KDD Challenge 2000

Guide to the Collagen Diseases Data Set

Domain

The database was collected at Chiba University hospital. Each patient came to the outpatient clinic of the hospital on collagen diseases, as recommended by a home doctor or a general physician in the local hospital.

Collagen diseases are auto-immune diseases. Patients generate antibodies attacking their own bodies. For example, if a patient generates antibodies in lungs, he/she will chronically lose the respiratory function and finally lose life. The disease mechanisms are only partially known and their classification is still fuzzy. Some patients may generate many kinds of antibodies and their manifestations may include all the characteristics of collagen diseases.

In collagen diseases, thrombosis is one of the most important and severe complications, one of the major causes of death. Thrombosis is an increased coagulation of blood, that cloggs blood vessels. Usually it will last several hours and can repeat over time. Thrombosis can arise from different collagen diseases. It has been found that this complication is closely related to anti-cardiolipin antibodies. This was discovered by physicians, one of whom donated the datasets for discovery challenge.

Thrombosis must be treated as an emergency. It is important to detect and predict the possibilities of its occurrence. However, such database analysis has not been made by any experts on immunology. Domain experts are very much interested in discovering regularities behind patients' observations.

Goals

1. Search for patterns which detect and predict thrombosis.
2. Search for temporal patterns specific/sensitive to thrombosis. (Examination date is very close to the date on thrombosis. If we can find specific/sensitive patterns before/after the thrombosis, they are very useful.)
3. Search for features which classifies collagen diseases correctly.
4. Search for temporal patterns specific/sensitive to each collagen disease.

Domain experts told us that if useful patterns are discovered then they are acceptable in major journals on rheumatology (collagen diseases.)

Evaluation Scheme

One of the domain experts, who is well known for rheumatology, will attend PKDD'99 conference and evaluate all the results. The results will be also evaluated in the clinical environment in the future.

Database

Database consists of three tables. (TSUM_A.CSV, TSUM_B.CSV, TSUM_C.CSV). The patients in these three tables are connected by ID number.

TSUM_A.CSV

Basic information about patients (input by doctors). This dataset includes all patients (about 1000 records).

item	meaning	remark
ID	identification of the patient
Sex
Birthday		YYYY/M/D
Description date	the first date when a patient data was recorded	YY.MM.DD
First date	the date when a patient came to the hospital	YY.MM.DD
Admission	patient was admitted to the hospital (+) or followed at the outpatient clinic (-)
Diagnosis	disease names	multivalued attribute

TSUM_B.CSV

Special laboratory examinations (input by doctors) (measured by the Laboratory on Collagen Diseases). This dataset does not include all the patients, but includes the patients with these special tests.

item	meaning	remark
ID	identification of the patient
Examination Date	date of the test	YYYY/MM/DD
aCL IgG	anti-Cardiolipin antibody (IgG) concentration
aCL IgM	anti-Cardiolipin antibody (IgM) concentration
ANA	anti-nucleus antibody concentration
ANA Pattern	pattern observed in the sheet of ANA examination
aCL IgA	anti-Cardiolipin antibody (IgA) concentration
Diagnosis	disease names	multivalued attribute
KCT	measure of degree of coagulation
RVVT	measure of degree of coagulation
LAC	measure of degree of coagulation
Symptoms	other symptoms observed	multivalued attribute
Thrombosis	degree of thrombosis	0: negative (no thrombosis) 1: positive (the most severe one) 2: positive (severe) 3: positive (mild)

Examination date is very close to the date on thrombosis. In negative examples, these tests are examined when thrombosis is suspected.

TSUM_C.CSV

Laboratory Examinations stored in Hospital Information Systems (Stored from 1980 to March 1999) All the data include ordinary laboratory examinations and have temporal stamps. The tests are not necessarily connected to thrombosis.

