Classification of PBC Disease Profiles in a Contextual Setting

An Analysis of Primary Biliary Cirrhosis (PBC) Patients Survey Data

W.M. van den Bergh[a] and M.I. de Kruijff [b]

Preliminary Version, May 2001

Comments welcome!

Introduction

Primary Biliary Cirrhosis (PBC) is an infrequent chronic liver disease that is increasingly brought into focus. In March 1998 the PBC Support Group performed an Internet Survey on possible factors that patients suspected could lead to PBC. The 231 questions in the Survey were answered by 131 patients.[1] The resulting data matrix was too broad to encourage direct conclusions in a statistical report summarizing the results.[Krivy,1998]

Aware of this problem we wondered if advanced data-mining techniques would be able to improve the results of this Survey. CELA (Competitive Exception Learning Algorithm) [cf. Van den Bergh et al., 2001] is a recent application of a so-called fuzzy identification approach. See [Babuska, 1997] for an overview of such techniques. It constructs a fuzzy model from data where no prior knowledge about the system under study is used to identify system rules. The CELA model has been developed in the field of financial analysis on stock market data. It has the advantage of considering a context of many variables simultaneously and relate it to a target (here a disease profile) that may as well be multi-dimensional.

The epidemiology, the etiology, varieties and treatment of PBC are discussed in recent literature (cf. Parikh-Pathek et al., 2001, Kim WR, 2000, Poupon R, 2000, Heathcote EJ, 2000, van Hoogstraten HJ, 1999). Generally, traditional methods of Multivariate Linear Regression direct to certain relationships in the development of PBC. In a large sample of PBCers compared to control groups without PBC Parikh-Patel, et al. 2001 reported a higher incidence of other auto immune diseases, smoking, tonsillectomy, and female vaginal or urinary tract infections. [2] The smaller sample we consider here is drawn from the same source. We do not use exactly the same variables, e.g. we consider ‘Diet’ but ‘Other Auto Immune Diseases’ is not taken up.

· Our first hypothesis is that we will expect to find that smoking, tonsillectomy and female or UTI infections affect the disease negatively.

Why PBC affects more women than men is unknown. A suggested serological difference was not found. [Nalbabdian G, et al, 1999]

· Our second hypothesis is that we expect to mark a difference between men and women with PBC.

The importance of oestradiol in PBC has been stressed as transport proteins inhibiting bile salt export [Jansen, PL, et al, 2000].

· In our third hypothesis we expect to find a relation between hormone medication and PBC.

In accordance with common understanding that a daily glass of wine is not harmful to your health. Dr. Gerald Minuk of the University of Manitoba, Canada [AASLD, 2000] found that the equivalent of one or two drinks a day appeared to help the damaged livers of lab rats repair themselves. He suggests that if the results are found to apply to humans one or two drinks a day might indeed be beneficial rather than harmful to the liver.

· Our fourth hypothesis is that we expect to find that more than one or two drinks a day make a difference.

We have chosen a number of other contextual variables (like diet and exercise) from the survey data of which the relation with PBC is unexplored. We incorporate them also in our analysis.

Data collection and Method

Only part of the data files of the PBC March 1998 Survey could be retrieved for our analysis. As it turned out 95 respondents had been kept on file. After sorting out double entries and incomplete cases the remaining data set of only 63 respondents was used for this analysis. Although general statistics for this group proved roughly similar to the PBC March 1998 Survey Summary Results, it must be noted that our population is less than half of the 131 respondents used in the original.[3]

Considerable re-coding of the data coming from different parts of the world, using different laboratory values was necessary for use by the CELA model. Apart from inconsistent answers on some indeed suggestive questions the data material was remarkably good.[4] The study population consists of PBC patients only. No control groups were formed.

We ended with 50 endogenous variables built from answers to the questions on LFT's, latest known stage of PBC and symptoms. The contextual information, trying to explain the occurrence of PBC is contained in 53 exogenous variables. They are constructed from answers to the questions in the Survey on Personal factors as: Sexe, Age, Height, Weight, Time since diagnosis, Time since onset menopause. On Life Style factors as: Current Diet, Exercise, Smoking and Alcohol habits. Response on questions concerning: Childhood diseases, Tonsillectomy, Pneumonia, Urinary Tract Infections, Female Infections, Sinus Infections, Gall Bladder Problems and information on Prescribed and Alternative Medication. Given the limitations of the sample CELA will explore on remaining open questions by focussing on the effect of different factors in various stages of the disease.

