Comparison of Analytic Algorithms
for
Detecting
Glaucomatous Visual Field Loss Joanne Katz,
MS; Alfred Sommer, MD; Douglas E. Gaasterland, MD; Douglas R. Anderson, MD
\s=b\ The
sensitivity and specificity of alanalytic strategies for recognizing glaucomatous visual field loss from automated threshold perimetry (C-30-2 test of the Humphrey Field Analyzer) were compared among one eye each of 106 patients with glaucoma and 249 normal subjects. Algorithms included commercially available global indexes and cross-meridional differences (Statpac 1 and Statpac 2), as well as cross-meridional and cluster analyses that were developed independently for natural history studies and clinical trials. The sensitivity of most algorithms was high, except for those that used only diffuse loss as an indicator of abnormality. Specificity was acceptably high for all algorithms. Subjects who failed to meet the manufacturer's standard for reliability had much reduced specificity, but sensitivity was also affected. Algorithms that were ternate
based on any of the alternate definitions of localized reduction in retinal sensitivity performed equally well, which suggests that any of these approaches is useful in searching for glaucomatous visual loss as typified by this database. Availability, familiarity, and convenience may govern the selection of any one analytic approach for use in a particular
setting. (Arch Ophthalmol. 1991;109:1684-1689)
Accepted for publication April 22,
1991. From the Dana Center for Preventive Ophthalmology, Wilmer Eye Institute and the School of Hygiene and Public Health, The Johns Hopkins University, Baltimore, Md (Ms Katz and Dr Sommer), the Center for Sight, Georgetown University, Washington, DC (Dr Gaasterland), and the Bascom Palmer Eye Institute, Department of Ophthalmology, University of Miami (Fla) School of Medicine (Dr Anderson). Reprint requests to Wilmer Eye Institute, Room 120, The Johns Hopkins Hospital, 600 N Wolfe St, Baltimore, MD 21205 (Ms Katz).
utomated static threshold
peri-
used to iden¬ "^metry is increasingly and visual field
tify glaucomatous
loss,
it may detect visual field defects at an earlier stage than manual testing.1"7 One widely used test is the C-30-2 program of the Humphrey field analyz¬ er, which tests an evenly spaced grid of 76 points in the central 30° field.
See also
1690.
Several analytic strategies have been developed independently for recogniz¬ ing glaucomatous visual field loss from the quantitative output of such pro¬ grams.8"16 The Humphrey field analyzer provides global visual field indexes and
associated values that aid in the detection of field loss.1617 The deviation from age-corrected normal sensitivity and the values associated with indi¬ vidually depressed points are also pro¬ vided by the Humphrey Statpac out¬ put in the form of probability maps.18 To provide entry criteria into clinical trials, glaucomatous field loss is typi¬ cally defined by some criterion, eg, a certain-sized cluster of points for which sensitivity is depressed by more than a predefined decibel amount or probability level when compared with that expected for normal subjects of a similar age. The latest version of the Humphrey Statpac software also pro¬ vides an analysis based on differences of probability scores across the hori¬ zontal field meridian.19 This is similar to an algorithm based on cross-meridi¬ onal differences in threshold values that were developed during a natural history study of glaucoma.14,15 The availability of several researchgenerated, as well as commercial algo¬ rithms, has prompted us to evaluate the sensitivity and specificity of these alternate interpretive strategies from
Downloaded From: http://archopht.jamanetwork.com/ by a Yale University User on 05/18/2015
database of normal subjects and patients with glaucoma (elevated intraocular pressure and visual field defects previously documented by de¬ tailed manual perimetry). a common
SUBJECTS AND METHODS
Subjects who were enrolled in the Glau¬ Screening Study (GLASS) at the Wil¬ mer Eye Institute, Baltimore, Md, provid¬ ed the population sample for comparing the various visual field algorithms. This study has served as the basis for designing and evaluating visual field algorithms for the diagnosis of glaucoma.20. On enrollment into the study, threshold-related suprathreshold static and kinetic perimetry was per¬ formed on all subjects.20 A detailed medical and ocular history was taken, and each subject underwent a comprehensive ocular examination. Subjects were classified as normal if they had intraocular pressures of 21 mm Hg or lower, entirely normal visual fields on manual perimetry, and no personal or family history of glaucoma. Both eyes had to meet this definition. Subjects were considered to have glaucoma if their intrao¬ cular pressures were consistently greater than 21 mm Hg and if they had at least one visual field defect on manual perimetry that was reproducible on a second manual test. coma
