[Skip to Content]
Access to paid content on this site is currently suspended due to excessive activity being detected from your IP address 54.159.197.114. Please contact the publisher to request reinstatement.
[Skip to Content Landing]
Download PDF
Figure.
The final decision tree with decisionrules and number of patients.

The final decision tree with decisionrules and number of patients.

Table 1. 
Meibomian Gland Dysfunction (MGD) Rules*
Meibomian Gland Dysfunction (MGD) Rules*
Table 2. 
Medians of Physiologic and Blepharitis Variables by 9 Groupsof Classification Tree
Medians of Physiologic and Blepharitis Variables by 9 Groupsof Classification Tree
Table 3. 
Final Decision Tree
Final Decision Tree
Table 4. 
Cross-Tabulation Between the Groups From the Cluster Classificationand the Physiologic Diagnosis*
Cross-Tabulation Between the Groups From the Cluster Classificationand the Physiologic Diagnosis*
1.
Mathers  WDBinarao  GPetroll  M Ocular water evaporation and the dry eye: a new measuring device. Cornea 1993;12335- 340
PubMedArticle
2.
Mathers  WDDaley  TE Tear flow and evaporation in patients with and without dry eye. Ophthalmology 1996;103664- 669
PubMedArticle
3.
Mathers  WDLane  JAZimmerman  MB Tear film changes associated with normal aging. Cornea 1996;15229- 234
PubMedArticle
4.
Mathers  WDLane  JA Meibomian gland lipid, evaporation, and tear film stability. Sullivan  DADartt  DAMeneray  MAedsLacrimalGland, Tear Film, and Dry Eye Syndromes 2 New York, NY Plenum Press1998;349- 360
5.
Mathers  WDLane  JASutphin  JEZimmerman  MB Model for ocular tear film function. Cornea 1996;15110- 119
PubMedArticle
6.
Mathers  WDShields  WJSachdev  MSPetroll  WMJester  JV Meibomian gland dysfunction in chronic blepharitis. Cornea 1991;10277- 285
PubMedArticle
7.
Ward  JH Hierarchical grouping to optimize an objective function. J Am Stat Assoc 1963;58236- 244Article
8.
Cleveland  WSGrosse  EShyu  WM Local regression models. Chambers  JMHastie  TJedsStatisical Modelsin S. New York, NY Wadsworth & Brooks/Cole1991;
9.
Breiman  LFriedman  JHOlsen  RStone  CJ Classification and Regression Trees.  New York, NY Wadsworth & Brooks/Cole1984;
10.
Ihaka  RGentleman  R R: a language for data analysis and graphics. J Comput Graph Statist 1996;5299- 314
11.
Thygeson  P The etiology and treatment of blepharitis: a study in military personnel. Arch Ophthalmol 1946;98191- 203
12.
McCulley  JPDougherty  JMDeneau  DG Classification of chronic blepharitis. Ophthalmology 1982;891173- 1179
PubMedArticle
13.
Mathers  W Why the eye becomes dry: a cornea and lacrimal gland feedback model. CLAO J 2000;26159
PubMed
14.
Tetz  MRKlein  UVolcker  HE  et al.  Staphylococci-associated blepharo-kerato-conjunctivitis [in German]. Ophthalmologe 1997;94186- 190
PubMedArticle
15.
Seal  DFicker  LRamakrishnan  MWright  P Role of staphylococcal toxin production in blepharitis. Ophthalmology 1990;971684- 1688
PubMedArticle
16.
McCulley  JPDougherty  JM Bacterial aspects of chronic blepharitis. Trans Ophthalmol Soc U K 1986;105314- 318
PubMed
17.
Dougherty  JMMcCulley  JP Comparative bacteriology of chronic blepharitis. Br J Ophthalmol 1984;68524- 528
PubMedArticle
18.
Shine  WEMcCulley  JP The role of cholesterol in chronic blepharitis. Invest Ophthalmol Vis Sci 1991;322272- 2280
PubMed
19.
Shine  WESilvany  RMcCulley  JP Relation of cholesterol-stimulated Staphylococcus aureus growth tochronic blepharitis. Invest Ophthalmol Vis Sci 1993;342291- 2296
PubMed
20.
Tsubota  KYamada  M Tear evaporation from the ocular surface. Invest Ophthalmol Vis Sci 1992;332942- 2950
PubMed
21.
Mathers  WD Ocular evaporation in meibomian gland dysfunction and dry eye. Ophthalmology 1993;100347- 351
PubMedArticle
Epidemiology
November 2004

Cluster Analysis of Patients With Ocular Surface Disease, Blepharitis,and Dry Eye

Author Affiliations

Author Affiliations: Department of Ophthalmology,Casey Eye Institute (Dr Mathers), and Division of Biostatistics, Departmentof Preventive Medicine (Dr Choi), Oregon Health and Science University, Portland.

