1
Victor Babes UNIVERSITY OF MEDICINE AND
PHARMACY TIMISOARA

• DEPARTMENT OF
• MEDICAL INFORMATICS AND BIOPHYSICS
• Medical Informatics Division
• www.medinfo.umft.ro/dim
• 2004 / 2005

2
MEDICAL DECISION SUPPORT (I)

• COURSE 11

3
ELEMENTARY CYCLE OF MEDICAL ACTIVITY
4
1. MEDICAL DECISION

• 1.1. DIRECTIONS
• COMPUTER ASSISTED DIAGNOSIS
• INVESTIGATION SELECTION
• THERAPY OPTIMISATION
• HEALTHCARE MANAGEMENT

5
1.2. ELEMENTARY CYCLE OF MEDICAL ACTIVITY
6

• 1.2. METHOD CLASSIFICATION
• a) LOGICAL
• TRUTH (SYMPTOM) TABLES
• DECISION TREES
• b) STATISTICAL
• BAYES RULE
• PATTERN RECOGNITION
• c) HEURISTICAL
• EXPERT SYSTEMS

7
2. LOGICAL METHODS

• 2.1. CONSTRUCTIVE PRINCIPLES
• Based on bivalent logics
• Knowledge represented symbolically by
• Yes/No (1/0)
• b) Knowledge Base Symptom table
• c) Patient STATE VECTOR (PAT)
• d) Sequential comparison
• e) Diagnoses list (sorted)

8
Ddisease, SsymptomPATpatient state vector
S1 S2 S3 ......Score
D1 1 0 1 2/8 D2 0
1 1 .... 3/6 .... .........................
.................... PAT 0 1
0
9
2.2. Types of logical methods

• According to PAC vector construction
• A) Symptom tables (truth tables)
• Symptom selection from a menu
• B) Decision trees
• Set of questions with Y/N answers
• Avoiding useless questions
• Patient involvement

10
2.3. DISADVANTAGES OF LOGICAL METHODS

• CANNOT QUANTIFY SYMPTOM INTENSITY (ex
  high/moderate fever)
• CONSIDER ALL SYMPTOMS AS EQUALLY WEIGHTED FOR
  DIAGNOSIS
• SOME SYMPTOMS MIGHT NOT BE PRESENT
• DISREGARD DISEASE PREVALENCE

11
3. STATISTICAL METHODS

• 3.1. BAYES RULE assumes as known
• p(D) Probability of disease D within a
  population (disease probability prevalence)
• p(S/D) Probability of symptom S to occur IF
  the disease D is present
• p(S) probability to encounter symptom S,
  which may be computed or estimated from p(S) and
  p(S/D)

12
b) For each pair D/S (Disease / Symptom) we build
a 2x2 table
S S -
D n11 n12 R1 D - n21 n22 R2
C1 C2 N
13
c) PROBABILITIES- unconditional p ( D) R1 /
N probability of disease D to be present-
conditional p ( S / D- ) n21 /
R2 probability of symptom S to be present if
disease D is absentd) BAYES RULE
P(S/B) x P(B)
P(B/S)
P(S)
14

• e) Importance
• - possibility to compute p(D / S) without the
  table
• f) For several symptoms
• - it can be applied only for INDEPENDENT
  symptoms
• independence test (chi-square)
• g) Application
• P(S/D) n11/R1
• P(D) R1/N
• P(S) C1/N
• gt P(D/S) n11/C1

15
Example

• In a study 3000 medical records are analyzed. 500
  patients had virosis and 400 of them presented
  fever. But fever has been present also in other
  600 patients. Calculate
• A) probability that a patient had virosis
• B) did not have fever
• C) had fever if did not have virosis
• D) had virosis if had fever

16
3.2. PATTERN RECOGNITION

• Establishing a diagnosis as a RECOGNITION process
• Notion of PATTERN
• NAME OF A CLASS DISTINGUISHED BY A SET OF
  CHARACTERISTIC FEATURES
• Discriminant power of various features
• Process of RECOGNITION

17
(No Transcript)
18
c) CLASSIFICATION METHOD

• a set of classified objects is given (with their
  numerical characteristic features)
• a representation in a multidimensional space is
  made and each class is delimited
• Question to which class does a new object belong
  
• the new object is classified according to its
  position
• two working phases
• learning (supervised)
• classification
• advantage similarity with real situations

19
d) CLUSTERING METHOD

• a (large) set of unclassified objects is given
• an n-dimensional graphical representation is
  performed
• Question can these objects be divided into
  different classes ?
• two phases
• learning (unsupervised)
• classification (cluster defining)

20
(No Transcript)
21
e) BUILDING A PATTERN

• FEATURE SELECTION
• class delimitation (disjunct classes)
• projection function
• methods - vectorial
• principal components analysis
• discriminant analysis
• - structural
• feature hierarchy
• CLASSIFIER SYNTHESIS (decision function)
• geometrical / statistical / syntactic rules

22
E n d