�ݺ�ߣ

By
Amos Otieno Olwendo
Thesis Research (May 2014),
Masters of Science(MSc) in Medical Informatics
06/10/15 1

.1 1INTRODUCTION
 WHO lists COPD as the 4th
leading cause of death
worldwide
 90% of COPD mortalities experiences in middle and
low-income countries
 COPD diagnosis is prone to
 under-diagnosis and
 misdiagnosis (reported in UK, Australia, Canada e.t.c)
 Known risk factors:-
 Smoking of tobacco,
 anti-1-trypsin (A1At), and
 air pollution are the known major risk factors

.1 2Previous Studies?
 Previous studies have focused on using PFTs
 Identify COPD phenotypes using variables
 classify COPD cases based on severity (Stage 1 - 4) with
Stage 1 being Mild and Stage 4 Very Severe (Figure 2.1)
 Use methods ;data-driven phenotyping techniques
such as:-
 Cluster analysis
 PCA
 Factor analysis
 Discriminant analysis to define COPD phenotypes
06/10/15 Amos Otieno Olwendo 3

.1 3What’s new?
 Our model attempts NOT to use PFT
 Phenotyping is related to disease classification
 classifies COPD phenotypes based on
 Morphology (appearance)
 Function
 Behavior
 We use a Bayesian Network
 We achieved a classification of 98.75% on the test data
set

Phenotype?

.1 5Phenotypes in action?

.1 6Research Scope
1. Determine the essential variables and parameters
2. Design the probabilistic model used in this research
3. Determine whether a given patient case has COPD
4. Identify the consequent COPD Phenotype
4.1 Emphysema
4.2 Chronic bronchitis
4.3 General COPD (amalgamation of bronchitis and
emphysema)
4.4 Asthmatic COPD (amalgamation of asthma and any
other phenotype(s))
5. Ascertain whether the given patient has Asthma
6. Severity ; NEXT as in Figure 2.1
7. Determine cause-effect relationships among variables
7

.2 4Spirometry & Barriers
 FEV1 decline predicts the future of the patient
 Equipment and training costs
 Low confidence in the use and
 interpretation of the results
 Perceived lack of utility
 Quality assurance issues
 Physical demand from the patient to use the
spirometer – esp. by the elderly and those
experiencing respiratory challenges
11

.2 5What is Modeling?
research technique that
connects empiricism to theory,
and
experiments to theory construction
and validation
Here: BN used as the knowledge base
laws of probability theory as the
reasoning engine

.2 6Why PGMs?
1. Ability to handle vagueness as a result of:
i. Biased or incomplete understanding of the event at
hand
ii. World of Noisy observations
iii. Phenomena not represented
iv. Randomness of events in real-life
1. Human reasoning -based on facts and assumptions
2. Probabilistically – degrees of belief are adjustable
based on evidence
3. Intuitive with a compact data structure
13

2.7 Why BNs?
14
 Representation: a directed acyclic graph (DAG)
 Composed of random variables X1, …, Xn organized as
 Query,
 Non-query, and
 Evidence variables
 Each variable/factor has a corresponding CPT
 Parent-child relationships of variables are represented
as CPDs [ P(X1, …, Xn) ]
 Inference: exact and approximate
 Learning: both parameters & structure, with complete
or incomplete data (through MLE)
 Employs the use of Chain rule for BN

Knowledge Engineering
 Knowledge acquisition
-elicitation,
-collection,
-analysis,
-modeling, and
-validation
 Knowledge representation and
 Reasoning
Types of CPDs
Noisy-Or CPD– common for medical diagnosis applications
 Sigmoid CPD – best design approach (personal opinion) 15

.2 9Circuit design and Noisy-OR

.3 1METHODOLOGY
 Non-interventional(Observational)
 retrospective study [Experimental study]
 Conducted at Loghman Hakim:
 Heart &Lung Division
 Tehran- a city with high levels of air pollution
(especially during winter )
17

.3 2Methodology
 This study was conducted from
 August 2013 to January 2014
 This unit receives approximately
 420 COPD patients and
 4200 Asthma patients monthly
18

.3 3Methodology
 Sample Size: 100 COPD + 100 Asthma
 The environment composed of
 Dr. Agin,
 2 resident physicians (worked with Dr. Agin), and
 2 nurses (1 translator + Dr. Agin-Patient contact)
 Amos
 conducted a structured interviews and
 data was recorded using
 a structured checklist
19

