- The document describes a study that constructed a disease network using clinical data from claims records. - Researchers extracted disease-disease pairs from patient records to represent relationships between previous and later diagnosed diseases. - The resulting disease network had 775 nodes and over 4,000 edges, with edges colored to represent sex-dominant relationships. - Analysis found the network followed a scale-free structure and that degree of older patients and diseases were higher. Communities were identified. - The disease network was compared to a gene-disease network, finding some novel relationships not previously identified.

1. 대한생물정신의학회 2018 춘계 학술대회
Construction and Interpretation of
Disease Network using Clinical Data
차의과학대학교 의학전문대학원 정보의학교실
한현욱 (M.D, Ph.D)
stepano7@gmail.com

2. C.V
 Position
 Head Professor, Dept. of Biomedical Informatics, CHA University
 Head, Healthcare Big Data Laboratory, CHA University (CHABI)
 Scientific Program Committee Member, Translational Bioinformatics Conference
 Instructor, GDA and CPBMI, KOSMI
 Advisor Medibloc, Syntecabio, and MISOinfo
 Education
 CPBMI, Certified Physician in BioMedical Informatics, KOSMI
 Ph.D, Biomedical Informatics, Graduate School of Medicine, CHA University
 M.D, Gradate School of Medicine, CHA University
 M.S, Dept. of Electrical & Computer Engineering, Seoul National University
 B.S, Dept. of Electrical & Computer Engineering, Hanyang University
 Work
 Research Assistant Professor, Dept. of Biomedical Informatics, Ajou University
 Clinical Assistant Professor, Dept. of Preventive Medicine, CHA University
 Research Scientist, Systems Biomedical Informatics Research Center, SNU
 R&D Engineer, LG Electronics R&D Campus

3. Book
베스트
셀러

4. Network Science
초연결
초지능

5. Network Science
나 오바마
삼라
만성

6. Network in biology
Pathway Network

7. Biological Network
Protein Interaction
Network
Network biology
Network Medicine
Regulatory
Network
Metabolic
Network
Cell Signaling
Network
Perturbation
Sensitivity Network
Social Network
Disease Network
Drug Target
Network

8. Network Science
 20C Network (Complex System) Science
 Opposing to Reductionism
 Barabasi et al.
 Network Topologies
Random Network Scale-free Network

9. Characterization of Real Network
P(k) ∝ λ-γ (2 < γ < 3)
Hub node vs. peripheral node
Evolution
Small world phenomenon
Six degrees of separation
The 80/20 rule
Rich get richer
Achilles' Heel
 Scale-free network
 Network growth
 Preferential attachment
Social Network
Internet Network
Biological Network
log-log

10. Previous Researches of Biological Network
PPI is scale-free
(Barabasi, 2000)
Centrality & Lethality
rule (Jeong, 2001)
Hubs evolve slowly
(Fraser, 2002)
Disease : clustering,
tissue specificity &
Periphery (Goh, 2007)
Drug target : similar
with disease genes
(Yildirim, 2007)

11. Disease Network

12. Data source for Construction of Disease Network
Disease A – Gene 1
Disease B – Gene 1
 Disease A – Gene 1 -- Disease B
 Disease A- Disease B

13. Previous Disease Network
 Human Genes – Disease network (Koh et al. 2007,
PNAS) : Gene Base

14. Previous Disease Network
 Human Symptoms – Disease Network (Barabasi et al. 2014,
Nature Communication) : EMR Base
Disease A – Symptom 1
Disease B – Symptom 1
 Disease A – Symptom 1 -- Disease B
 Disease A- Disease B

15. Previous Disease Network
 Disease Comorbidity Network (Yang Chen, 2015, AMIA)
상병명 : Disease A ,Disease B
 Disease A- Disease B

16. Motivation
 Definition of Disease (Wiki)
 질병(疾病)이란 유기체의 신체적 기능이 비정상적으로 된 상태를
일컫는다.
 인간에게 있어서 질병이란 넓은 의미에서는 극도의 고통을
비롯해 스트레스, 사회적인 문제, 신체기관의 기능
장애와 죽음에까지를 포괄한다.
 질병이란 개인만에 한정되는 것이 아니어서 사회적으로 큰
맥락에서 이해되기도 한다.
 더 넓게는 사고나 장애, 증후군, 감염, 행동 장애 등을 모두 나타낼
수 있다.
 질병의 종류에는 약 30000가지 정도가 있다고 한다.