Arch Ophthalmol. 2004;122(11):1700-1704. doi:10.1001/archopht.122.11.1700
Abstract

Objective  To develop a classification system for blepharitis and dry eye basedon a classification-tree model of a large group of subjects who were givena variety of objective physiologic tests.

Methods  We evaluated 513 subjects, some healthy and some with blepharitis anddry eye,with tests for tear volume, tear flow, and tear turnover and the Schirmertest for dry eye. Meibomian gland function was evaluated by meibomian glandlipid expression for lipid volume and lipid viscosity, evaporation, and eyelidtransillumination for meibomian gland drop out. We subjected these data tocluster analysis and formulated a classification tree.

Main Outcome Measure  The outcome measure of this study was the statistically valid groupsof subjects with and without ocular surface symptoms identified by their physiologiccharacteristics.

Results  Cluster analysis most successfully grouped subjects by initially dividingthem into 2 groups based on the presence or absence of gland drop out andthen by lipid viscosity and volume, Schirmer test results, and evaporation.The analysis created 9 categories. This division created an objective classificationsystem that was found to have clinical relevance. Normal subjects were distributedacross several groups.

Conclusions  Using a classification tree, blepharitis and dry eye can be classifiedwith objective physiologic tests into clinically relevant groups that havecommon characteristics. The analysis establishes the central role of meibomiangland dysfunction in blepharitis and demonstrates the diverse characteristicsof the normal population.

We undertook a long-term project to measure a series of physiologicparameters in normal subjects and in patients with ocular surface diseaseincluding dry eye, blepharitis, and other related symptom complexes, suchas rosacea and Sjögren syndrome. We performed a statistical analysisof the data to assess whether our current classification system was justifiedor whether we needed to search for new methods of examining patients and collectingdata that could group and classify patients successfully. We hoped such ananalysis would also clarify pathophysiologic mechanisms that have not beenpreviously considered or appreciated.

METHODS

The 513 subjects in this study were taken from our clinic populationat the University of Iowa, Iowa City, and at Oregon Health and Science University,Portland. Most were referred with a diagnosis of dry eye and/or blepharitis.For each subject, we recorded an initial diagnosis of dry eye, blepharitis,rosacea, Sjögren syndrome, normal, or other. These diagnoses, exceptthe diagnosis of normal, were not mutually exclusive. Of 662 subjects tested,149 were eliminated for missing values, leaving 513 subjects in the finaldata set. The study was reviewed and approved by the institutional reviewboard of each institution.

Patients were recruited for the normal group without reference to symptomsof dry eye or blepharitis. Thus, they were categorized as normal only to theextent that they believed themselves to be normal; they were using no eyedropsor eye medications, and they had no history of eye surgery or eye diseaseand were not contact-lens wearers. Some of these patients may have had occasionalsymptoms of dry eye. A separate group of women who did not have symptoms wasrecruited and tested to determine mean and standard deviation values for asymptomaticsubjects.

The largest single group referred for evaluation and testing came withthe initial diagnosis of dry eye, and all of these subjects complained ofdry eye symptoms. We did not use a standardized questionnaire to stratifythe severity of their dry eye complaints. The initial diagnosis of Sjögrensyndrome was based on a clinical history of a systemic inflammatory disease,dry eye, and dry mouth. Many patients were also referred with an initial diagnosisof blepharitis based on the subject’s symptoms and the referring physician’sclinical assessment. There was no attempt to identify the type of blepharitisor its severity since there is no common diagnostic method and most physiciansdid not differentiate blepharitis into subsets of any kind. Patients who werereferred with an initial diagnosis of rosacea were given this diagnosis basedon dermatologic findings. Seventy-one patients also had other conditions thatmight contribute to their ocular surface problems. For our analysis, we assignedthis group an additional initial diagnosis, which we designated as “other.”The 2 most common diagnoses were basement membrane dystrophy (10 subjects)or ocular allergies and atopic disease (12 subjects).

