The primary goal of this workshop is to help health professionals gain a critical understanding of the various types of AI technologies available so they can make wise decisions and invest AI for healthcare improvement.
Chen & Decary, ITCH 2019
AI in Healthcare:
From Hype to Impact
Workshop presented at ITCH 2019: Improving Usability, Safety
and Patient Outcomes with Health Information Technology
Victoria, BC, Canada
Feb 14, 2019
Cogilex R&D Inc.
Mei Chen & Michel Decary
Chen & Decary, ITCH 2019
Achievement, excitement,
hype, & fear
Google’s AlphaGo has defeated
human Go master
Self-driving cars are as safe
as human drivers
AI defeated human doctors in
contest to diagnose tumors
AI will replace doctors,
especially radiologists
Machine has surpassed
human intelligence
AI promises a new paradigm for healthcare & it
will revolutionize the industry
Chen & Decary, ITCH 2019
A reality check
In general
● 60-85% all big data strategies have failed due
to a combination of challenges ( Walker, 2017,
Digital Journal).
In healthcare
● Rapid AI development in the industry;
● Many success stories, but mostly related
small-scale pilots or research projects;
● Some heavily invested projects have failed;
● Early hype around AI promised miracle cures
and delivered few results (Hyde, 2018,
Forbes).
Chen & Decary, ITCH 2019
Challenges for
using AI in
Healthcare
• Inadequate understanding about what a
particular type of AI technology can or
can’t do
• Lack of good implementation strategies
• Incompatibility with legacy technologies
and data
• Shortage of trained workforce
• Pre-existing corporate biases
Chen & Decary, ITCH 2019
Objectives of the workshop
Understand the real potential of AI for healthcare and gain
insights about how to invest wisely in AI .
● Role of AI in healthcare
● Types of AI technologies useful for healthcare
What is it?
How does it work?
What is the context of its proper use?
● Insights about how to use AI to support best
medical decision making and clinical practice
Exemplar AI applications in healthcare
Next-gen AI-powered EHR systems
Chen & Decary, ITCH 2019
Artificial intelligence(AI)–A subfield of
computer science
● AI is machine that simulates aspect of learning or any other feature
of human intelligence (John McCarthy, 1956).
● The theory and development of computer systems able to perform
tasks normally requiring human intelligence.
● Being able to pass the Turing test
Chen & Decary, ITCH 2019
AI in healthcare: Augmented Intelligence
AI as a powerful tool and partner
*
Man + machine = enhanced human capabilities (AMA, 2018)
AI can help human
• Unlock the power of big data and gain insight into patients
• Support evidence-based decision making, improving quality, safety,
and efficiency
• Coordinate care and foster communication
• Improve patient experience and outcomes
• Deliver value and reduce costs
• Improve health system performance & optimization
Chen & Decary, ITCH 2019
Human-machine Partnership in Healthcare
AI-powered
Automation
Improving
Effectiveness
AI-powered Automation
Man Machine
Improving effectiveness
● Quality
○ Experience
○ Outcomes
● Safety
○ Ways to ensure
patient safety
● Efficiency
○ Usability
○ Productivity
● Access to care
Controlling costs
AI-powered automation
• Medical robotics
o Surgical robots
o Rehabilitation robots
o Smart pills
• Machine learning
o Supervised learning
o Unsupervised learning
o Reinforcement learning
• Natural language processing
o Statistical vs. rule-based NLP
• AI voice technology
o Clinical voice documentation
o Voice nursing assistants
Chen & Decary, ITCH 2019
10 Promising AI Applications in Health Care
Source: Harvard Business Review, Kalis, Collier, Fu, 2018
https://hbr.org/2018/05/10-promising-ai-applications-in-health-care
E
H
R
Chen & Decary, ITCH 2019
Types of AI technologies for Healthcare
● Medical robotics
Surgical robots, rehabilitation robots, smart pills, senior’s robotic companion
● Machine learning/Deep learning
Supervised learning-->unsupervised learning-->reinforcement learning
● Natural language processing (NLP)
Statistical vs. rule-based NLP
● AI voice technology
Medical voice documentation; AI nursing assistants (voicebots)
Chen & Decary, ITCH 2019
Arthur Samuel, 1950
The field of study that gives computers the ability to learn without being explicitly programed.
Google dictionary
Machine learning is a program or a system that builds or trains a predictive model from input data.
