- 20, hCG, PT, APTT, INR…etc. e.g. Coumadin vs INR e.g. Wafarin vs angiogram e.g. Penicillin vs PCN allergy e.g. Retinoids vs pregnancy Data Interaction Model for Adverse Event detection 2x 2x 2x 2x 1x
  • 26.
    Input Variables forAIHC with the Temporal Dimension Patient Profile Diagnosis /Problem ProceduresMedication Lab/Exam 12 9 7 8 5 11 10 4 3 2 Birth YC (Jack) Li et. al., 2016 Phenotype (Environmental)
  • 27.
    Output Variables ofAIHC Death YC (Jack) Li et. al., 2016 Treatment Rehabilitation Prognosis Management Diagnosis Prediction Early Detection Suggestion/Recommendation
  • 28.
    Conclusion • AI andHealthcare should go hand-in- hand • AI is opening a whole new page of preventive & earlier medicine • AI has to change the future of medicine (or we may not have one) •  because we deserve it!
  • 29.
    Thank you foryour attention