CELA and PBC

The CELA model discloses a conditional relationship between vectors of patient information in a multi dimensional input space (X) and a multi dimensional output space (Y). As explained earlier, data points in Y refer to the PBC disease itself and in X to the contextual setting of the disease.

First a characterization of the disease profiles is made in the Y-space by applying a clustering method (the number of clusters was arbitrarily set to 5). The method ties similar cases together as much as possible. To get some intuition for the method imagine that all patients compete for a classification that is as crisp as possible. Each data point fights to attract one of the cluster centroids. The higher the local density of data points, the higher the probability of getting a cluster centroid nearby. Once these optimal cluster locations have been established each cluster describes a ‘prototype’ disease profile. Now the patients can be classified in a fuzzy manner, meaning that each patient is, to a certain extent, member of all clusters.[5] The memberships values add up to one, so a patient belongs for p1% to prototype 1, p2% to prototype 2…….and for p5% to prototype 5.

Next the algorithm averages the membership values of all patients per prototype (cluster). In this way we arrive at a frequency distribution of the patients over the optimized clusters and in absence of any contextual information we expect a patient to behave as the average patient with the average cluster membership distribution (the Unconditional Y-cluster Distribution, UYD). From this unconditional frequency distribution we can identify a hypothetical patient that exactly satisfies the UYD and serves as Most Representative Case (MRC).[6] The MRC will be used as a benchmark for conditional Y distribution estimates.

Finally the contextual information from the (53) exogenous variables in the X-space is clustered again using a competitive learning approach. This time however we try to optimize the location of X-clusters in such way that they capture systematic deviations from the MRC, i.e. deviations from the average disease record in the Y-space. Thus, the competition is for ‘hot spots’ in the X-space showing exceptional local behavior with respect to their Y distribution. Once this optimization step is completed, we can determine an X-cluster membership distribution for each individual data point.[7] And since we know the exceptional Y-behavior in the X-cluster centroids, we can intrapolate on this and estimate the Y-behavior for each X-location, i.e. the disease profile for each specific contextual setting. By averaging the individual membership values per cluster we arrive at the Conditional Y-Distribution (CYD) in contrast to the Unconditional Y-Distribution (UYD) which is relevant when no prior context information is available. In a similar manner as we determined a MRC in the Y-space, we can compute an MRC in the X-space (an ‘average’ X-context) which will be used as a benchmark for ‘exceptional’contextual settings.

Analysis of the Patients Classification

Although we arbitrarily set the number of X-clusters to 5, the clustering of hot spots in risk factors and disease profile of PBC brings about a clear dichotomy. One cluster is large and visibly different from the other clusters. The other clusters intertwine to form a second group. This is apparent in the next figure.

In this picture all cases have been sorted (in decreasing order) following their membership to the largest cluster. The general picture is that the membership for each of the smaller clusters increases when the membership for the largest cluster decreases. Accordingly we analyze the four smaller clusters as one group (55% of all cases).

The other group of 45% (the large single cluster) has higher scores on the lighter aspects of PBC. This group seems to have a lighter disease profile in PBC compared to the remaining population in the sample. Thus we will distinguish a PBCLite group and a PBCPlus group. First the distinction between the groups in the Y-space (the disease itself) is analyzed. Note that this distinction is the result of a joint effect of disease profile and contextual setting as discovered by CELA.

In all following charts the cases on the x-axis are sorted in decreasing order according to their PBCLite group membership. Thus the crispest PBCLite cases are situated on the left side and the crispest PBCPlus cases on the righthand side. The values on the y-axes are plotted in deviation from the benchmark (MRC).

Y-space: disease profile

PBC Stage

The Stage of PBC is found to be lower in the PBCLite group. Generally the values for this group are below average of Stage 1-2. Low values are more frequent but we observe also some higher values. In contrast PBCPlus group averages are above stage 1-2.

Liver Function Tests

Alkaline Phosphatase, AST, Calcium, CGT, LDH, Sodium and Total Bilirubin values are below average in PBCLite group and above average in PBCPlus group.

Albumin, ALT, AMA, Creatinine, Potassium, Protime and Total Protein values are above average in the PBCLite group and below average in the PBCPlus group.

Values for Cholesterol and Triglycerides show no discrimination between the two groups.

Alkaline Phosphatase is under benchmark (left) in PBC Lite and above benchmark (right) in PBCPlus.

Protime is high (left) in PBCLite above benchmark and low (right) in PBCPlus.

