I. Evolutionary Knowledge Machine (EKM) II. AI Doctor Platform (IMP) - - PowerPoint PPT Presentation
AI Doctor based on Department of Computer Science & Engineering, Evolutionary Knowledge Machine (EKM) Kyung Hee University KOREA February 19 th , 2019 Prof. Sungyoung Lee http://uclab.khu.ac.kr 2 Contents / I. Evolutionary Knowledge
Department of Computer Science & Engineering, Kyung Hee University KOREA
AI Doctor based on
Evolutionary Knowledge Machine (EKM)
February 19th, 2019
http://uclab.khu.ac.kr
/
2
Contents
/
4
What is our Research Goal?
/
5
Challenges of AI Doctor
Evolutionary Knowledge Big Knowledge
/
6
What Problems of Big Knowledge?
Knowledge Representation
/
7
How to obtain Qualified Medical Knowledge?
Data Driven Knowledge Acquisition Expert Driven Knowledge Acquisition
/
8
source: http://www.legaltechnology.com/latest-news/artificial-intelligence-in-law-the-state-of-play-in-2015/
Research Areas of AI
/
9 White Box and Black Box Model (Supervised Learning for classification)
White Box Models Black Box Models Machine learning algorithms which produce decision models in such a form that are interpretable for the domain experts Description Decision Tree Machine learning algorithms which produce decision models in the form of a set of mathematical functions that are non- interpretable for the domain experts
Each attribute is a node, important attributes are placed higher in the decision tree
Decision Rules
A model is determined piece-wise by a set of 'rules' that each cover part
Root node Rule 1 Rule 2 Rule 3 Rule 4 Ordered Rules List
Neural Networks
A set of Neurons are stacked in a multi-layer form to generate a non- linear mapping from input to
Graphical Models
A probabilistic model for reflecting dependency between a set of random variables
Input Layer H(1) H(2) Output
A C B D
Random Variables and their probabilistic graph
Description
/
10 Abstract of EKM
Structured Data
Knowledge Extraction
Machine Learning
Big Data
Verification? Inference Engine Adaptive Recommendations Alerts Reminders
1 2 3 4
Intervention of Domain Knowledge Expert
Engineering Support Tool
Evolutionary Knowledge base
/
11
Hybrid Knowledge Acquisition CNN Knowledge Authoring Tool
RDR Editor DT Rule RDR Rule DT Editor Rule Black Box White Box Feature Selector (UFS, FCBF, PSO-FS)
Image Data EMR Data (Structured) Expert Heuristics Legacy Knowledge Base
Case Generation (RDR Transformation) Knowledge Consolidation & Inferencing
RDR Rules
Research Concept & Scope
Classification Label Feature Vector Data Driven
Output Input
Expert Driven concatenation J48, Random Forest, DT
White Box
*RDR: Ripple Down Rules *DT: Decision Tree
Hybrid knowledge Acquisition Model
/
12 Features of EKM Knowledge Base
/
14
AI Doctor (Intelligent Medical Expert System)
Evolutionary Knowledge Base -> Big Knowledge Management
AI Doctor Platform
Intelligent Medical Services (Silo)
/
15 AI-Clinical Decision Support System (CDSS)
/
16
be, to be able to be present in most commercial reliability medical expert system knowledge base does not have these characteristics
Integrity Adaptability Freshness
Reliability
There should be no Rule (knowledge generation) is flawed — Rule should reflect the complete medical knowledge — the Rule is an actual medical environment there should be no shortage
Rule must be customized according to hospital environment - It should be customizable according to the situation of available resources (medical equipment, inspection equipment, etc.) by medical environment
Easy to update new knowledge
derived, the rule should be updated on these matters.