item	meaning	normal range
ID	identification of the patient
Date	Date of the laboratory tests (YYMMDD)
GOT	AST glutamic oxaloacetic transaminase	N < 60
GPT	ALT glutamic pylvic transaminase	N < 60
LDH	lactate dehydrogenase	N < 500
ALP	alkaliphophatase	N < 300
TP	total protein	6.0 < N < 8.5
ALB	albumin	3.5 < N < 5.5
UA	uric acid	N > 8.0 (Male) N > 6.5 (Female)
UN	urea nitrogen	N < 30
CRE	creatinine	N < 1.5
T-BIL	total bilirubin	N < 2.0
T-CHO	total cholesterol	N < 250
TG	triglyceride	N < 200
CPK	creatinine phosphokinase	N < 250
GLU	blood glucose	N < 180
WBC	White blood cell	3.5 < N < 9.0
RBC	Red blood cell	3.5 < N < 6.0
HGB	Hemoglobin	10 < N < 17
HCT	Hematoclit	29 < N < 52
PLT	platelet	100 < N < 400
PT	prothrombin time	N < 14
Note	comment for the test PT
APTT	activated partial prothrombin time	N < 45
FG	fibrinogen	150 < N < 450
AT3	marker of DIC, one of the most important complications of collagen diseases	70 < N < 130
A2PI	marker of DIC	70 < N < 130
U-PRO	proteinuria	0 < N < 30
IGG	Ig G	900 < N < 2000
IGA	Ig A	80 < N < 500
IGM	Ig M	40 < N < 400
CRP	C-reactive protein	N= -, +-, or N < 1.0
RA	Rhuematoid Factor	N= -, +-
RF	RAHA	N < 20
C3	complement 3	N > 35
C4	complement 4	N > 10
RNP	anti-ribonuclear protein	N= -, +-
SM	anti-SM	N= -, +-
SCl70	anti-scl70	N= -, +-
SSA	anti-SSA	N= -, +-
SSB	anti-SSB	N= -, +-
CENTROMEA	anti-centromere	N= -, +-
DNA	anti-DNA	N < 8
DNA-II	anti-DNA	N < 8

This database was donated by dr. Katsuhiko Takabayashi and prepared by prof. Shusaku Tsumoto
For possible questions on the data and task description contact Shusaku Tsumoto or Dr. Takabayashi. All questions and answers will be published as appendixes to this document.

Asked Questions in PKDD'99 Discovery Challenge

• The description of attributes (and their normal ranges) doesn't correspond to the data. (21.7.1999)
  Following several questions, I checked the original database in hospital, we found several errors about attribute information. We found errors from PT to TAT2. Please replace the first line with the second one.
  Old: PT APTT FG PIC TAT TAT2 U-PRO
  New: PT Note APTT FG AT3 A2PI U-PRO
  The Normal Range is:
  AT3 70 < N < 130
  A2PI 70 < N < 130
• While most values fall into the normal range (as given by the Guide to the Medical Data Set) this is not the case for item UN and CRE. (16.7.1999)
  These are experts' mistakes. UN: N<30 and CRE: N<1.5 are normal values. Sorry.
• For the following items the normal range does not fit to the data: WBC: values 0.1<=N<=119.5, normal range 3500< N <9000,
  RBC: values 0.01<=N<=6.57, normal range 350< N <600, (16.7.1999)
  Yes, they are also expert's mistakes. Please change the normal range to: WBC: normal range 3.5 < N < 9.0 RBC: normal range: 3.5 < N < 6.0
• Are the values of diagnosis ordered lists or just sets? ie, is 'RA, SJS' equal to 'SJS, RA'? (14.7.1999)
  The values are just sets, so RA,SJS is equal to SJS,RA.
• What means the word 'susp' that comes after many of diagnosis, like 'SLE susp'? And what about the words that comes between parenthesis, like 'BEHCET (entero)', 'EN (r/o BEHCET)' and 'RA (seronegative)'? All diagnosis like 'SLE susp' can be grouped into a higher level like 'SLE'? (14.7.1999)
  Susp stands for "suspected". So, their diagnosis have not been confirmed. 'BEHCET (entero)': entero stands for enterocolitis type of Bechet disease. It is one type of Bechet diseases in which colon is the main target of autoimmune process. Bechet have several types. In case of BECHET (neuro), the main target will be neuron.
  'EN (r/o BEHCET)' :This means that this entercolitis case in which Bechet is strongly suspected.
  'RA (seronegative)': From the observations (symptoms), this case can be diagnosed as RA. But, serum tests (laboratory examinations) are negative. We have had such strange cases in real clinical practice. So, this case is clinically RA, but negative from the labo tests. (So, from the viewpoint of labo tests, they are "true-negative" cases.)
• We came across some attribute values in table TSUM_C which puzzled us a bit. (30.6.1999)
  I found one error in attribute information in Tsumoto_c.csv. There is one laboratory examination between TAT and U-PRO. All the questions about CRP,IGM, RF, IGG are coming from this error. So, please replace the attribute-list:
  ID Date GOT GPT LDH ALP TP ALB UA UN CRE T-BIL T-CHO TG CPK GLU WBC RBC HGB HCT PLT PT APTT FG PIC TAT U-PRO IGG IGA IGM CRP RA RF C3 C4 RNP SM SC170 SSA SSB CENTROMEA DNA DNA-II
  with:
  ID Date GOT GPT LDH ALP TP ALB UA UN CRE T-BIL T-CHO TG CPK GLU WBC RBC HGB HCT PLT PT APTT FG PIC TAT TAT2 U-PRO IGG IGA IGM CRP RA RF C3 C4 RNP SM SC170 SSA SSB CENTROMEA DNA DNA-II
• Another little puzzle has been attribute ANA Pattern in table TSUM_B. First of all, are its values ordered lists or just sets, if any? ie, "P,S" = "S,P"? Are there any other consideration you think important as to this attribute? (30.6.1999)
  This values are just sets, so {P,S} = {S,P}.
• Attributes RNP, SM, SC170, SSA, SSB, CENTROMEA, which are expected to assume [-, +-], are often seen to have numbers as values. How are they supposed to be interpreted? (30.6.1999)
  Usually, these test have two kinds of measurements: qualitative and quantitative. We thought that they are measured by qualitative methods. I will check the normal range.
• We found values such as "<30" and ">=1000" for some numerical attributes. How should they be interpreted? Could they be replaced with some number? (30.6.1999)
  It means that these values are too small or too large. For example, you can set some values to each case. Say, "<30" can be transformed to "10" and ">=1000" to 1500.
• New attribute U-PRO (after TAT-2) sometimes shows to assume value TR. What does it mean?
  TR means that due to some problems with blood serum, the laboratory cannot measure. So, it means that "error in measurement due to the problems with submitted blood serum".
• We have downloaded and started to analyze the medical data set. It seems that 350 of the patient IDs in TSUM_B.csv have no correspondent entry in TSUM_A.csv. Therefore, data of BOTH examinations exist only for about 400 of the 1200 patients in TSUM_A.csv ? (30.6.1999)
  Tsumoto_a.csv includes all the data of patients who are followed by doctors at outpatient clinic in University Hospital at least several months. On the other hand, Tsumoto_b.csv includes the data of two types of patients. The first one is a patient followed at University Hospital. The second one is a patient who is not followed at University Hospital, but specific laboratory examinations are made (even in this case, we will register that patient and provide ID number of university hospital.) So, tsumoto_a.csv and tsumoto_b.csv includes three types of patients:
  1. First type: a patient followed at outpatient clinic in University hospital, but no special examinations are made for this patient. (Patients in the first type do not suffer from thrombosis: that it, they are negative with respect to throbmosis.)
  2. Second type: a patient followed at University hospital and special examinations are made for this patient.
  3. Third type: a patient who is not followed at University hospital, but special examinations are made for this patient.

