AI IN MEDICINE

1
AI IN MEDICINE
Pic Google

• Tejashree Aher (06011011)
• Akhil Deshmukh (06D05007)
• Anshul Maheshwari (06D05009)
• Narendra Kumar (06D05008)

2
Overview

• Motivation
• AIM
• What is AIM?
• Goals of AIM
• Applications of AIM
• Clinical expert system MYCIN
• Introduction
• How it works
• Specification of the therapy selection problem
• Representation of Goals
• Certainty factor
• Partial derivation of the algorithm

3
Motivation

• Doctors in a box to diagnose diseases.
• Community of computer scientists and healthcare
  professionals set a research program - Artificial
  Intelligence in Medicine (AIM) with the aim of
  revolutionize medicine.

4
What is AIM?

• Clancey and Shortliffe Medical artificial
  intelligence is primarily concerned with the
  construction of AI programs that perform
  diagnosis and make therapy recommendations.
• Medical AI programs are based on symbolic models
  of disease entities and their relationship to
  patient factors and clinical manifestations
• AI specialized to medical applications
• Employ human- like reasoning methods in the
  programs

5

• AIM use machine learning to use and create
  knowledge.
• Machine learning - computers that can learn from
  experience
• Use stored data used in diagnosis
• Creation - analyse the relationship within the
  data to come up with new results
• Used in Drug discovery

6
Goals of AIM?

• Expert computer programs for clinical use
• Dissemination of the best medical expertise to
  geographical regions where that expertise is
  lacking
• Making consultation help available to
  non-specialists not within easy reach of expert
  human consultants.
• To formalize medical expertise

7
Applications of AIM

• Knowledge based systems
• Diagnostic and educational systems

8
Knowledge based systems

• Use the medical knowledge stored for reasoning
• Store information about a specific task
• Knowledge represented in the form of set of rules
• Support healthcare workers in the normal course
  of their duties -manipulation of data and
  knowledge
• Examples Generating alerts and reminders -warn
  changes in a patient's condition (in less
  critical cases, through a email)
• Agents for information retrieval- software
  agents are sent to search for and retrieve
  information

9
Diagnostic and educational systems

• Most research systems were developed to assist
  clinicians in the process of diagnosis.
• Expert System
• A program that contains a large amount of
  knowledge in one specific area.
• Rules for organizing and expressing its knowledge
• Approaches to integrate the recommendations

10
MYCIN

• Created in the mid-1970s,helps doctors choose the
  correct antibiotics for patients with severe
  infections (and the best ones !!!!)
• It is given large amounts of information on
  meningitis and bacteremia
• This information represented as -if A and B are
  true(evidently), then there is evidence that C is
  true.
• Dynamic computation
• Same recommendation with different certainty
  factors, MYCIN integrates them by means of a
  numerical function.

11
How MYCIN Works ???

• Diagnose for infectious diseases.
• Identify infection that requires therapy,
• What is the identity of the organism(s) by
  clinical and laboratory evidence.
• primary, secondary.
• What are the potentially useful drugs
• chloramphenicol (0.95)
• clindamycin (0.95)
• erthromycin (0.77)
• tetracycline (0.41)
• carbenicillin (0.25)
• 5. Which will be best ? (yes, it suggests the
  best one!)

12
Fig by M. Chandra and V. K. Sonkar
13
Example

• Joe shows the following disorders
• Headache
• Bodyache
• Nausea

What exactly is wrong with Joe??
MYCIN has the answer.
Pic Google
14
Organization of MYCIN
MYCIN
PATIENT
RULE BASE
PATIENT DB
MEDICAL EXPERT
Fig by M. Chandra and V. K. Sonkar
15
The Knowledge Base

• Inferential knowledge stored in decision rules
• If Premise then Action (Certainty Factor CF)
• If AB then C (0.6)
• The CF represents the inferential certainty
• Static knowledge
• Natural language dictionary
• Lists (e.g., Sterile Sites)
• Tables (e.g., primary, gram stain, morphology,
  aerobicity)
• Dynamic knowledge stored in the context tree
• Patient specific
• Hierarchical structures Patient, cultures,
  organisms

16
Fig by Yuval Shahar
17
Specification of the therapy selection problem

• Given a diagnosis (one or more organisms
  suspected of infecting the patient), choose the
  therapy (set of drugs) that best satisfies the
  following medical goals
•  
• Maximize drug sensitivity.
• Maximize drug efficacy.
• Continue prior therapy.
• Minimize number of drugs.
• Give priority to covering likelier organisms.
• Maximize number of suspected organisms covered.
• Dont give two drugs from the same general class.
• Avoid contraindications for the patient.
•  

18

• How to choose the best therapy???
• It subject all the therapies to the following
  three tests -
• Coverage test.
• Classes of selected drugs in a therapy.
• Contra-Indication.
• A therapy is suggested or rejected , Explanation
  !!!