The physiologic parameters tested were the same for all patients andhave been previously reported. The tests for dry eye included tear osmolarity,fluorophotometric tear volume, tear flow, and tear turnover (decay constant)and the Schirmer test without anesthesia.13 Tearosmolarity was measured with freezing point depression using a Clifton NanoliterOsmometer (Clifton Technical Physics, Hartford, NY). Fluorophotometry wasperformed with a Fluorotron Master Fluorophotometer (OcuMetrics, Moutainview,Calif). The tests for meibomian gland dysfunction and blepharitis includedmeibomian gland expression for lipid volume and viscosity, meibomian glanddrop out, and tear film evaporation. A description of these tests has beenpresented in previous publications.1,46

Lipid volume was assessed as the average diameter of the expressiblelipid measured at the slitlamp following 5 seconds of digital pressure onthe lower lid. Lipid viscosity was based on a scale of 1 to 4, with 1 representingclear lipid and 4 representing very thickened, opaque viscous lipid with theconsistency of toothpaste.

CLINICAL ASSESSMENT AND CLASSIFICATION

We assigned 1 or more physiologic diagnoses of physiologic dry eye andseborrheic or obstructive meibomian gland dysfunction, based on physiologicparameters.4,5 We also assigneda clinical diagnosis of rosacea, Sjögren syndrome, normal, or other basedon the subject’s initial clinical manifestation. We developed a setof rules to define physiologic dry eye based on data from a separate set ofnormal women. For this, a separate group of 76 normal women, aged 35 to 60years, without any symptoms of dry eye or other eye disease, was used. Wetested these subjects to determine the mean and standard deviation for 5 dryeye tests: the Schirmer test, fluorophotometric tear volume, flow and tearturnover, and tear osmolarity. A value at or outside 1 standard deviationwas used to define physiologic dry eye. No adjustment was made for age orsex. This is a relatively subjective cutoff to signify values that constitutephysiologic dry eye. The range of values found in the normal, asymptomaticpopulation is, however, wide. By using 1 standard deviation, we produced definingvalues that seemed reasonable and were close to those usually associated withdry eye. The mean ± SD data for normal subjects for turnoverof tear volume per minute was 18% ± 7% and the defining valuewas ≤11.0%; for tear volume, 1.91 ± 1.40 μL and ≤0.48 μL;for tear flow, 0.36 ± 0.33 μL/min and ≤0.03 μL/min;for osmolarity, 308 ± 10 mmol/mL and ≥ 318 mmol/mL;and for the Schirmer test, 16.4 ± 10.7 mm of wetting in 5minutes without anesthetic and ≤5.7 mm of wetting in 5 minutes withoutanesthetic.

We categorized patients as having either obstructive or seborrheic meibomiangland dysfunction, based on their lipid expression results. Rules definingobstructive and seborrheic meibomian gland dysfunction are listed in Table 1. The presence of a single indicator wassufficient to assign that diagnosis. Our diagnosis of rosacea was based onlyon examination of the facial skin, not on eye examination findings or meibomiangland expression.

STATISTICAL METHODS

Hierarchical clustering (Ward minimum variance method) was initiallyapplied to all subjects by using all physiologic variables (Schirmer test,fluorophotometric tear volume, flow and tear turnover, and tear osmolarity)and all blepharitis variables (evaporation, gland drop out, lipid viscosity,and lipid volume).7 Potential age effects inthe physiologic variables were removed by using local polynomial regressionbefore the cluster analysis.8

The results of cluster analysis were used as a starting point for aclassification-tree model. This classification-tree model was fitted with10-fold cross-validation by using all original physiologic and blepharitisvariables.9 All computation was done with Rstatistical computing language.10

RESULTS

Using cluster analysis, we attempted to create groups that correspondedto clinically relevant and identifiable diagnoses. The size and propertiesof the data set ultimately determine the extent to which this process canbe carried. We also selected for the simplest classification method that deriveduseful clinical information. Six clusters were initially identified from thecluster analysis. Using the 6 groups as a starting point, the final fittedclassification tree divides 513 subjects into 9 categories. The Figure shows the final decision tree with decision rules and numberof patients. We found we needed to use only 5 of the 13 variables to establisha classification system that corresponded, at least partially, to clinicallyderived systems. These variables were, in order of their use, gland drop out,lipid viscosity, evaporation, Schirmer test value, and lipid volume. Furthersubdivisions did not produce additional groups that had clinical significance.