 

	Benchmark		PBCLite	PBCPlus
	US units	SI units
Alk Phos	227 U/L	227 U/L	-	+
AST	58 U/L	58 U/L	-	+
Calcium	9.65 mg/dl	2.41 mmol/L	-	+
Choles	247 mg/dl	6.42 mmol/L	0	0
GGT	379 U/L	379 U/L	-	+
LDH	406 U/L	406 U/L	-	+
Sodium	144 mEq/l	144 mmol/L	-	+
T Bilirub	1.185 mg/dl	20 ¼mol/L	-	+
Albumin	3.9 g/dl	39 g/L	+	-
ALT	43 U/L	43 U/L	+	-
AMA	1:250 titer	1:250 titer	+	-
Creatinine	0.70 mg/dl	61.9 ¼mol/L	+	-
Potassium	4.3 mmol/L	4.3 mmol/L	+	-
Protime	7.5 seconds	7.5 seconds	+	-
T Protein	5.77 g/dl	57.7 g/L	+	-
Triglycer	142 mg/dl	1.56 mmol/L	0	0

Symptoms

The PBCPlus group suffer above average from symptoms associated with PBC as Itching, Fatigue, Joint Pain, Jaundice, Dry Eyes, Sleep Disturbance, Pain in ribs, Pressure pain in liver area, Abdomen swelling, Edema (Ascites), Bleeding gums, Dental problems, Mouth ulcers, Mood swings, Shortness of breath, Memory problems, Spider veins, Skin rashes, Tingling in hands or feet, Red palms, Very dry skin, Nail abnormalities, Nausea and Hair loss. [8] PBCLite group members are often asymptomatic or have only suffered symptoms prior to diagnosis.

Symptoms as Dry Mouth, Small Fat Deposits (bumps) around eyes, Change in color and skin texture, Ankles and Leg Swelling and Leg Cramps are very lightly found with a few PBCLite members and very strongly with a few PBCPlus members.

PBCLite group under benchmark 1.3604. No nausea and stomach problems or prior to diagnosis.

PBCLite group under benchmark 0.4164. No Ankles or legs swelling

X-space: Contextual Setting

Personal characteristics

The few men in the sample (10%) are over-represented in the PBCPlus group. PBCPlus group is over 50 years. Women in this group are more than 4 years post menopausal. Their height is under 5.8 ft (1.77 m). Their weight tends to be more than 199 lbs (90 kg).

As the two groups mirror each other the PBCLite group includes more women. They are younger than 50 years. Less than 4 years post menopausal. Taller than 5.8 ft and weighing under 199 lbs.

Diseases

Childhood diseases found more in PBCPlus cases are:

Mumps, Rubella, both Mononucleosis and Polio (very clearly), Typhoid Fever, Scarlet Fever, Rheumatoid Fever and Heart Mummer. Remarkably is that they had less Chicken Pox, Strep Throat. These exceptions Chicken Pox and StrepThroat are more found in the PBCLite group. Day Measles does not differ much between groups.

Later diseases more found in the PBCPlus group are: Pneumonia which they had more than 3 times. Urinary Tract Infections (frequent) and Female Infections.

Sinus problems are slightly higher in PBCPlus. The Lite group averages on drainage and infections problems; in the Plus group also parents with sinus problems were found.

The PBCLite group however had more Tonsillectomies.

Mononucleosis is clearly more found in the PBCPlus group

Chicken Pox is more found in the PBCLite group.

Gall Bladder problems are more found in the PBCPlus group.

In PBCPlus more Gall Bladder problems are found that lasted for more than 5 years before diagnosis

PLUS strong family history of Gall Bladder problems. Both parents and grandparents are involved

Medication

It is remarked that other than Ursodeoxycholic acid (UDCA) medication is considered only. The PBCPlus group use more Methotrexate, Prednisone, Colchicine, Questran, Rifampin, Fosamax, Anti Inflammatory medication, Daily Aspirin and Vitamins. The same goes for Lactulose and Milk Thistle but here the difference between groups is sharper. [9]Calcium and Hormone Replacement Treatment (HRT) are slightly more used by PBCPlus but do not differentiate very well between groups. Pain medication is used evenly by both groups. Birth Control Pills have (ever) been more taken by the PBCLite group. However, PBCPlus members used the Pill longer (than 10 years).

Life Style

PBCLite diet is average to high fat junk food. The PBCPlus group however eat healthy low fat meals. Eating beef does not seem to make a difference between groups. PBCLite exercise on average less than once a week. PBCPlus exercise more.