A visual field defect was defined as a para¬ central or full arcuate scotoma at least 0.4 log units deep or a nasal step at least 10° wide and present to at least two isopters. If one eye met these criteria, the patient was considered to have glaucoma. Patients with hemispheric or greater loss were not eligi¬ ble. Thus, a defect that was found by manual perimetry and accompanied by ele¬ vated intraocular pressure was considered to be the "gold standard" for the diagnosis of open angle glaucoma. From 1984, new recruits and subjects who were already enrolled began visual field testing with the C-30-2 program of the Humphrey Field Analyzer (Humphrey Al¬ lergan, San Leandro, Calif), but manual perimetry remained the standard for diag¬ nosis of disease in subjects. The first auto-
mated visual field test session was evalu¬ ated in this report. Among normal subjects, the first eye that was tested in the session was selected for analysis. Among those with glaucoma, the first eye that was tested in the session was selected if both eyes had glaucoma. If not, the eye with glaucoma was chosen, even if it was the second eye that was tested. Although some subjects had not previously undergone automated testing, all were familiar with manual
perimetry. The Statpac 1 software of the Humphrey field analyzer provides a total deviation and a pattern deviation plot.16'17 The total devi¬ ation plot displays the difference between the observed threshold and that expected
subject of the same age at each location, except for two in the region of the blind spot. The pattern deviation plot is similar to the total deviation plot, except for a normal
that the difference between the observed and expected threshold values is adjusted for the sensitivity level in the most normal region of the visual field. Probability maps are printed below each deviation plot. The values associated with each location for which the deviation is "outside normal lim¬ its" are indicated on these plots. A series of "global" indexes are also printed on the Statpac output. The mean deviation (MD) is a variance-weighted average difference of observed threshold sensitivity from the age-specific normal sensitivity. A large neg¬ ative MD is a measure of overall abnormali¬ ty of the visual field. The pattern standard deviation (PSD) is a weighted SD of the differences of observed threshold from the age-specific normal thresholds. A large PSD suggests the presence of localized de¬ fects because the variance of retinal sensi¬ tivity at different locations is greater than expected for a normal subject of the same age. The short-term fluctuation represents
test-retest variability as a weighted mean of the SD of repeated measurements, as estimated at 10 test locations where thresh¬ olds are determined twice. A large shortterm fluctuation indicates a high degree of test-retest variability. The corrected pat¬ tern standard deviation (CPSD) estimates the extent of localized visual field defects, adjusted to reflect only irregularities in the field that are greater than those generated by measurement error alone. Subjects who constituted the manufacturer's normal da¬ tabase were experienced in automated peri¬ metry and "reliable" (
for
Detecting
Glaucomatous Visual Field Loss Joanne Katz,
MS; Alfred Sommer, MD; Douglas E. Gaasterland, MD; Douglas R. Anderson, MD
\s=b\ The
sensitivity and specificity of alanalytic strategies for recognizing glaucomatous visual field loss from automated threshold perimetry (C-30-2 test of the Humphrey Field Analyzer) were compared among one eye each of 106 patients with glaucoma and 249 normal subjects. Algorithms included commercially available global indexes and cross-meridional differences (Statpac 1 and Statpac 2), as well as cross-meridional and cluster analyses that were developed independently for natural history studies and clinical trials. The sensitivity of most algorithms was high, except for those that used only diffuse loss as an indicator of abnormality. Specificity was acceptably high for all algorithms. Subjects who failed to meet the manufacturer's standard for reliability had much reduced specificity, but sensitivity was also affected. Algorithms that were ternate
based on any of the alternate definitions of localized reduction in retinal sensitivity performed equally well, which suggests that any of these approaches is useful in searching for glaucomatous visual loss as typified by this database. Availability, familiarity, and convenience may govern the selection of any one analytic approach for use in a particular
setting. (Arch Ophthalmol. 1991;109:1684-1689)
Accepted for publication April 22,
1991. From the Dana Center for Preventive Ophthalmology, Wilmer Eye Institute and the School of Hygiene and Public Health, The Johns Hopkins University, Baltimore, Md (Ms Katz and Dr Sommer), the Center for Sight, Georgetown University, Washington, DC (Dr Gaasterland), and the Bascom Palmer Eye Institute, Department of Ophthalmology, University of Miami (Fla) School of Medicine (Dr Anderson). Reprint requests to Wilmer Eye Institute, Room 120, The Johns Hopkins Hospital, 600 N Wolfe St, Baltimore, MD 21205 (Ms Katz).