The median values of the 9 groups are presented in Table 2. Meibomian gland drop out, found only in groups 1 and 4,differentiates these 2 groups from the others. Drop out is not listed as acolumn in Table 2. Although the numberof glands missing in each lower lid was recorded and evaluated as a variable,we found it was not necessary to use the number of glands missing. Only thepresence of any gland drop out was needed to differentiate groups 1 and 4from the others.

The median values of each variable and the statistical decision treewere then used to create a simpler decision tree, presented in Table 3. These tables show that by using only 5 clinical tests,meibomian gland drop out, lipid volume, Schirmer test value, evaporation,and lipid viscosity, subjects can be placed in their diagnostic group. Theonly test we used that is not routinely performed was evaporation, and thiswas required only to discriminate group 2a from group 2b and group 3a fromgroup 3b. All other subjects could be categorized without testing for evaporation. Table 3 also lists a clinical description foreach group that identifies its main characteristics.

We then compared the statistical classification tree with our previousclinically derived classification system that was based primarily on physiologicparameters, a dermatologic examination for rosacea, and a clinical history.The results of this comparison are presented in Table 4. Included also are the data on the presence of global ocularsurface symptoms. In this comparison, each subject could have had more than1 physiologic diagnosis, such as dry eye and obstructive meibomian gland dysfunction.

COMMENT

There have been many classification systems proposed to separate subjectswith blepharitis and dry eye into relevant groups.11,12 Mostof these have been based on examination techniques that were not easily quantifiableor comparable between physicians. They also relied on subjective symptomsand the clinical history of each subject. Our analysis creates a classificationsystem that is based only on measurable and objective parameters. It has beenour contention that most, if not all, subjects with blepharitis demonstratesome meibomian gland dysfunction, with or without dry eye.5,6 Ourcluster analysis validates this approach.

These results suggest that data can be used to classify subjects withsimilar characteristics that have clinical relevance and statistical validity.We believe this approach places ocular surface disease classification on astronger scientific footing. There are several clinical implications thatcan be drawn from this study. The study strongly validates the fundamentaldivision between subjects with gland drop out and those without. Meibomiangland drop out appears to be an important marker for a disease process thatis relevant to this group of subjects. The study also validates the fundamentalrole of lipid volume since we could make statistically valid divisions betweensubjects with increased lipid volume and those with normal or low lipid volume.It was on the basis of lipid volume that we previously proposed differentiatingpatients with meibomian gland dysfunction between those with obstructive meibomiangland dysfunction (low lipid volume) and those with seborrheic meibomian glanddysfunction (high lipid volume). The cluster classification system was derived,however, through an independent and objective analysis of the data withoutregard to our, or other, previous classification schemes. The result is notidentical to our past intuitive efforts, and we did not expect it to be. Itis, however, reasonably similar.

Groups 1 and 4 contained those subjects with obstructive meibomian glanddysfunction. Group 1, which we named Rosacea, Obstructive Meibomian GlandDysfunction, and Dry Eye, differs from group 4 by lipid volume and viscosity.Fourteen of the 19 subjects with rosacea fell into group 1, which was characterizedby gland drop out and high lipid volume. Group 4 was also characterized bygland drop out but they had low lipid volume. Thus, rosacea appears to createa fairly consistent set of physiologic conditions. These 2 groups representa primary disease process that is distinct from other forms of blepharitis.Their degree of dry eye is very high, possibly from evaporation or from cornealstimulation and secondary inhibition of the lacrimal gland.13 Manyof these subjects might, in other systems, be classified as having staphylococcalblepharitis. It is our contention that bacteria play a role in the developmentof some types of meibomian gland dysfunction since Staphylococcusaureus is found somewhat more frequently in patients with blepharitisand the quantity of bacteria in the lid margin of patients with blepharitisis increased in comparison with the normal population.1417 Everyeyelid, however, is colonized by bacteria. The mere presence of these bacteriais therefore not sufficient to constitute disease. The role of bacteria onthe lid margin in the development of blepharitis is most likely importantbut it is still not well understood. It appears, however, that blepharitisis not primarily a process of eyelid infection. It is, rather, a process thatprimarily involves meibomian gland function that is probably affected by bacterialaction. Our study demonstrates that a meibomian gland orientation can successfullycharacterize a large number of subjects with blepharitis on the basis of meibomiangland function without a separate category for staphylococcal blepharitis.