  Thus, about 400 patients in Tsumoto_b.csv are belonging to third type. But, they are not followed at University hospital, they do not have temporal data. So, please use first type and second type patients for the analysis of thrombosis.

• Are for a given patient the values of attribute "Diagnosis" in table TSUM_A the same as the values of attribute "Diagnosis" in table TSUM_B ?
  Yes. If not, please use the diagnosis in TSUM_A. That is the most recent updated file about diagnosis.
• Are "Diagnosis" (table TSUM_A, TSUM_B) concepts, the contributors to the discovery challenge should consider or, is the only target attribute the "Thrombosis" ?
  Diagnosis is also the target attribute. My colleagues are not only interested in "Thrombosis", but also in "Diagnosis".
• In table TSUM_C, you gave the normal range of values. What is the possible range of all values (e.g. I have found value "+" for attributes with normal range "-", "+-", but I have found value "-" also in attribute with normal range N<8)
  Okay, first, {-,+-,+} is a usual notation in medical "qualitative" tests. "-" is negative (in normal range), "+-" is not negative but at the border of normal range, "+" means positive, or abnormal.
  {-,+-,+} can be observed in a simple test: usually, each symbol corresponds to a range of "quantitative" values. In the case of "-" in Normal range (N<8), "-" means that the value of this test is in the normal range (N<8).

This page is originally written by Peter Berka, PKDD99 Discovery Chair. The donators would like to thank him for his effort and his kindness. Without his efforts, the workshop would neither have had such an impact nor have made us continue using this data for future discovery challenge workshops.

Shusaku Tsumoto and Katsuhiko Takabayashi

Shusaku Tsumoto

Department of Medical Informatics, Shimane Medical University.

E-mail: tsumoto@computer.org

For more details, refer to PKDD99 Discovery Challenge Home Page.

Asked Questions

• question. (21.7.1999)
  answer.