19
MYCIN Algorithm

• Representation of goals
• Set of axioms
• Partial ordering
• Preference order
• Linear ordering
• Metric representation
• Partition
• Yes/no predicate

• Certainty Factor.

20
Contd.

• Certainty Factor
• What is Certainty factor?
• How does it combine?
• Proceeds as
• Several rules single hypothesis.
• Several propositions together.
• Following the chaining rule.

21
Cont.

• Measure of belief MBh, e.
• Measure of disbelief MDh, e.
• Certainty factor CFh, e MBh, e MDh, e.

22
Cont.

• Combination of evidences
• MBh, s1 s2 0 if MDh, s1 s2 1
• MBh, s1 MBh, s2(1- MBh, s1) else
• MDh, s1 s2 0 if MBh, s1 s2 1
• MDh, s1 MDh, s2(1- MDh, s1) else
• Combination of hypothesis
• MBh1 ? h2,e min(MBh1,e ,MBh2,e )
• MBh1 ? h2,e max(MBh1,e ,MBh2,e )

23
Certainty Factors

• Certainty Factor (CF) with its conclude
  functions,
• Conclude function-
• Say the current CF value is x, and a new evidence
  with CF y is supporting the same hypotheses
  comes, then
• F(x,y) xy(1--x) if
  x, y 0,
• xy(1 x) if x,
  ylt0, x, y 1.
• (x y)/(1 - min(x,y)) else.
• Conclude derives a conclusion including the CF of
  the result
• E.g., There is suggestive evidence (0.7) that
  the identity of the organism is streptococcus.
• It is always true that -1 CF 1
• If CF 1 then all other hypotheses are rejected

24
Example

• Joe has a disease A
• bodyache headache-gtyes (0.7) ...e1
• headche weakness -gt yes (0.8) ...e2
• no weakness -gt no (0.6) .e3
• weakness nausea -gt yes (0.6) ....e4
• Joe comes to doctor-
• headache? yes
• bodyache ? yes
• weakness ? no
• nausea ? yes

25
CF(headache (Joe, yes)) 0.7 CF(weakness (Joe,
yes)) 0.65 CF(nausea (Joe, yes))
0.4 CF(bodyache (Joe, yes)) 0.8 MD(joe, e3)
CF(e3) max(0, CF(weakness)) 0.6
(1-0.65) 0.210 get MB(joe, e1)
CF(e1) max (0, min(CF(bodyache),
CF(headache))) 0.7 0.7 0.49 MB(joe,
e2) CF(e2) max (0, min(CF(weaknes8),
CF(headache))) 0.8 0.65
0.52 MB(joe, e4) CF(e3) max (0,
min(CF(weaknes8), CF(nausea))) 0.6 0.4
0.24 MB(joe, e3) CF(e4) max (0, min(CF(no
weaknes))) 0.4 0.6 0.24 MB(joe,
e1,e2) 0.49 0.52 (1-0.49)
0.7552 MB(joe,e1,e2,e4) 0.7552 0.24 (1-
0.0.7552) 0.813 MD(joe,e3) 0.6 0.24
0.144 CF (joe, fever) MB(joe, fever) -
MD(fever) 0.813-0.144
0.669 . Chances of Joe having fever !!
Pic Google
26
Partial derivation of the algorithm

• Representing Goals
• Linear ordering ltfewer
• Matric scale 100-1000
• Considering the above example
• Drug (A) ltfewer Drug (B)

27
Preference ordering and Partition
Preference ordering CONDENSE, a many to one
function F(x). F(x)ltF(y) gt xlty PARTITION M(x)
-gt F(x) F(x) ?(x) i ti-1 ltp M(x) ltp ti t0
p M(x) p tn1 -gt re-formulation of
constraint.
Drawback of CONDENSE
F(x) lt F(y) significant difference
28
EXTENDsion and CONJOINing

• EXTEND An ordering on individual items to an
  ordering on bags of items, follows
• x lt y iff x lt y.
• If X lt Y and X lt Y, then XX lt YY, where
  denotes bag union. For example 1lt2 implies 1 lt
  2 and l, l lt 1,2.