The system then uses a learned model to make useful predictions from new, never-before-seen data drawn
from the same distribution as the one used to train that model.
What is machine learning (ML)?
Chen & Decary, ITCH 2019
Supervised learning
Unsupervised learning
Semi-supervised learning
Reinforcement learning
ML can be applied to all data types: images, audios, text, and videos
Machine learning algorithms:
● Given input (a data set);
● known outcomes (labeled);
● Look for patterns correlated
outcomes with input to make
prediction
Chen & Decary, ITCH 2019
Identifying the underlying structure/patterns of data and trying to map input
variables into discrete categories
Most AI successes are achieved through supervised learning
Data-driven clinical decision support:
● Diagnostic analysis using medical imaging (e.g., CT-scan, x-rays, MRI results)
● Making medical diagnosis based on the symptoms
● Designing and selecting treatments based on biomarkers or other attributes
● Identifying at-risk groups based on health-related factors and social determinants
Supervised learning: Classification problem
Chen & Decary, ITCH 2019
Classification problem: Classifying disease severity
Source: Lily Peng, Google AI Healthcare
Chen & Decary, ITCH 2019
Classification problem:
Identifying different types
of diseases
CheXpert
A large dataset of chest X-rays
and competition for automated
chest x-ray interpretation
--Launched in 2019 by Stanford university
Image Source: Unspecified
Chen & Decary, ITCH 2019
Supervised learning: Regression problem
Predicting results within a continuous output.
Predictive analytics in healthcare:
● Predicting inpatient mortality and long length of stay based on EHR Big Data
(Google and partners, 2018)
● Predicting the number of ER patients during a specific period, needed staff
and beds based on past data
● Predicting patients’ survival rates based on health conditions, age, and other
characteristics
Chen & Decary, ITCH 2019
Monitoring
tumor response
to treatment
Beyond diagnostics: Machine learning on medical
images for real-world clinical research
Image source: Unspecified
Chen & Decary, ITCH 2019
Natural language processing (NLP)
Natural language processing (NLP) deals with building computational
algorithms to automatically analyze and represent human
Language, particular text.
Statistical NLP vs.
Rule-based approach
Chen & Decary, ITCH 2019
Steps Involved In rule-based NLP:
1. Morphological Processing:
1. Syntax Analyzing:
1. Semantic Analysis:
1. Pragmatic Analysis:
Rule-based NLP
Uses:
● Text Mining (entities, relationships)
● Classification
● Text summary
● Machine Translation
● Question and Answering
● Text creation
● Semantic search
Semantic search is the transition “from being an information
engine to a knowledge engine (Google, 2012)
Chen & Decary, ITCH 2019
A cognitive-based semantic approach to NLP:
• Using semantic rules in conjunction with a world model and
cognitive frameworks to semantically analyze, rank, select, and
extract web content from trusted sources;
• Analyzing health information in relation to user’s goals, tasks and
the information needs in specific situations;
• Using knowledge models and rule-based NLP analysis as a base
for creating patient-centered digital health technologies (e.g.,
smart medical search engine, medical voice chatbots)
Cogilex’s rule-based NLP
Chen & Decary, ITCH 2019
Key features of Cogilex’s semantic search
technology
1. Identifying selfcare information
related to:
● Disease entity > 25,000
● Symptom entity > 4,500
● Injury and accident entity > 1,500
● Medical procedure entity > 9,500
● Drug entity > 8,000
● Other health related object classes
3. Using machine learning on
big data to extract relationships
among:
● Symptoms
● Tests
● Treatment modalities
● Drugs
● Dietary plans
● Etc.2. Classifying health information and
generating knowledge maps to guide
user’s search
Chen & Decary, ITCH 2019
Seenso
medical
search
engine:
Analyzing web
content from
trusted sources
and providing
knowledge maps
to guide user's
search
Chen & Decary, ITCH 2019
NLP applications in Healthcare
● IBM’s Watson Health
● Amazon Comprehend Medical
● Google Deepmind health
● United Health Group Inc.’s Optum
● MetaMap (NLM)
● cTakes (Mayo clinical text analysis and knowledge extraction system)
● Linguamatics
● CLAMP (Clinical Language Annotation, Modeling, and Processing) Toolkit
Quotes
Chen & Decary, ITCH 2019
IBM Watson Cognitive Computing
The goal of cognitive
analytics is to generate
new insights using
artificial intelligence (AI)
and machine learning
algorithms that can
understand, reason and
learn.