17. Motivation
 Genetic Disease Vs. Non-Genetic Disease
 Many diseases have no genetic basis at all. Usually a physical
injury such as a bone fracture is not caused by genetics when it
is caused by something else. (Of course, there could still be an
underlying genetic cause of weak bones or osteoporosis that
really caused the fracture.) Similarly, a virus or bacterial
infection is caused by an external microbe
 예방의학교과서
 유전체로 설명될 수 있는 질환은 20~30% 정도로 알려져 있음
Most diseases are non-genetic

18. Motivation
The Same Symptom – Different Disease
DDx is very important!!

19. Motivation
 Sex
Sex is one of the clinical factors of disease

20. Motivation
 Age
Age is a important clinical factor of disease

21. Motivation
History
- Disease
- Drug
- Family
- Socio-economy
- Operation
- Mensturation (F)
A disease is risk factor of another disease

22. Motivation
 Causality and Risk
Clinically, Disease Network is Directed and Weighted
D1
D2
D4
D5
D3
D1
D2
D4
D5
D3
60
10
20
12
7
32
Non-directed and non- weighted Directed and Weighted

23. Motivation
 Network Visualization
A clinician likes incidence-based node Presentation
D1
D2
D4
D5
D3
D1
D2
D4
D5
D3
60
10
20
12
7
32
Centrality-based Node Presentation Incidence-based Node Presentation

24. Motivation
 Navigation of Disease Network
Help for a Clinician and a Cheat Sheet for Research

25. Problem of Clinical Data
Data
Fragmentation

26. Claim Data

27. NHIS Cohort Data
A B C
A B, BC, AC
Sex, Age, Direction, Weight, Duration

28. Method
• Assumption
• we extracted all combinations of disease-disease pairs
from each patient’s transition based on the assumption
that “previous incidence would become a cause (or act as a
risk factor) of later incidence” .
• Step 1. 3 types of event sequence:
• Incidence sequence
• Ex: A  B  C  D
• Step 2. Extracting disease-disease pairs:
• Disease-disease pairs : (A  B), (A  C), (A  D), (B  C),
(B  D), (C  D)
• Step 3. Record the frequency of each disease-
disease pair

29. Method
Statistical analysis and constructions of the human disease directed network. A: databases of the sample cohort data. B: 5 variables from
the sample cohort data and the disease codes selection process. C: Extracting the disease-disease relationships and the frequency records
in the Dn x Dm table

30. Method
Method – Fisher exact test
• Cutoff
• False discovery rate(FDR) p < 0.001
• Relative risk > 4

31. Result
 Network Construction
775 node
4,100 edge
Edge color - Blue : male-dominant – 329
Red : female-dominant – 1,868
Green: not sex-dominant – 3,539

32. Results
0-20 20-40
40-60>60
Sex and Age are important factors in determining the structural dynamics of
disease networks.
 Network Construction

33. Results
 Degree Analysis
• DPN is a typical scale-free network
• The positive correlation between
the in- and out-degrees of diagnoses

34. Results
• ICD-10 categories are determined according to in- and out-degrees of diagnoses
• Old age has high in- and out-degree
 Degree Analysis

35. Results
• Community detection in the DPN formed
88 clusters.
• Six giant cluster had distinct properties
(Sex, Aage, ICD-10 categories)
 Community Analysis

36. Results
• Long-term disease-pair association within 10%
 Duration

37. Results
 DPN vs. gene-disease network (GDN) of Cancer
 29 disease categories (DPN : 27 pair, GDN : 261 pair)
 14 of 27 links In the DPN overlapped with the links in the GDN
 What is the 13 links?
Example of 13 links Evidence
“Colon cancer” (C18) and
“Prostate cancer” (C61)
Fitzgibbons, R. J. Jr., Lynch, H. T. &
Salerno, G. M. Hereditary colon
cancer syndromes
“Bladder cancer” (C67) and “Lung
cancer” (C34)
Kantor, A. F. & McLaughlin, J. K.
Second cancer following cancer of
the urinary system in Connecticut,
1935–82. Natl Cancer Inst Monogr
68, 149–159 (1985)

38. Results
 Practical Usability
• Supporting Network visualization according to Sex, Age, Disease-pair, Duration,
Relative Risk and Directionality

39. Results
Subnetwork
 Usability

40. Conclusion
 We built a directional weighted network with duration information using
claim data
 We showed that our network had both in-degree and out-degree
distributions following a power law (Scale-free Network)
 Older patients are more likely to have been exposed to various diseases
 Disease Network is grouped by gender, age, and ICD-10 categories.
 Our network presented clinically meaningful connectivity and also identified
connectivity that were not previously found in the gene-disease network
(the macroscopic level, such as the metastasis of cancer)
 The network presented here may potentially serve as a predictive tool for
the diagnosis and treatment of diseases

41. Paper

42. 대한생물정신의학회
경청해 주셔서 감사합니다.