• CONJOIN We combine the preference ltfewer
  for fewer drugs with the preference lteffective
  for more effective therapy by Conjoining them.
• x pq Y iff x P y and x q y //x is
  atleast as good as y
• x ltpq iff (x p y and x ltq y) OR (x ltp y and x
  q y) //x is preferable
• (Note that A lteffective B means therapy A is more
  is more effective than therapy B, ie. More
  preferable with respect to the effectiveness.)

29
Combine coverage preferences

• The therapy goals listed in above include
  maximizing the number of organisms covered and
  giving priority to those the patient is likelier
  to have. Lets see how these two goals are
  integrated
•  
• Classify organisms as most likely or less
  likely.
• 2. Relax the coverage goal by ignoring less
  likely organisms.
• 3. Reformulate the coverage goal as the
  constraint that all the most likely organisms
  be covered.

30
Domination of Preferences1. Letting one
preference -- ltprimary , ltsecondary using
ltsecondary only to resolve ties .X lt primary Y
Or (X primary Y and X ltsecondary Y). 2. A
preference can simply be IGNORED. For example,
ignoring ltsecondary ltprimarysecondary to
lt primary ' This particular case of IGNORE is
appropriate if ties with respect to ltprimary are
too rare to worry about, or if violating
ltsecondary in the event of such a tie wouldnt
do much harm. It is unlikely for two therapies
to be equally effective on the likeliest
organisms but different on the less likely ones,
so it is reasonable to ignore the less likely
organisms altogether. 3. The Condensed
preference compares therapies based on the number
of most likely organisms covered. This
preference is now reformulated into a constraint
by THRESHOLDING.  THRESHOLD (tmin) M(x)
µ(X) , (M(x) tmin),
31
Maximizing therapy effectiveness appears more
important than minimizing the number of drugs, in
the sense that increasing therapy effectiveness
by 1 rank is considered more desirable than
reducing the number of drugs by 1.
32
-Why MYCIN-

• Addresses the problems of reasoning.
• Provide clear and logical explanation of
  reasoning.
• Explore how human experts make these rough (but
  important) guesses.
• Useful for junior or non-specialized doctors.

33

• 
  -MYCIN-
• Does it always thinks like an Expert??
• But not always good to use drugs with high
  effectiveness .
• So it is always preferred by professional doctors
  to start with low concentration ( low mg) drugs,
  than increase it step by step if effects are not
  significant.
• At the time of the first study, MYCIN rules
  included only bacteremia (meningitis and
  endocarditis were added later), thus never tested
  in a real clinical environment with general
  infections

34
Summary

• Reduction in Medication Errors and Adverse Drug
  Events.
• Computerassisted - fewer errors than
  handwritten prescriptions and to be five times
  less likely to require pharmacist clarification
• Prompt to use a cheaper generic drug when a more
  expensive drug was initially ordered
• Cannot model common-sense
• Cannot be completely relied upon ( loss of
  confidence !! )
• The knowledge-acquisition bottleneck remained
  significant (additional effort from already busy
  individuals !!!)

35
Contd.

• Rely on human knowledge
• The program acts as advisor to a person
• Medical practitioners serve as a critical layer
  of interpretation between an actual patient and
  the expert systems
• Limited ability of the program to make a few
  common sense inferences is enough to make them
  usable and valuable

36
References-

• Peter Szolovits , Artificial Intelligence and
  Medicine, Westview Press,1982.
• Towards Explicit Integration of Knowledge in
  Expert Systems An Analysis of MYCINs Therapy
  Selection Algorithm, Bill Swartout, Jack Mostow,
  AAAI-86 ,1986.
• http//www.openclinical.org/gmm_ardensyntax.html.
• Peter Szolovits, William J. Long, The Development
  of Clinical Expertise in the Computer, Westview
  Press,1982.
• Athanasios K. Tsadiras, Konstantinos G.
  Margaritis, The MYCIN certainty factor handling
  function as uninorm operator and its use as a
  threshold function in artificial neurons, Fuzzy
  Sets and Systems 93,1998.
• Yuval Shahar, Diagnostic Systems (I),Medical
  Decision support systems, Stanford
  Univarcity,2007.