A three-step approach for STULONG database analysis: characterization of patients

1
A three-step approach for STULONG database
analysis characterization of patients groups
Discovery Challenge (PKDD 2004)

• O. Couturier, H. Delalin, H. Fu, E. Kouamou, E.
  Mephu Nguifo
• Computer Science Research Center of Lens (CRIL)
• CNRS - Université dArtois IUT de Lens

2
Goal

• What are the relations between social factors
  (social characteristics) and the other
  characteristics of men in the respective groups?

3
Overview

• Discovery process
• Techniques and results
• Clustering
• Classification
• Association rules
• Conclusion and further work

4
Discovery Process

• Hypothesis on data
• ENTRY table
• Groups provided by expert
• Merging groups 1 and 2 Normal group
• Merging groups 3 and 4 Risk group
• Ignoring group 6
• Characteristics
• Considering previous work of LRI ML research team
  at previous PKDD Challenges

5
Discovery Process

• Can we find a model that fits with the provided
  groups ?
• Are there strong similarities among instances of
  different groups ?
• Which kind of relations exist among group
  characteritics ?

6
Discovery Process
Data Tasks Knowledge

Clustering Generated clusters vs provided ones

Entry data groups Supervised classification Similarities among instances, and groups

Association rules search Affinity among groups characteristics
7
Techniques and Results Clustering

• Goal do the initials groups can be considered as
  they were defined?
• Data groups 12, 34 and 5
• Clustering systems (WEKA package)
• COBWEB 2 groups
• EM 4 groups
• KMEANS 2 groups
• Results difficulty to identify properties which
  allow to retrieve the initial groups

8
Techniques and Results Supervised Classification

• Risk group patients similar to those in Normal or
  Pathological group ?
• Data
• Training set group 12 and group 5
• Test set group 34 (Risk)
• System (WEKA package)
• Decision tree C4.5

9
Techniques and Results Supervised Classification

• Training results
• HT descriptor are one of the most relevant
  factors of the disease
• Thirdteen instances of Pathological group are
  classified as Normal

Confusion Matrix a e
lt--classified as 276 0 a 12 13 101
e 5
10
Techniques and Results Supervised Classification

• Test set Risk Group34
• Health district number is not a relevant factor
• 2/5 Risk patients similar to Normal group patients

Confusion Matrix a c d e lt--
classified as 0 0 0 0 a 12 177
0 0 250 c 3 (odd) 197 0 0 235 d
4 (even) 0 0 0 0 e 5
11
Techniques and Results Association rules search

• Goal Find relations that exist among group
  characteritics
• Data 1417 patients of groups 12, 34 and 5
• System
• Apriori B. Goethals implementation
• Preprocess
• Binary conversion of the 27 characteristics
• Frequent Itemsets Search
• Results
• Frequent itemsets common to different groups

12
Techniques and Results Association rules search

• Preprocessing Binary conversion
• BMI weight / size² (m)
• If bmi gt 27 then 1 else 0
• Age
• If age gt 45 then 1 else 0
• Smoker
• If smokerconsumption!0 OR duration then 1 else 0

13
Techniques and Results Association rules search

• Pre-processing Binary conversion
• Bolhr (chest pain)
• If bolhr1 or bolhr6 then 0 else 1
• Chol
• If (chol gt 2(age/100)) then 1 else 0
• Tg
• If tglt150 then 0 else 1

14
Techniques and Results Association rules search

• Frequent itemsets search
• Support threshold (Minsup) 0.10
• significant for at least 10 of the population
• Search was done with no MinSup (i.e MinSup value
  0)

Itemsets Itemsets Itemsets
Class 12 Class 34 Class 5
Support value Support value Support value
15
(No Transcript)
16
Techniques and Results Association rules search

• Frequent itemsets search Results
• Support value of Alcohol attribute 1
• 1-itemsets
• Attribute IM is false for each patient of group
  12 and 34. The value is true for 33 of patients
  of group 5.
• HT is false for each patient of group 12.
• STUDY is more frequent in group12 than in group5
• 3-itemsets
• AGE SMOKER CHOL is less frequent in group12
  than in group5
• etc
• SupportValue Group 34 is between SupportValue
  Group 12 and SupportValue of Group 5.

17
Conclusion

• RG similarity with NG and PG.
• 3 steps
• Clustering initial groups are not found
• Classification some attributes which
  characterize the pathological group but already
  known
• Frequent itemsets search difficult to highlight
  concrete results but interesting informations

18
Further work

• Upgrade the binary conversion
• Refining the data set on the population
• for instance, 12 patients died because of
  atherosclerosis while they were in the NG.
• Refining our hypothesis
• Data set of ENTRY table
• Look at the CONTROL table

19
Thanks !