Medizinische Experten und wissensbasierte Systeme

1
Fuzzy Diagnostic and Therapeutic Decision
Support Klaus-Peter Adlassnig Department of
Medical Computer Sciences University of Vienna
Medical School
2
Objectives

• Support of medical decision making for the
  single patient
• provide a correct diagnosis
• selection of an optimal therapy
• correct assessment of prognosis
• optimal patients management in medical
  institution

3
Knowledge-Based Methodology

• knowledge level
• modelling of mental processes linguistically
  communicated
• modelling based on symbols (linguistic concepts
  abstract concepts)
• objective and subjective knowledge (definitional,
  causal, statistical, and heuristic knowledge)
• measurements and observational level
• measured and observed data
• data-to-symbol conversion

4
patients medical data
physicians medical knowledge
induction
symptoms signs test results biosignals images 3D-v
isualizations diagnoses therapies standardiz
ation telecommunication chip cards
anatomy physiology pathophysiology pathology nosol
ogy therapeutic knowledge (e.g.,
pharmacology) subjective experience intuitio
n
? ? ?
many patients
general knowledge
deduction
single patient
general knowledge


HIS, MIS, LIS, PDMS, RIS, PACS
expert and knowledge-based systems
telemedicine
5
(No Transcript)
6
differential diagnosis
7
(No Transcript)
8
(No Transcript)
9
Fuzziness in Medicine

• vagueness of medical concepts
• gradual transition from one concept to another
• uncertainty of medical conclusions
• uncertainty of co-occurrence of vague medical
  concepts
• incompleteness of medical data and medical theory
• partially known data and partially known
  explanations of medical phenomena

10
FuzzyKBWeanA Fuzzy Control System forWeaning
from Artificial Ventilation

• C. Schuh, M. Hiesmayr, K.-P. Adlassnig
• Department of Medical Computer Sciences
• Department of Cardiothoracic and Vascular
  Anaesthesia and Intensive Care
• University of Vienna Medical School and Vienna
  General Hospital

11
Objective

• mechanically ventilated patients after
  cardiothoracic surgery in an Intensive Care Unit
  (ICU)
• proposals for changes of the ventilator settings
  during the three phases of mechanical ventilation
  (stabilization, weaning, and finally extubation
  of the patient)
• open-loop system integration into the patient
  data management system (PDMS) time resolution
  1 minute
• closed-loop system as a long-term objective
  integration into the ventilator (auto-mode)

12
(No Transcript)
13
(No Transcript)
14
Structure of FuzzyKBWean
15
Methods

• phase-dependent fuzzy sets
• linguistic If/Then rules
• If patients physiological parameters and
  ventilator measurement parameters (in a defined
  context)
• Then proposals for changes of ventilator
  settings
• fuzzification step
• arithmetic, statistical, comparative, logical,
  temporal, and control operators
• defuzzification step
• center of gravity method
• verification by the attending physician, i.e.,
  open-loop

16
(No Transcript)
17
(No Transcript)
18
Fuzzy Control
inference process with linguistic fuzzy rules

defuzzification with center of gravity method
fuzzification with fuzzy sets
measurement and observational level
PaO2 mmHg PaCO2 mmHg
FiO2-change
19
Knowledge Base
fuzzy sets
PaO2
1

very low
low
normal
high
75
85
90
95
100
70
mmHg
0
PaCO2
low
high
1
very low
normal
very high
35
30
40
45
50
55
0
mmHg
linguistic fuzzy rules
rule 1 If PaO2 very low and PaCO2 low then
FiO2-change 10 rule 2 If PaO2 normal and
PaCO2 normal then FiO2-change 5 rule 3 If
PaO2 high and PaCO2 normal then FiO2-change
10 etc.

20
PATIENT
PaO2 96 mmHg, PaCO2 42 mmHg
fuzzification
PaCO2
96
low
high
1
very low
normal
very high
35
30
40
45
50
55
0
mmHg
42
inference process
rule 2 If PaO2 normal and PaCO2 normal then
FiO2-change -5 rule 3 If PaO2 high and PaCO2
normal then FiO2-change -10 rule 2 min (PaO2
normal, PaCO2 normal) min (0.8, 1)
0.8 rule 3 min (PaO2 high, PaCO2 normal)
min (0.2, 1) 0.2

defuzzification
21
(No Transcript)
22
(No Transcript)
23
Results 1

• 23 variables
• 74 fuzzy sets (phase-dependent)
• 16 If/Then rules
• 4 rules checking for measurement errors and
  validity
• 3 rules for ventilation (normal range,
  hypoventilation, hyperventilation)
• 4 rules for oxygenation (stabilization,
  oxygenation normal, hypoxia, severe hypoxia)
• 4 rules for intermediate states (increased EtCO2,
  decreased EtCO2, phase changes)
• 1 rule for extubation

24
Weaning Rule Wean_1

• If
• mean EtCO2 during the last 30 minutes
• is contained in fuzzy set
  EtCO2_wean_normal
• and if
• rule Wean_1
• has not been activated
• during the last 30 minutes
• and if
• EtCO2 is valid
• and if
• EtCO2 is normal
• Then
• PIP -3

25
Results 2

• 10 prospectively randomized patients
• FuzzyKBWean reacted correctly 131 (SEM 47)
  minutes earlier than the attending physician
• adjustment of ventilation parameters was more
  reliable than adjustment of oxygenation (EtCO2 is
  more reliable as SpO2)
• phase-specific rules often proposed too small
  changes of the ventilator settings
• temporal rule blocking, fuzzy set adaptations,
  employing thresholds to avoid oscillations

