Temporal Analysis of Platelet Data in Chronic Viral Hepatitis Dataset

1
Temporal Analysis of Platelet Data in Chronic
Viral Hepatitis Dataset

• Shoji Hirano Shusaku Tsumoto
• hirano_at_ieee.org tsumoto_at_computer.org
• Department of Medical Informatics, Shimane
  University, School of Medicine, Japan

2
Outline

• Introduction
• Cluster analysis system for time-series medical
  data
• Experimental results
• 1. Cluster analysis of platelet (PLT) sequences
• 2. PLT-based analysis of the progress speed of
  liver fibrosis
• Conclusion Future Work

3
Introduction

• Platelet (PLT) decrease in chronic viral
  hepatitis
• The liver plays an important role in generating
  platelets
• Bleeding is often observed on the patients at the
  terminal stage of liver cirrhosis
• Liver fibrosis -gt dysfunction -gt PLT decrease -gt
  bleeding
• Use of PLT count as an noninvasive index for
  evaluating liver fibrosis F0 (no fibrosis), F1,
  F2, F3, F4 (liver cirrhosis)
• PLT counts correlate with fibrotic stage
  (Matsuura et al., J. Viral Hepat. 2000),
  according to PLT counts at the time of biopsy
• However,
• Time-series analysis of the PLT data have rarely
  been performed
• Features of temporal patterns/courses have not
  been examined

4
Introduction (contd)

• Difficulties in time series comparison
• Irregular sampling intervals, irregular sequence
  length
• Existence of events with different length and
  phase

Difficult to find the corresponding pairs of
points be compared
4 years
9 years
3-6M
1-2 Weeks
IFN 6m
5
Cluster Analysis System
Time-series Medical Data
Multiscale Comparison
Cluster representation of time series
Grouping
6
Experiment 1 Cluster Analysis of PLT Sequences

• Objective
• Discovery of common temporal courses in platelet
  sequences
• Discovery of relations between platelet courses
  and fibrotic stages
• Used data
• Chronic hepatitis dataset provided by Chiba
  University
• Total 488 cases (Type B 193, Type C w/o IFN
  99, Type C with IFN 196) (Male 343, Female 145
  (14548, 6138, 13759))
• Procedure
• Generate a dissimilarity matrix using the
  modified multi-scale matching
• Generate a dendrogram from the matrix using
  agglomerative hierarchical clustering (average
  linkage) and perform cluster analysis

7
Experiment 1Cluster Constitution w.r.t. Fibrotic
Stages

• Type C, IFN (total 196 cases)

? clusters of N lt 3 omitted
F4 Major
F1 Major
8
Experiment 1Grouped Sequences
Number of cases (F0/F1/F2/F3/F4)

• F4, F3 major clusters (Type C, IFN)

Cluster 5 N11 (0/2/1/2/6)
Cluster 8 N40 (0/9/6/9/16)
0
5
10
15Y
Normal L (120 x103 /ul)
Normal H (350 x103 /ul)
?first 16 cases
9
Experiment 1Grouped Sequences

• F1, F2 major clusters (Type C, IFN)

Cluster 11 N46 (2/24/9/9/3)
Cluster 12 N42 (1/22/10/6/3)
?First 16 cases
0
5
10
15Y
10
Experiment 1Grouped Sequences

• Clusters containing remarkable increase/decrease
  cases

Cluster 4 N4 (0/3/1/0/0)
Cluster 10 N5 (0/1/0/1/3)
Male 2, Female 2
Cluster 6 N3 (0/1/0/1/1)
Male 4, Female 1
0
5
10
15Y
Male 2, Female 1
11
Experiment 1Relationships between PLT and GPT

• F4, F3 major clusters (Type C, IFN)

Cluster 5 N11 (0/2/1/2/6)
PLT
GPT
0
5
10
15Y
0
5
10
15Y
Normal L (120 x103 /ul)
Normal L (7 IU/L)
Normal H (350 x103 /ul)
Normal H (40 IU/L)
12
Experiment 1Relationships between PLT and GPT

• F1, F2 major clusters (Type C, IFN)

Cluster 11 N46 (2/24/9/9/3)
PLT
GPT
?First 16 cases
?First 16 cases
0
5
10
15Y
13
Experiment 1Relationships between PLT and GPT

• Clusters containing remarkable increase/decrease
  cases

Cluster 4 N4 (0/3/1/0/0)
Cluster 10 N5 (0/1/0/1/3)
PLT
GPT
PLT
Cluster 6 N3 (0/1/0/1/1)
PLT
GPT
GPT
0
5
10
15Y
14
Experiment 1Summary

• Fibrotic stage and PLT course
• F4/F3 major clusters
• Lower PLT level, below the normal range
• Decreasing, or flat courses
• Some F1/F2 cases may follow similar courses
  (ex. Type C IFN cluster 5)
• F1/F2 major clusters
• Moderate PLT level within the normal range
  (ex. Type C with IFN clusters 11, 12, 23)
• Mildly decreasing, or flat courses
• PLT level may correspond to fibrotic stages
• (ex. Type C with IFN clusters 23 gt 12 gt 11
  )

15
Experiment 1Summary

• Fibrotic stage and PLT course
• Fast increase/decrease cases
• Some cases rapidly descent from normal range
  regardless of fibrotic stage (Type C with IFN
  clusters 4,6)
• Some cases rapidly ascent toward normal range
  (after IFN therapy)(Type C with IFN cluster 10)
• Relationships between PLT and GPT
• Chronically high GPT level may induce the descent
  of PLT, progressing the fibrotic stage
• Difficult interpretation of GPT patterns high
  level followed by flat low level
• Cured cases or terminal cases ?