Group 2 included those subjects with seborrheic meibomian gland dysfunction,with or without dry eye. Previous studies have suggested that an excess oflipid, which is a defining characteristic of this group, was responsible fortheir generally low evaporative rate. In the present study, we were surprisedto discover there are actually 3 separate seborrheic groups. Only 1 of these,group 2a, had the expected high lipid volume and low evaporation rate. Thesecond group had a high lipid volume, normal viscosity, high evaporation,and dry eye. The third group appeared to have elements of both obstructiveand seborrheic disease because their high lipid volume was combined with ahigh lipid viscosity and evaporation. These patients would be classified asmixed by many other clinically based classification systems.

Dry eye without meibomian gland dysfunction is also a distinct entity,and these subjects were clustered in group 5. Half of those with Sjögrensyndrome were found in this group. The other groups were mostly variationsof normal, some with high evaporation and some with low evaporation. The analysisillustrates the difficulty that all classification systems have with the widespectrum of findings in normal individuals. There are several reasons forthis. Blepharitis and dry eye symptoms are very common in the general population,and some elements of meibomian gland dysfunction are frequently found in thepopulation that considers itself to be normal. More than half of groups 2aand 2b were normal subjects, and one third of groups 3b and 6 were normalsubjects as well. Analyzed differently, 25% of the normal subjects were allocatedinto group 3a and 23% were allocated into group 6 while 17% were allocatedinto group 1. Even the population of symptom-free women demonstrated a fairlywide range of variability in their physiologic measurements that overlappeddisease states. This is a good indication that symptoms are not tightly linkedto these physiologic measurements. Therefore, it is not surprising that wefound a few normal subjects in many of these groups. Our data describing theasymptomatic, normal population were based only on women, and this representsa potential limitation of our study. This was a sample of convenience basedon the much greater availability of women for study purposes.

This analysis uncovered data that may shed some light on groups to befound within the normal population. We found 2 groups with a relatively highpercentage of normal subjects with high evaporation (groups 3b and 6), whileother normal subjects (part of group 3a) had low evaporation. Thus, thereappear to be 2 distinct types of normal subjects, 1 with high evaporationand 1 with low. This may be explained by lipid composition. There is a reasonablelikelihood that evaporation is partly controlled by components of the lipidlayer. One component of particular interest is cholesterol ester, becauseShine and McCulley18 and Shine et al19 found a basic division in the general populationbetween those with low levels of cholesterol ester in their meibomian glandsecretions and those with high levels. This is the only consistent divisionthey have found in the normal population.

We found evaporation was a key variable in differentiating some clusters.This suggests that evaporation is a key process and that it plays an importantrole in ocular surface disease.1,2,20,21 Evaporationis not intrinsically difficult to measure but it requires a practiced technicianand specialized equipment that is not generally available. Few physicians,therefore, have much experience with this fundamental process, and its usehas been limited to research. Although we found no real substitute for itsmeasurement at the present time, without evaporation, subjects can still beusefully classified up to this final point. There are, however, fundamentaldifferences between groups that only evaporation can evaluate at present.

Evaporation may not be of critical importance in determining the careand therapy that patients need. This cluster analysis does not necessarilytranslate directly into a therapeutic algorithm and is not presented as such.The current analysis is exploratory by nature and should instead provide insightsinto ocular surface disease processes that lie behind the symptoms and thesubjective clinical manifestations that we usually use to identify and treatdry eye, blepharitis, and ocular surface disease. Physiologic parameters mostlikely permit clustering for reasons that are based on pathophysiologic mechanisms,although these remain poorly understood. A more complete explanation of glanddrop out, lipid viscosity, seborrhea, and high evaporation will lead to abetter understanding of ocular surface disease and more directed and effectivetreatments. The use of simple and objective measurements to classify subjectsalso aids this process because it permits the valid identification and comparisonof treatments between different clinical centers. This has been a major problemin dry eye and ocular surface disease research.

