Transcript Natural Language Processing COMPSCI 423/723

Artificial Intelligence in Medicine
HCA 590 (Topics in Health Sciences)
Rohit Kate
1. Introduction and Overview
References
• Chapter 1, Artificial Intelligence: A Modern
Approach by Stuart Russell and Peter Norvig
• Paper: Patel et al. The coming of age of
artificial intelligence in medicine. Artificial
Intelligence in Medicine (2009) 46, 5—17
• Paper: Ramesh et al. Artificial Intelligence in
Medicine. Ann. R. Coll. Surg. Engl. 2004
Sep;86(5):334-8
• Guest Editorial articles from Artificial
Intelligence in Medicine journal
What is Artificial Intelligence (AI)?
• Intelligence, generally speaking, is the ability to do
the right or the best thing in a given situation to
achieve a goal
• Humans and some animals exhibit intelligent
behavior; this is intelligence that exists in nature
• Artificial Intelligence: Capability of a man-made
machine to exhibit intelligent behavior
– A chess playing machine
– A robot or a car that finds its way around
– A computer that answers natural language questions
about a topic
– A computer that diagnoses a patient from the symptoms
– ...
What is AI?
• Various other definitions of AI are possible ranging
from philosophical to engineering perspectives
– Machines with minds…
– Study of design of intelligent agents…
– A good working definition: Study of how to make
computers do tasks at which currently humans are better
(Rich and Knight, 1991)
– AI is a study of duplicating human faculties like creativity,
self-improvement and language use
• “Artificial” Intelligence may seem to mean that the
intelligence is not real or that it just simulates natural
intelligence, perhaps a better term could have be
“Computational Intelligence” or “Machine
Intelligence”
What is NOT AI?
• Word processing softwares
• Graphics design using computers
• Entering or searching for an entry in a
database
• Computer networks, Internet protocols
• Computer security
• Design of a processor
These tasks do not require the computer being
used to be intelligent.
•
•
•
•
What Makes Humans
Intelligent?
Knowledge
Reasoning
Learning
Language and Perception
All these are reflected in the major sub-areas
of AI
How to Make Computers
Intelligent?
• Should we model human intelligence
– A good idea but difficult to model
– Human brain is different from computer processor
• Humans are good at remembering and recognizing patterns;
computers are good at crunching numbers
• A compromising approach: Model human
intelligence as much as possible but also utilize
computer’s ability to crunch numbers
– Airplanes have wings like birds but they don’t flap them,
instead they use engine technology
Different Sub-Areas of AI
• Knowledge Representation and Reasoning
– How to encode knowledge and reason from it
• Humans do it all the time (common-sense knowledge
and reasoning, expert knowledge and reasoning)
• First-order logic, ontologies, rules, knowledge-bases
• Provision for Uncertain knowledge, probabilistic
reasoning
– Encode knowledge about diseases, symptoms etc.
and then predict diagnosis
Different Sub-Areas of AI
• Machine Learning
– Improve performance by learning from examples
• Humans do it all the time (learn to walk etc., develop
special skills)
• Rule-based methods: e.g. decision-trees
• Statistical methods: e.g. neural networks, support
vector machines, maximum entropy models
– Learn to diagnose a disease from previous
examples of patient data
Different Sub-Areas of AI
• Computer Vision
– “Eyes of computer/robot”
• Recognize objects (e.g. face, human) from an image
• Reconstruct a 3D model from 2D images, e.g. track an
object
– Distinguish a cancerous from a non-cancerous
radiology image
Different Sub-Areas of AI
• Natural Language Processing
– Understand and process natural languages like
English, Chinese etc.
• Natural language is the preferred medium of
communication for humans
• Follow natural language commands
• Answer natural language questions
• Find required information from several documents
• Translate from one natural language to another
– Answer clinical questions using a repository of
research articles
Different Sub-Areas of AI
• Robotics
– Physical agents that act in the physical world
– Surgical robots
• Planning
– Coming up with a best sequence of interdependent tasks to perform (e.g. wear socks
before shoes)
– Planning and scheduling in a hospital environment
Turing Test
• How to decide whether a machine is
intelligent?
• Should we make a list of qualities of
something that is intelligent? How to come up
with such a list? Will everyone agree?
• Instead Alan Turing (1950) proposed a test of
indistinguishability from humans as an
operational definition of intelligence of a
machine
Turing Test
• A machine passes this test if a human cannot
distinguish whether it is conversing in writing
with a human or a computer behind closed
doors
Turing Test
• To pass the Turing test a machine will need
the capabilities of
– Natural Language Processing
– Knowledge about the world
– Automatic Reasoning
– Machine learning
• Total Turing Test: Also test the subject's
perceptual abilities through video and passing
physical objects; additional capabilities of
– Computer Vision
– Robotics
Turing Test
• Passing Turing test requires capabilities of all
the major sub-areas of AI
• Loebner Prize: Current contest for restricted
form of the Turing test
– Usually dominated by hacks to fake human
conversations
– Not of much interest to real AI researchers
• Emphasis of AI research is not in passing this
test but on doing well on various tasks that
require intelligence, this will eventually lead to
a system that will also faithfully pass this test
A Turing Test for Medicine??