16
Experiment 2 PLT-based Analysis of Progress
Speed of Liver Fibrosis

• Assumption Diminishing PLT (below the normal low
  limit NL) for long time -gt F4 stage
• Using PLT-based stage estimation and biopsy
  information, we analyzed
• Years for reaching F4 stage
• Years elapsed between stages

Normal High (NH)
PLT diminishing below the NL
(7.92Y)
Normal Low (NL)
1Y
F1(biopsy)
F4(PLT)
MID 466
17
Experiment 2Preprocess Sequence Selection

• Excluded sequences
• 1. No biopsy information
• 2. Short of exam lt 3, or duration of exam lt
  2Y
• 3. Inhomogeneous SD of exam intervals gt 1Y

Normal High (NH)
Inhomogeneous seq.
Normal Low (NL)
Interval 14 Y
Interval 1M
MID 215
18
Experiment 2Preprocess Sequence Smoothing

• Removal of short-term changes
• Convolution with Gaussian kernel with 6M
  width(s2.8)

19
Experiment 2Discrimination of Diminishing Cases

• Criteria
• 1. Keeps diminished PLT counts (below the normal
  level) at least for 6 month
• 2. Once diminished, never maintain normal PLT
  level for 6 month at any time until the last
  examination

Normal High (NH)
Diminished Case
Normal Low (NL)
6 M
1Y
MID 466
20
Experiment 2Discrimination of Diminished Cases

• Criteria
• 1. Keeps diminished PLT count (below the normal
  level) at least for 6 month
• 2. Once diminished, never maintain normal PLT
  level for 6 month at any time until the last
  examination

Normal High (NH)
Not Diminished Case
Recovery
Normal Low (NL)
6 M
6 M
MID 37
21
Experiment 2Selection Results

• Selection Results (Total 720 Cases)

Numbers in () ratio ()
Subject forAnalysis
Not diminished below NL
SD of exam interval gt 1Y
of exam lt 3, or Duration of exam lt 2Y
Excluded fromAnalysis
PLT data exist, but no biopsy information
available
22
Experiment 2Years for Reaching F4 Two Measures

• First-exam basis Years from the first PLT
  examination
• First-biopsy basis Years from the first biopsy
• Fibrotic stage at starting date stage observed
  at the first biopsy
• Years 0 if the date for diminishment is earlier
  than the first biopsy

Normal High (NH)
First PLT exam F1 (stage at first biopsy)
Diminished date (below NL)
Normal Low (NL)
First-exam basis
First-biopsy basis
F4 (PLT)
First biopsyF1 (biopsy)
MID 216
23
Experiment 2Years for Reaching F4 Summary for
97 Diminishing Cases (First-exam basis)
24
Experiment 2Years for Reaching F4 Summary for
97 Diminishing Cases (First-biopsy Basis)
25
Experiment 2Years Elapsed between Stages
Summary for 97 Diminishing Cases (First-biopsy
Basis)
26
Experiment 2Summary

• PLT-based estimation about years for reaching F4
  and years elapsed between stages
• Assumption Diminishing PLT (below NL) over long
  time -gt F4 stage
• Keeps diminished PLT below NL at least for 6
  month
• Cannot maintain normal PLT level over 6 month at
  any time from the time it diminished below NL to
  the last examination
• Analyzed a total of 97 diminished cases
  (31 Type B, 40 Type C with IFN, 26 Type C w/o IFN
  cases)
• Years for F4 2.95 Y (av. 97 cases, first-exam
  basis) 2.22 Y (av. 97 cases,
  first-biopsy basis)
• Years between stages 1.60 Y/stage(av. 97
  cases, first-biopsy basis)

27
Conclusion

• Presented the results of temporal analysis of
  platelet data in chronic viral hepatitis dataset
• Findings
• Temporal courses of PLT might be classified into
  some patterns according to their levels and
  trends which might be further related to fibrotic
  stages.
• In some exacerbating cases, liver fibrosis may
  proceed a few times faster than the natural
  courses.
• Further validation of the clinical reasonability
  of the results
• Application of the analysis scheme on other
  items/datasets.

Future Work
28
Cluster Constitution w.r.t. Fibrotic Stages

• Type C w/o IFN (total 99 cases)

29
Grouped Sequences

• F4, F3 Major Clusters (Type C, w/o IFN)

Cluster 3 N47 (1/27/6/7/6)
Cluster 1 N22 (0/4/4/5/9)
0
5
10
15Y
?First 16 cases
?First 16 cases
30
Relationships between PLT and GPT

• F4, F3 major clusters (Type C, w/o IFN)

Cluster 1 N22 (0/4/4/5/9)
PLT
GPT
0
5
10
15Y
?First 16 cases
?First 16 cases