Back to top
Article Information

Correspondence: William D. Mathers, MD,Casey Eye Institute, Oregon Health and Science University, 3375 SW Terwilliger,Portland, OR 97201 (mathersw@ohsu.edu).

Submitted for Publication: May 20, 2003; finalrevision received May 6, 2004; accepted May 27, 2004.

Financial Disclosure: None.

Funding/Support: This study was supported inpart by a grant from Research to Prevent Blindness, New York, NY, and by grantRO1 EY10164 from the National Eye Institute, Bethesda, Md (Dr Mathers).

References
1.
Mathers  WDBinarao  GPetroll  M Ocular water evaporation and the dry eye: a new measuring device. Cornea 1993;12335- 340
PubMedArticle
2.
Mathers  WDDaley  TE Tear flow and evaporation in patients with and without dry eye. Ophthalmology 1996;103664- 669
PubMedArticle
3.
Mathers  WDLane  JAZimmerman  MB Tear film changes associated with normal aging. Cornea 1996;15229- 234
PubMedArticle
4.
Mathers  WDLane  JA Meibomian gland lipid, evaporation, and tear film stability. Sullivan  DADartt  DAMeneray  MAedsLacrimalGland, Tear Film, and Dry Eye Syndromes 2 New York, NY Plenum Press1998;349- 360
5.
Mathers  WDLane  JASutphin  JEZimmerman  MB Model for ocular tear film function. Cornea 1996;15110- 119
PubMedArticle
6.
Mathers  WDShields  WJSachdev  MSPetroll  WMJester  JV Meibomian gland dysfunction in chronic blepharitis. Cornea 1991;10277- 285
PubMedArticle
7.
Ward  JH Hierarchical grouping to optimize an objective function. J Am Stat Assoc 1963;58236- 244Article
8.
Cleveland  WSGrosse  EShyu  WM Local regression models. Chambers  JMHastie  TJedsStatisical Modelsin S. New York, NY Wadsworth & Brooks/Cole1991;
9.
Breiman  LFriedman  JHOlsen  RStone  CJ Classification and Regression Trees.  New York, NY Wadsworth & Brooks/Cole1984;
10.
Ihaka  RGentleman  R R: a language for data analysis and graphics. J Comput Graph Statist 1996;5299- 314
11.
Thygeson  P The etiology and treatment of blepharitis: a study in military personnel. Arch Ophthalmol 1946;98191- 203
12.
McCulley  JPDougherty  JMDeneau  DG Classification of chronic blepharitis. Ophthalmology 1982;891173- 1179
PubMedArticle
13.
Mathers  W Why the eye becomes dry: a cornea and lacrimal gland feedback model. CLAO J 2000;26159
PubMed
14.
Tetz  MRKlein  UVolcker  HE  et al.  Staphylococci-associated blepharo-kerato-conjunctivitis [in German]. Ophthalmologe 1997;94186- 190
PubMedArticle
15.
Seal  DFicker  LRamakrishnan  MWright  P Role of staphylococcal toxin production in blepharitis. Ophthalmology 1990;971684- 1688
PubMedArticle
16.
McCulley  JPDougherty  JM Bacterial aspects of chronic blepharitis. Trans Ophthalmol Soc U K 1986;105314- 318
PubMed
17.
Dougherty  JMMcCulley  JP Comparative bacteriology of chronic blepharitis. Br J Ophthalmol 1984;68524- 528
PubMedArticle
18.
Shine  WEMcCulley  JP The role of cholesterol in chronic blepharitis. Invest Ophthalmol Vis Sci 1991;322272- 2280
PubMed
19.
Shine  WESilvany  RMcCulley  JP Relation of cholesterol-stimulated Staphylococcus aureus growth tochronic blepharitis. Invest Ophthalmol Vis Sci 1993;342291- 2296
PubMed
20.
Tsubota  KYamada  M Tear evaporation from the ocular surface. Invest Ophthalmol Vis Sci 1992;332942- 2950
PubMed
21.
Mathers  WD Ocular evaporation in meibomian gland dysfunction and dry eye. Ophthalmology 1993;100347- 351
PubMedArticle
×