• A medical doctor can’t distinguish whether
conversing with another medical doctor or
with a “medically expert” computer
Foundations of AI
• AI is extremely inter-disciplinary; its foundations
come from several older disciplines
– Philosophy
• Where does intelligence come from?
– Mathematics
• How to infer logically? How to reason under uncertainty?
– Economics
• How should we make (intelligent) decisions that maximize payoff ?
– Neuroscience
• How do brains process information?
– Psychology
• How do humans and animals think and act?
– Linguistics
• How does language relate to thought? How do we process
language?
History of AI
• Relatively brief history, only 50-60 years old
• Interesting with many ups and downs
• A look at its history helps to understand how
the current focus and methodologies in AI
have emerged
History of AI
• Beginnings
– McCulloch and Pitts (1943) proposed a model of
artificial neurons could compute any computable
function
– Marvin Minsky (1951) built the first neural
network computer using vacuum tubes
– Alan Turing (1950) introduced Turing test,
machine learning, genetic algorithms and
reinforcement learning
History of AI
• Birth of AI
– Dartmouth conference (1956)
– Organized and attended by some of those who are
now regarded as founders of AI: John McCarthy,
Marvin Minsky, Allen Newell, Herb Simon
– Coined the term “Artificial Intelligence”
– Presentation of a reasoning program, "Logic
Theorist" which could automatically prove many
mathematical theorems
History of AI
• Early Years (1950s and 60s)
– Several interesting and impressive AI work that
people earlier did not believe that computers
could ever do
– General Problem Solver: Could solve limited
classes of puzzles thinking like humans
– Geometry Theorem Prover: Could prove theorem
that were tricky for students
– Checkers player: Disproved the idea that
computers can only do what they are told to do,
soon the program learned to play better than its
creator
History of AI
• Early years (1950s and 60s)
– SAINT: Solved freshman calculus problems
– ANALOGY: Solved IQ test analogy problems
– SIR: Answered simple questions in English
– STUDENT: Solved algebra story problems
– SHRDLU: Obeyed simple English commands in the
blocks world
History of AI
• Limitations of Early Systems
– Could only work on "toy" problems which were
not at the scale of real-world problems, for two
main reasons
• Difficult to formalize and encode real-world knowledge
– For example, they tried to build an MT system from Russian to
English using dictionaries and syntactic transformations but
due to lack of world knowledge: "the spirit is willing but the
flesh is weak" -> Russian -> "the vodka is good but the meat is
rotten“
• The systems used simple search to find a solution out
of all the potential solutions, this would not work with
more complex problems which have a combinatorially
large space of potential solutions
History of AI
• Knowledge-based Systems (1970s)
– Realization that domain-specific knowledge could
help finding the solution led to several expert
systems for specific domains
– Encoded rules that human experts would used
and so these systems could perform like human
domain experts
– DENDRAL: First knowledge-intensive system to
infer molecular structure from mass spectometer
data
History of AI
• Knowledge-based Systems (1970s)
– MYCIN: A medical expert system, could diagnose
blood infections from symptoms
• Encoded 450 rules
• Could perform as well as some experts and better than
junior doctors
• But was never used in actual practice because of nontechnical reasons
History of AI
• AI Industry (1980s)
– Several expert systems built and deployed, every
major U.S. company had its own AI group to use
or to investigate expert system
– R1: Helped configure orders for new computers,
saved $40 million a year
– Japanese started a project to build intelligent
computers running Prolog (logic programming
language)
– In U.S. the company MCC was formed with the
same goal
History of AI
• AI Industry (1980s)
– However, limitations of expert systems became
apparent
• Brittleness (won't work with slightly different input),
too domain-specific
• Difficult to acquire all the knowledge even for a specific
domain (knowledge-acquisition bottleneck)
• A brief period of "AI Winter"
History of AI
• Recent years
– Focus on learning from examples to address the
knowledge-acquisition bottleneck
– To counter brittleness: shift of focus from rule-based and
logical methods to probabilistic and statistical methods
(e.g. Bayesian networks, Hidden Markov Models)
– AI has become a science:
• Show real-world applications and not success on toy problems
• Base claims on hard experimental evidence and not on
intuitions
• Analyze results for statistical significance, make data and tools
available to replicate experiments
History of AI
• Recent years
– Instead of remaining isolated like early years, AI
has embraced other disciplines like statistics,
optimization, formal methods etc., whichever
areas are needed for success
– Increased interest in useful applications at the
large scale of the Internet
• search engines, recommender systems, Web site
construction systems
• data mining (find useful patterns in huge amounts of
data)
State of the Art in AI
• Deep Blue beats Kasparov (1997)
State of the Art in AI
• Spirit, and Opportunity explore Mars (2003)
State of the Art in AI
• DARPA grand challenge: Autonomous vehicle