Image source: Unspecified
Chen & Decary, ITCH 2019
Amazon Comprehend Medical
● Support for entity extraction and entity traits on a vast vocabulary of medical terms:
anatomy, conditions, procedures, medications, abbreviations, etc.
● An entity extraction API (detect_entities) trained on these categories and their subtypes.
● A Protected Health Information extraction API (detect_phi) able to locate contact details,
medical record numbers, etc.
Chen & Decary, ITCH 2019
● Medical dictation for
professionals: Nuance
Dragon Medical Practice
● Voice interfaces for
EHRs: an essential
step to humanizing the
EHRs
● AI Assistants for
consumers
AI Voice technology
Mycroft: An open source AI voice assistant for Linux-
based operating systems
IBM Watson and NVIDIA AI Platforms
Chen & Decary, ITCH 2019
Medical dictation for professionals: Nuance Dragon Medical
Practice
Nuance® Dragon® Medical Practice Edition 4 accurately translates the
doctor's voice into a rich, detailed clinical narrative that feeds directly into
the EHR. Improve documentation, boost efficiency, increase physician
satisfaction, and eliminate transcription costs.
Chen & Decary, ITCH 2019
General AI voice assistants
AI Assistants for consumers
Text-based medical chatbots
Your.MDSensely
Mycroft
(open source)
Orbita Voice
IBM Watson
Nuancen
Conversa
● AI Assistants like Alexa, Siri,
Cortana, and Google
Assistant do general
question-answer matching
and do not perform tasks
directly related to patient care.
● Health chatbots such as
Babylon, Ada, and Buoy rely
on text-based and mostly
close-ended communications
Chen & Decary, ITCH 2019
Big data in EHR:
Quantitative Data
Vital signs, diagnoses, diagnosis codes, laboratory results,
medication
Qualitative Data (80%)
Clinical notes, medication order notes, discharge instructions
Challenges
-Many data types, many user types, large data size and high
complex
-Processing such data and generating knowledge is moving
beyond unassisted human capacity
EHR systems–the backbone of
digital health transformation
Chen & Decary, ITCH 2019
Barriers in using EHRs:
For patients
● Complexity of EHR systems
● Patients’ lack of basic health literacy
● Patients’ preference to discuss their
health issues with care providers in
person (76% responders, Heath, 2018)
● Perceived limited use of EHR portals by
patients (59% responders) (Heath, 2018)
● Physicians want EHRs to be designed to
facilitate digital and mobile patient
engagement
For healthcare professionals
● AMA demands EHR overhaul,
calls them 'poorly designed and
implemented’,
● Studies indicates that typing and
clicking consume more than half
the workday for doctors,
contributing to physician burnout
● The majority physicians think
EHRs need a complete overhaul
Chen & Decary, ITCH 2019
Next-gen EHRs: AI-powered EHR (1 of 2)
From “systems of records” to “systems of intelligence” and “systems of
engagement”:
● Better clinical decision support:
○ Diagnostic analytics using medical imaging;
○ Predictive analytics using big data;
○ Routine integration of medical imaging with other clinical data for triage
and critical care monitoring, diagnostic interpretation, and treatment
modification
○ Personalized treatment design: precision medicine and behavior
modification
● Smarter search algorithms
● Full integrated NLP capacity for narrative health data (e.g., critical summary of
patient info, physician’s notes);
Chen & Decary, ITCH 2019
Next-gen EHRs: AI-powered EHR (2 of 2)
● Integrated voice technology
-AI voice assistant for health professionals (e.g., clinical dictation,
voice interface, question asking and answering)
-AI voice assistants for productive patient engagement (e.g.,, AI
nursing assistants/chatbots)
● Integrating data information from multiple sources (including mobile
data)
● Population health monitoring and management (epidemiology, social
determinants of healthcare)
● Mechanism for safety keeping
-Preventing medication errors
-Preventing harmful drug interaction
-Preventing complications of treatment to pre-existing conditions
● Content and tools for productive patient engagement
Chen & Decary, ITCH 2019
Benefits of machine learning for healthcare
Deep learning has produced remarkable results for complex real-world problems
that involve big data. It has the potential to provide data-driven, evidence-based
clinical intelligence for advancing both biomedical research and service delivery
across the full spectrum of healthcare.