Must have sufficient credibility
papers, clinical trial data, or EMR inference data. Rule DB built without sufficient medical knowledge Nonsense
Depending on the equipment the hospital has, the test method / treatment method /
An update to a new rule Easy to handle
It should be based on papers, clinical trial data from pharmaceutical companies, and EMR inference data
Medical staff directly Knowledge be able to create and maintain Evidence Based
Knowledge base requirements Detail CASE Final requirements
Requirements of AI-CDSS Knowledge Base
/
17 Existing CDSS and their Features
Features Systems
Big Data Support Standard Compliant Storage HIS Integration Big Data Analytics Hybrid Knowledge Acquisition Rule-based Recommendation Adaptive UI Support McKesson: InterQual Clinical Decision Support
X X O O X O X
AllScripts:Knowledge- Based Medication Administration (KBMA)
X X O X X O X
Visual DX
X X O X O O X
Kinesia 360
O
X
O O
X X X
Reed Group: MDGuideline Intelligent DSS
O
X
O O
X X
O
Medaware System Personalized CDSS
O O O O
X X X
IBM Watson
O
x
O O
X
O O
White Box
Black Box
Gray Box
Required Features
Tool Support for Knowledge Acquisition Incremental knowledge Maintenance Dialogue Support Personalization Support
https://www.changehealthcare.com/solutions/interqual https://www.allscripts.com/news-insights/blog/blog/2016/01/safer-medications-with-closed-loop-delivery https://www.visualdx.com https://glneurotech.com/kinesia/products/kinesia-360/ https://www.mdguidelines.com/ https://www.ibm.com/watson/ /
18
Intelligent Medical Platform
Heterogeneous Input Data Lifelog Knowledge base
Blood pressure device Unstructured Text
Thyroid Cancer Silo Knowledge Engineering Tool UI/UX Authoring Tool
Smart Watch Sleep Monitoring Device Glucose Meter Medical PACS Patient Profile EMR/EHR Patient Healthlog
UX Expert Physician Patient
Analytics Tool
Physician Physician
Cardiovascular Silo Head & Neck Cancer Silo Diabetes Silo Epilepsy Silo Lung Cancer Silo Public Health Silos Evidence Support Tool
8 Medical Services Silo
Big data Storage Intelligent Medical Services
AI Doctor (Intelligent Medical Platform) Environments
ENT (Ear, Nose, Throat) Silo
/
19 AI Doctor Platform (IMP)
/
20 SaaS implementation for IMP
/
21 Uniqueness: Adaptive Services
Adaptive Recommendation Adaptive Education Adaptive Q &A
Evolutionary Knowledge Base Adaptive (Personalized) Services
R E
Q&A
/
22
01 02 03 04
UI/UX Authoring Tool Knowledge Authoring Tool Evidence Support Tool Data Analytics Tool
validation and verification
documents
UX Expert Physician Physician Patient Physician Easy to commercialize by providing development environment
Uniqueness: Platform + Engineering Tool
/
23 IBM Watson Oncology vs. EKM
Characteristics/Features IBM Watson Oncology Evolutionary Knowledge Machine Knowledge Acquisition Process
Based on pre-curated annotations Incremental Learning Model
Knowledge Modeling Approach
Primarily focused on medical Images and textual data (clinical notes, doctor notes, patient case report) Knowledge Generation from multi-modal data sources (EMR, clinical notes, medical images, expert heuristics)
Medical Expert Assistance in Knowledge Creation
Limited support for direct knowledge incorporation Expert-friendly knowledge authoring environment for incorporating expert heuristics
Knowledge Maintenance Capabilities
Complex and time consuming Seamless knowledge maintenance (RDR)
Evidence backed Treatment
Supported Supported
Knowledge Shareability
All learned knowledge is tightly coupled with the system Supports knowledge sharaeability by converting knowledgebase into medical logic module
https://www.ibm.com/us-en/marketplace/ibm-watson-for-oncology