Smoking does not discriminate very well between Groups. In the PBCPlus group members smoked on average more than 23 years more than 16 cigarettes a day. This relation does not seem to be masked by age since there was not much difference found in age between groups. In the environment of the PBCLite group parents are found to smoke prior to their birth only. PBCPlus group however had parents smoking prior and after birth as well as smoking partners.

Drinking alcohol does not discriminate very well between groups. Perhaps PBCLite take a little more. They drink for a shorter period than PBCPlus but still up to 35 years. The type of drink does not differ between groups. However the amount of glasses a week taken on average by the PBCLite group is less than 1.5 glass per day. More than 1.5 glass daily are taken in the PBCPlus Group.

Conclusions

CELA discovered two distinct groups of PBCers in a small sample of PBC patients. One groups is obviously less affected by the disease than the other group. The PBCLite group consists of more women, younger, shorter time post menopausal, taller and slimmer than the PBCPlus group. PBCLite group have a lower Stage of PBC: on average stage 1-2. They are symptom free or had symptoms prior to diagnosis. PBCLite have lower Liver Function Test values on Alkaline Phosphatase, AST, Calcium, CGT, LDH, Sodium and Total Bilirubin. Higher values on Albumin, ALT, AMA, Creatinine, Potassium, Protime and Total Protein.

PBCLite had less than average of the Childhood diseases Mumps, Rubella, Typhoid Fever, Scarlet Fever, Rheumatoid Fever, Heart Mummer. And clearly less Mononucleosis and Polio. However, they had more than average Chicken Pox and Strep Throat and Tonsillectomy. Approximately the same Day Measles as the PBC Plus group.

PBCLite have had less Pneumonia, Urinary Tract Infections, and Female Infections. The same Sinus Problems. On average PBCLite had no Gall Bladder problems while PBCPlus had Gall Bladder Problems longer than 5 years before diagnosis plus parents and grandparents with those problems.

Medication in less used in the PBCLite group. Birth Control Pills have (ever) been more taken by the PBCLite group, but for a shorter time. Above average years of using Birth Control Pills is found in the PBCPlus group. Hormone Replacement Treatment is more or less the same among groups.

Life styles are different. PBCLite uses more fat and exercises less than PBCPlus who eats healthy low fat and exercises more than once a week. Note that the data do not reveal whether (nor when) a patient eventually adopted a healthier life style as a result of PBC.

Not much difference is found in smoking and non smoking habits. But PBCPlus smoked longer and more. There is a difference in the smoking environment of the groups. The PBCLite group had parents smoking prior to their birth only. PBCPlus group have had an environment of smoking parents prior and after birth and smoking partners.

Alcohol consumption and the type of drink does not differ much between groups. Perhaps slightly more PBCLite members actually use alcohol. But PBCPlus used to drink more years and more glasses a day. While PBCLite remains under an average of 1.5 glass daily.

We will now return to the hypotheses raised in the introduction.

· Hypothesis I. We expected to find that smoking, tonsillectomy, female and UTI infections influence the disease. This was confirmed, however there was a difference between the groups.

Smoking or not smoking does not discriminate very well between groups. However, PBCPlus group smoked longer (than 23 years) and more (than 16) cigarettes a day. This relation does not seem to be masked by age since there was not much difference found in age between groups. Also the smoking environment of groups was different. PBCPlus group had parents smoking prior and after birth as well as smoking partners. This suggests that the correlation between smoking and PBC should be interpreted rather in terms of the duration, the amount of cigarettes smoked and the smoking environment. Also this relation is found in patients more strongly affected by the disease.

Tonsillectomy is found more often with members of the PBCLite group. Furthermore, we found that Sinus infections are present in both groups. Strep Throat and Chicken Pox are remarkably more found in the PBCLite group. UTI, Female Infections, but also Pneumonia and Gall Bladder problems are more found in the PBCPlus group. Moreover, the family history of Sinus and Gall Bladder problems are stronger in the Plus group. Furthermore, there was a sharp distinction between groups in diseases as Mononucleosis and Poliomyelitis, which means that PBCPlus had those diseases far more and PBCLite had them far less.

· Hypothesis II. We expected to mark a difference between men and women with PBC. The very few men in the sample were relatively more found in the PBCPlus group. Thus, men in this sample are relatively more affected by the disease than women.