26
Results 3Delay of Staff Reaction in Case of
Hyperventilation
27
Discussion

• methodology
• minimal number of therapeutically significant
  classes per variable
• gradual transition between variable classes
• adequat consideration of inherent vagueness of
  medical concepts
• intuitive If/Then rules on the knowledge level
• physicians medical knowledge was transfered to
  FuzzyKBWean
• clinical trial
• periods of deviation from the target parameters
  are shorter
• contribution to patients safety and comfort
• closed-loop recognition of artifacts and
  information obtained by direct observation of the
  patient

28
CADIAG-IIA Hospital-Based Consultation System
forInternal Medicine

• Klaus-Peter Adlassnig, G. Kolarz
• Department of Medical Computer Sciences
• University of Vienna Medical School

29
Objectives

• diagnostic hypotheses, confirmed and excluded
  diagnoses
• indication of rare diseases
• proposals for further examinations
• ranked according to invasiveness and costliness
• pathological findings not yet accounted for
• search for further diagnoses
• correct and complete differential diagnoses
• at minimal risk for the patient and costs for the
  health care system

30
(No Transcript)
31
CADIAG-II Methods

• patient data
• symptom, signs, and test results are modelled as
  context dependent fuzzy sets
• diseases or diagnoses takes values in 0,1
• knowledge representation
• symptom and disease hierarchies
• crisp rule and fuzzy set based data-to-symbol
  conversion
• symptom/disease, symptom/symptom, disease/disease
  relationships, and complex diagnostic rules with
  frequency of occurrence and strength of
  confirmation
• inference mechanism
• manifold application of the compositional rule of
  fuzzy inference

32
glucose level in serum
?(x)
highly reduced ??
highly elevated ??
reduced ?
normal ?
elevated ?
1.00
µ?(x) 0.82
0.50
µ? (x) 0.17 µ?? (x) 0.00 µ? (x)
0.00 µ?? (x) 0.00
50
?0.00?
100
150
x
200 mg/dl
130 mg/dl
33
(No Transcript)
34
(No Transcript)
35
(No Transcript)
36
(No Transcript)
37
(No Transcript)
38
Examples

• Example 1 (indicating)
• IF elevated amylase level in serum
• THEN acute pancreatitis
• WITH (?O very often ?O 0.90, ?B strong
  ?C 0.70).
• Example 2 (necessary and sufficient)
• IF rheumatoid arthritis and
• splenomegaly and
• leukopenia less than 4 giga/l
• THEN Feltys Syndrom
• WITH (?O always ?O 1.00, ?C confirming
  ?C 1.00).

39
CADIAG-IIs Relationships
frequency of occurrence
strength of confirmation
40
Inference Mechanism
Compositional rule of inference
(0.8)
D3
1.0
1.0
D2
D1
(0.8)
(0.6)
0.4
0.6
0.8
0.4
0.8
and
S7
S3
S2
S6
S4
S5
(0.8)
1.0
0.9
1.0
0.0
1.0
1.0
S1
threshold for hypothesis generation
0.8
41
1
6

4
disease-to-disease inferences
finding-to-finding inferences
finding-to-disease inferences

5
diagnostic criteria-to-disease inferences
2
evaluation of intermediate criteria
3
evaluation of diagnostic criteria
inter- mediate criteria
diagnostic criteria
findings
diagnoses
42
(No Transcript)
43
(No Transcript)
44
(No Transcript)
45
(No Transcript)
46
(No Transcript)
47
(No Transcript)
48
(No Transcript)
49
(No Transcript)
50
(No Transcript)
51
Results

• Rheumatology
• more than 200 disease profiles, more than 2.000
  findings
• more than 50.000 finding-disease-relationships
• more than 160 complex rules
• Hepatology and Gastroenterology
• more than 100 disease profiles, more than 1.000
  findings
• more than 30.000 symptom-disease-relationships
• more than 40 complex rules

52
Evaluation
53
MedFrame/CADIAG-IVA Consultation System
Framework for Internal Medicine and Related Areas

• Klaus-Peter Adlassnig
• Department of Medical Computer Sciences
• University of Vienna Medical School

54
Objectives

• CADIAG-IV
• positive and negative diagnostic hypotheses,
  confirmed and excluded diagnoses
• positive and negative therapy proposals,
  necessary and excluded therapies
• MedFrame
• shell for medical knowledge-based systems
• knowledge-based telemedicine service

55
MedFrame/CADIAG-IVMethods

• extension of CADIAG-II
• symptoms, diseases, and therapies
• context-sensitive data-to-symbol conversion and
  patient specific adaptation of the knowledge base
• extended knowledge representation with
  step-by-step knowledge acquisition refinement
• MedFrame
• client/server architecture, WWW compatible
• integrated patient data and medical knowledge base

56
MedFrame Structure
57
(No Transcript)
58
(No Transcript)
59
Symptom-Disease Relationships
60
(No Transcript)
61
MedFrame/CADIAG-IVDiscussion

• mainframe CADIAG-II to Medframe/CADIAG-IV
• integrated patient data and medical knowledge
  base
• patient data and knowledge transfer
• theoretical extensions
• extension of the relationship, rule, and
  inference concept
• medical extensions
• rheumatology, hepatology, gastroenterology,
  radiology, neurology
• HEPAXPERT, TOXOPERT in MedFrame

62
Conclusions

• knowledge-based systems are becoming part of
  medical practice
• computational intelligence in medicine
• vagueness, uncertainty, and incompleteness of
  medical data and medical knowledge demand a
  flexible and extended formal framework
• medical knowledge representation and inference
• fuzzy set theory and fuzzy logic provide an
  appropriate solution