.3 4Initial BN Design

.3 5Checklist
 The checklist had 10 questions
 Parameter measures were conducted through
 self report and/or
 Observation
 Checklist design
 patients had to commit to their parameter choices by
choosing a number between 0 and 10
 Each interview result was cross -checked with the
reference standards
21

3.6 COPD & Asthma Diagnosis
 Patient History
 History of present illness
 Past medical history
 Family history of COPD and Asthma
 social history of the patient (exposure to irritants)
 Physical Exam
 Review of systems
 Visual examination (include palpation and percussion)
 listing to the lungs (stethoscope),
 physical activity
22

3.7 COPD & Asthma Diagnosis
 Pulmonary Function Test(PFT)s
 e.g. spirometry /
 Bronchodilators (Nebulizer)
 X-ray – if necessary
 Vital signs
 Examine O2 and CO2 in the blood (pulse oximeter )
 Blood pressure (sphygmomanometer exam.)
23

.3 8Data Collection…design

06/10/15 25Amos Otieno Olwendo

.3 10Neural Network setup

.3 11Data Analysis
 Primary tool: the Bayesian network
 Model Validation: NN based on LM algorithm
 The dataset was divided into
 60% for training
 40% test
 To ensure an even distribution and representation,
 we grouped cases based on phenotypes (per group:
target and control) then
 assigned identifications to case then
 Through simple random sampling, we determined what
cases to be used for training and testing respectively
27

.3 12Data Analysis
 Developed a C++ application –
 through cases analysis,
 assigns a real number between negative and positive
infinity to each patient case (using MLE)
 loaded these results to SQL Server and
 R Statistical software to obtain graphical outputs
28

.3 13Reliability Analysis: SPSS

.3 14Reliability Analysis: SPSS

.4 1RESULTS: BN
Bayesian Network Classification of 40 COPD Test Cases
COPD Phenotype Number of Cases
Asthma 1
Asthmatic COPD 5
Chronic bronchitis 13
General COPD 21

.4 2RESULTS: BN
Bayesian Network Classification of 40 Asthma Test Cases
Classification Number of Cases
Asthma 34
Asthmatic COPD 6
32

.4 3RESULTS: NN
Levenberg-Marquardt Algorithm Classification of 40 COPD Test Cases
COPD Phenotype Number of Cases
Asthma 2
Asthmatic COPD 2
Chronic bronchitis 25
General COPD 10
None 1
33

.4 4RESULTS: NN
Levenberg-Marquardt Classification of 40 Asthma Test Cases
Classification Number of Cases
Asthma 34
Asthmatic COPD 6
34

4.5 RESULTS: Summary
35
Category Bayesian Network
Percentage (%)
Classification of the
Test Data Set
Levenberg-Marquardt
Algorithm)
Percentage (%)
Classification of the
Test Data Set
COPD 97.50 92.50
Asthma 100 100
Overall 98.75 96.25

.4 7RESULTS: C++(group plot(
37

.5 1DISCUSSION
1. COPD burden worldwide is underestimated (could
be worse than it is)
2. COPD under-diagnosis and/or misdiagnosis should
not pose the challenges it currently does to clinicians
3. Increasing cases of COPD could be as
3.1 a result of the changes in some social behaviors that
3.2 affect COPD development and progression
3.3 Such behavior may include:
3.3.1 increasing number of female smokers and
3.3.2 increasing number of teenage smokers
3. Worst hit populations are in middle to low-income
countries (inadequate healthcare services)
39

.5 2SUGGESTIONS
1. Increased COPD Awareness at the community level
1.1 Anti-smoking campaigns
1.2 Reduced exposure to 2nd
hand cigarette smoke (creation
of designated smoking areas)
1.3 Cooking using firewood/cow dung in less ventilated
environments
1.4 Air obstruction symptoms
1.5 Legal measures- who can smoke and or buy cigarettes
40

.5 3SUGGESTIONS
2. Population-based screening (“Target Case Finding”)
2.1 whenever an individual shows up to a health care
worker
2.2 Maybe useful in identifying those at risk
2. Need for screening devices since
3.1 certain localities lack specialist and/or
3.2 equipment (PFT devices, other test materials like
bronchodilators, X-ray machines, maybe computers
and or internet)
41

.5 4SUGGESTIONS : Screening Criteria
42

End! Thank you
06/10/15 43Amos Otieno Olwendo

�ݺ�ߣ

v2 3rd (11-13 June 2015) KNH and UON Conference-Research as a Driver for Science & Technology Innovation for Heath

More Related Content

v2 3rd (11-13 June 2015) KNH and UON Conference-Research as a Driver for Science & Technology Innovation for Heath

Editor's Notes