· Hypothesis III. We expected to find a relation between hormone medication and PBC. PBC Plus group members are longer menopausal. Use of HRT is more or less the same, perhaps a bit higher in the Plus group. Fosamax is clearly more taken in the PBC Plus group. Birth Control Pills have (ever) been taken by more PBCLite members. However the difference is found in the duration that the 'Pill' has been taken: PBCPlus group took BC Pills for longer than 10 years.

· Hypothesis IV. We expected to find that more than one or two drinks a day make a difference. Drinking alcohol nor the type of drink does make a difference between groups. Perhaps PBCLite drink slightly more. However, the PBC Plus group drank more years and more daily glasses. Alcohol consumption in the PBCLite group averages on 1.5 glass per day.

The hypotheses are thus preliminary confirmed, be it in more refined terms.

This demonstrates that the context of a patient's record on childhood diseases, later diseases, medication and life style habits is instrumental to classify a patient in a phase of PBC with accompanying stage, symptoms, and LFT' s.

We conclude with the specific rule detected by the Competitive Exception Learning Algorithm (CELA):

IF the context of former diseases, medication and life style habits deviates in a specific manner from the average context, viz.:. more childhood diseases, later infection diseases Sinus and Gall Bladder problems. Less Chicken Pox, Strep Throat, Tonsillectomies. More medication. Less fat in diet and more exercise. More years of smoking, drinking and using Birth Control Pills plus more than 1.5 glasses of alcohol daily

THEN we expect a disease profile that systematically deviates from the average (UYD) profile in that it has a more advanced stage of PBC, more symptoms, higher LFT's for Alkaline Phosphatase and Bilirubin but lower values for Albumin and Protime. We have named this PBCPlus. Members of the PBCPlus group are, although PBC is predominated by women, more likely to be men than in the PBC Lite group, older, shorter and heavier.

Discussion

The available sample of 63 respondents was rather small. The disturbance of missing values has been avoided by incorporating them in the model as average or median values. CELA found an interesting dichotomy of patients that are more and patients that are less affected by PBC. These results confirm the very distinction of more active and severe types or different phases in PBC found in the literature on PBC. [Invernizzi et al, 2001, Masuda et al, 2001] Caution is called for since we do not yet have a clear significance test for the CELA estimates.[10] Prudence is needed when attributing the reported symptoms to PBC. The relation may be spurious, e.g. some symptoms are often also associated with aging or menopause.

We did not take up the variable UCDA treatment because nearly all patients use it and a lot of research has already been done on this subject. It is obvious that the UCDA treatment is somewhere hidden in the results of this analysis. It will be an objective for our next study on more elaborate data to have the dynamics of this relation come to surface.

An interesting and possibly interacting combination found in the context of this sample is the longer time smoking, drinking and taking Birth Control Pills. Perhaps surprising is that PBCPlussers are eating healthier and exercise more than the Lite group. Note that the data do not reveal whether a patient eventually adopted a healthier life style as a result of PBC.

This study must be understood as a pilot for further application on data from the PBC Support Group, preferably extended with data bases from European countries. Other large databases of PBC patients' data (and control groups) are invited to join for analysis by CELA.

Next objective would be to look into the dynamics of the disease. With a comparison of at least 3 moments in time with dosage of UDCA. Our questions for further research would be : Two types of PBC? Can we discern a turning point at stage 2?

Comments are invited!

Literature

AASLD, 2000, report to the annual meeting of the American Association for the study of Liver Disease.(AASLD) in Dallas. Reuters, November 8, 2000 "A little drink good for liver"

Angulo P, Patel T, Jorgensen RA, Therneau TM, Lindor KD, 2000, ‘Silimarin in the treatment of patients with primary biliary cirrhosis with a suboptimal responds to ursodeoxycholic acid.’ Hepatology 2000 Nov;32(5):897-900.

Babuska R, 1997, “Fuzzy Modeling and Identification”, PhD dissertation Technical University Delft, pp294, ISBN 90-9010153-5.

Erickson JM, Mawson, AR,2000, "Possible role of endogenous retinoid (Vitamin A) toxicity in the pathophysiology of primary biliary cirrhosis." Journal Theor Biol 2000 Sep 7;206(1)47-54.

Heathcote EJ, 2000, "Management of Primary Biliary Cirrhosis". Hepatology 2000:31:1005-1013.

Invernizzi P, Podda M, Battezzati PM, Crosignani A, Zuin M, Hitchman E, Maggioni M, Meroni PL, Penner E, Wesierska-Gadek J. 2001," Autoantibodies against nuclear pore complexes are associated with more active and severe liver disease in primary biliary cirrhosis." Journal of Hepatology, 2001 Mar;34(3):366-72.