navigates across desert and then urban environment
(2004-2007)
State of the Art in AI
• Automated speech/language systems for
airline travel
• Spam filters using machine learning
• Usable machine translation through Google
State of the Art in AI
• IBM supercomputer to compete with human
champions on the Jeopardy! TV show
• Feb 14-16th 2011
AI in Medicine
• AI systematizes and automates intellectual
tasks and is therefore potentially relevant to
any intellectual activity including Medicine
• Modern medicine is faced with the challenge
of acquiring, analyzing and applying large
amounts of knowledge necessary to solve
complex clinical problems, AI methods fit this
need
• AI in Medicine is a subfield of Biomedical
Informatics as well as Computer Science
Emergence of AI in Medicine
• Over the last few years medicine has become
a data-rich quantitative field because of
various electronic data capturing methods and
data management systems for both clinical
care and biomedical research, this is
transforming medicine from art to science
• The availability of data in electronic form
(documents, articles, clinical notes, electronic
health records etc.) has increased the
necessity and scope of their automatic
intelligent processing using AI techniques
Emergence of AI in Medicine
• Diagnosis, treatment and predicting outcome
depends on complex interactions of many
clinical, biological and pathological variables,
hence there is a growing need for analytic
tools to analyze them
• Note: Many AI in Medicine methods are
becoming more and more integrated within
medical applications often resulting into their
loss of visibility, sometimes not even labeled
as AI in Medicine methods; paradoxically, a
sign of success
AI in Medicine
• Applications of AI in Medicine
– Help in diagnosis and making therapeutic
decisions
– Predict outcomes
– Support healthcare workers in acquiring,
manipulating and searching data
– Guide researchers in making discoveries
AI in Medicine: Sub-Areas
• Knowledge Representation
– Design of good ontologies to enable data
exchange, standardization, communication, e.g.
UMLS, SNOMED-CT, Gene Ontology etc.
– Encode rules obtained from domain experts to
automate processing and reasoning
– Enable discovering new and useful knowledge and
refine existing knowledge
AI in Medicine: Sub-Areas
• Natural Language Processing
– Unlock the value buried in text and narrative
records so that content can be used for
automated processing
– Interact with computer in natural language, ask
clinical questions in natural language to search
research articles
– Information Retrieval: Find the required
information from a collection of documents,
answer questions
AI in Medicine: Sub-Areas
• Decision Support Systems
– Help in clinical diagnosis
– Combine uncertain evidences from multiple
sources and generate a probabilistic diagnosis
• Machine Learning
– image analysis and segmentation in radiology
– data interpretation, waveform analysis
– pattern recognition from medical data
• Data Mining
– Knowledge discovery from databases
– Clinical data mining
AI in Medicine: Example Systems
• MedLEE (http://www.nlpapplications.com/)
– MedLEE™ is a Natural Language Processing
(NLP) application that extracts clinical codes from
typed or dictated free-text medical narratives and
converts the data into computerized clinical
information automatically, quickly and error free
– Developed at Columbia University
– Automates analytics, reporting and alerting for
outflows such as Core Measures, PQRI, Patient
Summary review, Coding and Claims Adjudication,
Decision Support, Clinical Trials, Biological
Surveillance and more
AI in Medicine: Example Systems
• MedLEE (http://www.nlpapplications.com/)
– Supports multiple health care systems in the
hospital to enhance patient safety, quality
assurance, diagnosis and prognosis support, billing
and reimbursement administration. The physician
is not required to change work habits.
– Has been successfully tested by large hospital
systems and government agencies, including the
New York Presbyterian Hospital, the National
Cancer Institute and the U.S. Department of
Defense
AI in Medicine: Example Systems
• GIDEON (http://www.gideononline.com/)
– A global infectious disease knowledge management tool
– Easy to use, interactive and comprehensive web based tool
– Support for the diagnosis and treatment of infectious
diseases, knowledge base is updated weekly about
diseases and their trends
– Hundreds of customers from around the world, including
educational institutions, hospitals, public health
departments and military organizations, use it as their
diagnosis and reference tool for Infectious Diseases,
Microbiology and Occupational Toxicology
AI in Medicine: Example Systems
From: http://www.gideononline.com/
AI in Medicine: Example Systems
• HELP (Health Evaluation through Logical Processing)
– http://www.openclinical.org/aisp_help.html
– "HELP was the first hospital information system to collect
patient data needed for clinical decision-making and at the
same time incorporate a medical knowledge base and
inference engine to assist the clinician in making
decisions" [Gardner et al, 1999]
– Developed at University of Utah, operational at LDS
Hospital in Utah since 1967
– Decision support has been used to provide
alerts/reminders, data interpretation, patient diagnosis,
patient management suggestions and clinical protocols.