/
25
Published Research Expert Heuristics Clinical Guideline EMR/EHR Silos Mind Map
Silo Construction Process
Diagnosis Treatment Follow-up
A silo provides intelligent medical services for specific diseases
IMP Service (Silo)
/
지
26
Rules
4 3
Knowledge Engineer Mind Map Decision Tree
1 2 5
Knowledge Base Physician Recommendation Production Rules
Knowledge Creation and Diagnosis Recommendation
/
27
Enterprise Architect
Medical Experts Knowledge Engineer
Mind Map Decision Tree
Case #1: Cardio- Knowledge Acquisition (Decision Tree)
/
28
Intelligent Knowledge Authoring Tool of IMP
(I-KAT)
Medical Experts
Decision Tree Production Rules Total Rules: 1,309 Total Patient Data : 300 Initial Accuracy : 90%
Case #1: Cardio - Knowledge Acquisition (Production Rule)
/
29 Case #1: Implementation of Cardio - Dashboard
Dashboard: Shows all the patient data from EMR and EHR systems
Add New Patient Update Existing Patient Delete Existing Patient Search Patient Patient abstract information
/
30 Case #1: Implementation of Cardio - Patient Detail
Patient Detail Screen: Shows all detail of patient to add new patient of update existing patient
Patient Information And Cardio Information Patient clinical History Patient Symptoms Patient Physical exam information
/
31
CDSS Intervention: Shows the recommendation of a patient based on patient profile and symptoms
The decision comes from knowledge base
Final Decision Knowledge Rule triggered
Case #1: Implementation of Cardio CDSS Intervention
/
32 Case #1: Accomplishment of Cardio Silo
You have completed the confirmation process for Abstract Control Number: 13269:
Artificial Intelligence (Ai) Clinical Decision-Supporting System (CDSS) for Diagnosis of Heart Failure: Concordance With Expert Decision, accepted for
presentation at Scientific Sessions 2018. Dong-Ju Choi M, Jin Joo Park M, Youngjin Cho M, Seoul Natl Univ, Seongnam, Korea, Republic of; Sungyoung Lee M, Taqdir Ali M, Kyung Hee Univ, Suweon, Korea, Republic of Phone: +82 31 787 7007
American Heart Association (AHA) Scientific Sessions - 2018 10 - 12 Nov, 2018, McCormick Place, Chicago
/
33 Case #2: Thyroid Cancer (Treatment)
갑상선 진단 지식베이스 갑상선 진단 시스템 인터페이스 테스트 케이스 테스트 결과
처리한 경우 97.95%로 정확도가 향상되었음.
/
34 Case #2: Accomplishment
Manuscript ID amiajnl-2018-006960 entitled "Use of mind
maps and iterative decision trees to develop a guideline-based clinical decision support system (CDSS) for routine surgical practice: Case study in thyroid nodules" which you submitted to the Journal of the
American Medical Informatics Association, has been Accepted
(IF4.2).
HyungWon Yu, JY Choi, Ho Sung Han, Seoul Natl Univ, Seongnam, Korea, Republic of Korea; Maqbool Husain, Sungyoung Lee, Kyung Hee Univ, Suweon, Republic of Korea
Next Silo: Adrenal Tumor (Treatment)
/
35 Case #3: Eye (Retina) Silo (Follow up)
Deep learning model Best Follow up date
Rule based Expert system
Internal medicine data Eye data Deep learning input
Deep learning model :
Input: Patient history data and current eye data (value and video), internal medicine data (value) Output: Expected value for time t to decide best Follow up date (Ex: Blood sugar level, eyesight, intraocular pressure)
Rule based expert system :
Input: Expert knowledge and deep learning output Output: Best Follow up date (Ex: after 6 months)
Gender : male Age : 65 Retinal thickness : 0.7 Diabetes : 300
……
/
36 Case #4: 12 weeks Diabetes Management Program (follow up)
/
37
Adaptive Recommen dation Lifestyle status Behavior Status Adoptive Question
Case #4: Output of Diabetes Management Services (Wellness)
/
39
Will AI Replace Doctor?
/
40 Role of AI Doctor