Jansen PL., et al, 2000, "Bile formation and cholestasis." Ned. Tijdschrift voor Geneeskunde 2000 Dec 9;144(50):2384-91

Kim WR, Lindor KD, Locke GR, Therneau TM, Homburger HA, Batts KP, Yawn BP, Petz JL, Melton LJ, Dickson ER, 2000, " Epidemiology and natural history of primary biliary cirrhosis in a US community" , Gastroenterology 2000 Dec;119(6):1631-6.

Krivy KE, 1998, "PBCers March 1998 Survey Summary Results."

Masuda J, Omagari K, Matsuo I, Kinoshita H, Sakimura K, Hazama I, Ohba K, Isomoto H, Murase K, Murata I, Kohno S. 2001," Changes in titers of antimitochondrial and antinuclear antibodies during the course of primary biliary cirrhosis." J Gastroenterol Hepatol 2001 Feb;16(2):239-43.

Nalbabdian G, Van de Water J, Gish R, Manns M, Coppel RL, Rudich SM, Prindiville T, Gerhwin ME 1999, "Is there a serological difference between men and women with primary biliary cirrhosis?", American Journal of Gastroenterology 1999, sep;94(9):2482-6.

Parikh-Pathek A, Gold BE, Worman H, Krivy, KE, Gershwin ME, 2001, "Risk Factors for Primary Biliary Cirrhosis in a cohort of patients from the United States", Hepatology, vol 33, no 1.

Poupon R, Chazouilleres O, Poupon RE, 2000," Chronic cholestatic diseases", Journal of Hepatology 2000;32(1 suppl):129-40.

Tanaka A, et al., 1999 "Are infectious agents involved in primary biliary cirrhosis?", Journal of Hepatology, 1999, Oct;31(4):664-71.

Van Hoogstraten HJ, Hansen BE, van Buuren HR, ten Kate FJ, van Berge-Henegouwen GP, Schalm SW, 1999, " Prognostic factors and long-term effects of ursodeoxycholic acid on liver biochemical parameters in patients with primary biliary cirrhosis. Dutch Multi-Centre PBC Study Group", Journal of Hepatology 1999, Aug;31(2):256-62.

Van den Bergh WM,Van den Berg J,Kaymak U, 2001, “Detecting Noise Trading using Fuzzy Exception Learning” in: Fuzziness and Soft Computing in the New Millennium, Joint 9th IFSA World Congress and 20th NAFIPS International Conference of NAFIPS, July 25-28, Vancouver BC.

Notes

--------------------------------------------------------------------------------

[a] Department of Finance, Erasmus University Rotterdam, Netherlands

[b] Corresponding author: m.de.kruijff@freeler.nl

--------------------------------------------------------------------------------

1 PBC Support Group: http://pbcers.org. Initiated by Linie Moore in 1996. In March 1998 the Group consisted of approximately 300 members. In March 2001 more than 1400 members are registered

[2] It has been suggested that bacterial infections may play a role in the initiation of PBC. Tanaka A, et al., 1999 find no evidence for ongoing chronic infectious processes in PBC patients.

[3] In order to check comparability a report of the statistical analysis of the 63 respondents was made.

[4] It is remarkable that 100% of the respondents reported to be willing to help find a cure for PBC. Even after answering the long questionnaire with answer categories that were not always excluding and exhausting. Extra questions for testing purposes obviously confused respondents resulting in contradictory answers for fatigue, itching and memory problems.

[5] In the uncommon case that a case is located exactly on a cluster centroid it has a membership value of one for that cluster and of zero for all other clusters.

[6] This average is calculated by weighting the cluster centroids with their relative overall membership values. The MRC may, of course, deviate from overall statistical average.

[7] Again this is a fuzzy frequency distribution since each individual data point may belong to a certain degree to all clusters. So a given case belongs with p1% to prototype context 1, with p2% to prototype context 2…..etc.

[8] Erickson JM, et al 2000, suggest that major symptoms of PBC as pruritus, lethargy, sicca syndrome and osteoporosis might be the effect of hypervitaminosis A.

[9] Milk Thistle is clearly more used by PBCers in advanced stages of the disease. Discussion on usefulness of this alternative medication is ongoing. Angulo P, et al, 2000, Conclude that Silymarin did not provide benefit to patients with PBC responding suboptimally to Ursodeoxycholic acid.

[10]No appropriate test for defining the significance in CELA is currently available. This is one of the next steps to be taken by designers of the model.