Koski (2018):
• A new paradigm to derive insights on biological, diagnostic and therapeutic processes
and behaviors from data
• Accelerate the process of digesting and interpreting vast quantities of complex,
diverse information
• Enable new data-driven presentation, diagnostic, treatment, and management options
Chen & Decary, ITCH 2019
Machine learning is not an all-purpose solution
For tasks that require common-sense solutions or domain-specific expertise, and situations that are
outside of the ML training dataset, machine learning is less applicable.
ML weaknesses:
● Identifies superficial features and patterns, but lacks the understanding of meanings and concepts;
● Identifies correlations but not causal relations;
● Lacks common sense reasoning, general intelligence, and domain knowledge integration;
● Lacks explainability and it is hard to fix certain identified problems;
● Needs big data, machine learning models are as good as the training data (biases, noises, errors
often exist in real-world data);
● Difficulty to generalize the finding beyond its training dataset.
Limitations of machine learning for healthcare
Chen & Decary, ITCH 2019
Key requirements for AI success in healthcare
● Understanding what a particular type of AI technology can or
can’t do
● Specifying the context of its proper use
● Developing an AI strategy that will bring real value to your
organization and being able to implement it
● Establishing performance standards:
○ Evaluation criteria
○ Performance measure
○ Pilot-implementations and validation
● Ensuring privacy, security, and ethics
Chen & Decary, ITCH 2019
Medical imaging
● Mayo Clinic neuroradiologists are using AI to find genetic markers in MRI scans.
● Stanford researchers have developed an AI algorithm that can diagnose up to 14 types of medical
conditions and is able to diagnose pneumonia from medical images.
● Memorial Sloan Kettering Cancer Center is collaborating with an imaging company to improve the
diagnosis of prostate cancer.
● University of Warwick is using an AI system to analyze chest X-rays and spot patients who should
receive immediate care
● Google’s DeepMind, is working to develop a commercialized deep learning CDS tool that can
identify more than 50 different eye diseases – and provide treatment recommendations for each
one.
Examples of AI for healthcare
Chen & Decary, ITCH 2019
Machine learning, voice technology and EHR integration
● Epic, Cerner Allscripts and others are building EHRs that feature automation analytics, voice
dictation, and tools for patients
● IBM Watson health
● Apple mobile devices, data and EHR integration
● UK NHS Medical diagnosis and services using Babylon mobile app
● Google’s EHR analysis to forecast patient outcomes
● Canada Mackenzie Health EHR implementation that includes clinical voice documentation
● Mayo clinical text analysis and knowledge extraction system, voice integration in HER
Examples of AI for healthcare
Chen & Decary, ITCH 2019
AI platforms & services
These platforms typically provide
functionality for:
• -Natural language processing,
• -Image recognition,
• -Question-answer matching,
• -Voice recognition,
• -Predictive analytics, etc.
1. Amazon web service
2. Google cloud
3. Microsoft Azure
5. IBM Watson
4. MonkeyLearn
6. NVIDIA Platform
7. Nuance Platform
8. OpenAI
9. SAS
Chen & Decary, ITCH 2019
Discussion
1. Give an example of the best AI integration in healthcare
2. What are the priorities and challenges for AI in healthcare?
3. AI development and integration strategies:
○ Given the high complexity and costs involved in developing the next-gen EHRs, should each
country build a national AI-powered EHR system?
○ Should the next-gen EHRs be built with incremental improvements or a complete overhaul?
○ What are the keys to interoperability and data exchange?
○ How can AI transform healthcare? To support the existing clinical workflows or new ways of
practicing medicine?
4. You thoughts on the benefits, costs, and feasibility of developing AI in healthcare.
Chen & Decary, ITCH 2019
Key References
1. Artificial intelligence in healthcare: past, present and future
2. Artificial intelligence, bias and clinical safety
3. AI in healthcare - not so fast? Study outlines challenges, dangers for
machine learning
4. Artificial Intelligence and Machine Learning Workshop
5. AMA AI Policy
6. 3 charts show where artificial intelligence is making an impact in healthcare
right now
7. AI and machine learning: What cuts hype from reality?
https://www.healthcareitnews.com/projects/ai-and-machine-learning
Note: All references, including the unspecified image sources are to be
completed