– One trial suggested the program had a 94% success rate of
choosing an appropriate antibiotic regimen compared to a
77% success rate for physicians
AI in Medicine: Example Systems
• ATHENA (Assessment and Treatment of Hypertension: Evidence-Based
Automation)
– http://www.openclinical.org/aisp_athena.html
– The ATHENA Decision Support System (DSS) implements guidelines for
hypertension using Stanford Medical Informatics architecture
– Developed by Stanford Medical Informatics
– ATHENA DSS encourages blood pressure control and recommends guidelineconcordant choice of drug therapy in relation to comorbid diseases
– ATHENA DSS has an easily modifiable knowledge base that specifies eligibility
criteria, risk stratification, blood pressure targets, relevant comorbid diseases,
guideline-recommended drug classes for patients with comorbid disease,
preferred drugs within each drug class, and clinical messages
– ATHENA DSS is designed to allow clinical experts to customize the knowledge
base to incorporate new evidence or to reflect local interpretations of
guideline ambiguities.
• See http://www.openclinical.org/aiinmedicine.html for more AI in
Medicine systems in practice
Adapting AI to Medicine
• Medicine is a human endeavor, any AI system needs to
take into account human issues like usability, userfriendliness, user-supportiveness, organizational
change, workflow etc.
• Human expertise in medicine developed over centuries
cannot be discarded or replaced by re-discovering them
by analyzing data; build models that integrates human
expertise with machine learning methods
• The methods that can use the existing knowledge and
can refine or augment it are preferable over the
methods that completely based on data analysis
Adapting AI to Medicine
• Doctor: We need to amputate your finger.
• Patient: Why???
• Doctor: Because our expert system that uses
the most hot-shot state-of-the-art statistical
machine learning technique says so.
• Patient: #@$%&^#
Adapting AI to Medicine
• The output of an AI system must be humaninterpretable, sometimes preference for rulebased machine learning techniques over more
accurate but opaque statistical machine
learning techniques
• Reasoning in medicine is based on arguments,
it is not the accuracy of predictions but their
explanation and communication that matters
• Visualization methods are particularly helpful
for analyzing data
Adapting AI to Medicine
• Case-based reasoning/learning, a machine
learning technique, in which a new case is
processed based on its similarity with previous
cases is particularly useful and popular in AI in
Medicine because it resembles how medicine
is practiced
• Medical data is often imbalanced, one positive
disease-related example to hundreds of
negative disease-free examples; the machine
learning techniques should work well with
such imbalanced data
AI in Medicine: Issues
• Although there has been a remarkable
progress in AI in Medicine but adoption of
these methods have been slow, mostly
because of political, fiscal and cultural reasons
– If an computer makes a wrong diagnosis leading
to bad consequences, who should be held legally
responsible?
– Many learning methods need a lot of data, will
that compromise medical data confidentiality?
– All healthcare workers may not be computer savvy
– How much will doctors trust computers?
– Do doctors feel they are being replaced?
AI in Medicine: Issues
• AI applications are most suited in medicine in
the form of:
– Supporting tools instead of a stand-alone systems,
for example, in suggesting possible diagnoses and
their probabilities
– Covering human mental shortcomings/lapses
• Forgetfulness: reminders of certain tests or medications
• Detect possible errors
– Searching and mining huge amounts of data which
is not humanly possible and present results to
humans
Course Information and
Syllabus
http://www.uwm.edu/~katerj/courses/aim
AI in Medicine: Resources
• Artificial Intelligence in Medicine
– Journal published by Elsevier, accessible online through library’s website
• AIME: A European biannual conference of AI in MEdicine
• OpenClinical.org
– An online resource for knowledge management systems in healthcare
includes AI in Medicine (http://www.openclinical.org/aiinmedicine.html)
• Artificial Intelligence in Medicine, edited by Peter Szolovits
– An old outdated book but still interesting, entirely available online
• http://groups.csail.mit.edu/medg/ftp/psz/AIM82/
• Artificial Intelligence in Medicine Inc (http://www.aim.on.ca/)
– A Canadian company that makes AI based products for healthcare
applications
Homework 1 (Due next week)
• Which of the AI in Medicine systems you
found most impressive and why? As a patient
will you recommend its use on you?