utomated static threshold
peri-
used to iden¬ "^metry is increasingly and visual field
tify glaucomatous
loss,
it may detect visual field defects at an earlier stage than manual testing.1"7 One widely used test is the C-30-2 program of the Humphrey field analyz¬ er, which tests an evenly spaced grid of 76 points in the central 30° field.
See also
1690.
Several analytic strategies have been developed independently for recogniz¬ ing glaucomatous visual field loss from the quantitative output of such pro¬ grams.8"16 The Humphrey field analyzer provides global visual field indexes and
associated values that aid in the detection of field loss.1617 The deviation from age-corrected normal sensitivity and the values associated with indi¬ vidually depressed points are also pro¬ vided by the Humphrey Statpac out¬ put in the form of probability maps.18 To provide entry criteria into clinical trials, glaucomatous field loss is typi¬ cally defined by some criterion, eg, a certain-sized cluster of points for which sensitivity is depressed by more than a predefined decibel amount or probability level when compared with that expected for normal subjects of a similar age. The latest version of the Humphrey Statpac software also pro¬ vides an analysis based on differences of probability scores across the hori¬ zontal field meridian.19 This is similar to an algorithm based on cross-meridi¬ onal differences in threshold values that were developed during a natural history study of glaucoma.14,15 The availability of several researchgenerated, as well as commercial algo¬ rithms, has prompted us to evaluate the sensitivity and specificity of these alternate interpretive strategies from
Downloaded From: http://archopht.jamanetwork.com/ by a Yale University User on 05/18/2015
database of normal subjects and patients with glaucoma (elevated intraocular pressure and visual field defects previously documented by de¬ tailed manual perimetry). a common
SUBJECTS AND METHODS
Subjects who were enrolled in the Glau¬ Screening Study (GLASS) at the Wil¬ mer Eye Institute, Baltimore, Md, provid¬ ed the population sample for comparing the various visual field algorithms. This study has served as the basis for designing and evaluating visual field algorithms for the diagnosis of glaucoma.20. On enrollment into the study, threshold-related suprathreshold static and kinetic perimetry was per¬ formed on all subjects.20 A detailed medical and ocular history was taken, and each subject underwent a comprehensive ocular examination. Subjects were classified as normal if they had intraocular pressures of 21 mm Hg or lower, entirely normal visual fields on manual perimetry, and no personal or family history of glaucoma. Both eyes had to meet this definition. Subjects were considered to have glaucoma if their intrao¬ cular pressures were consistently greater than 21 mm Hg and if they had at least one visual field defect on manual perimetry that was reproducible on a second manual test. coma
A visual field defect was defined as a para¬ central or full arcuate scotoma at least 0.4 log units deep or a nasal step at least 10° wide and present to at least two isopters. If one eye met these criteria, the patient was considered to have glaucoma. Patients with hemispheric or greater loss were not eligi¬ ble. Thus, a defect that was found by manual perimetry and accompanied by ele¬ vated intraocular pressure was considered to be the "gold standard" for the diagnosis of open angle glaucoma. From 1984, new recruits and subjects who were already enrolled began visual field testing with the C-30-2 program of the Humphrey Field Analyzer (Humphrey Al¬ lergan, San Leandro, Calif), but manual perimetry remained the standard for diag¬ nosis of disease in subjects. The first auto-
mated visual field test session was evalu¬ ated in this report. Among normal subjects, the first eye that was tested in the session was selected for analysis. Among those with glaucoma, the first eye that was tested in the session was selected if both eyes had glaucoma. If not, the eye with glaucoma was chosen, even if it was the second eye that was tested. Although some subjects had not previously undergone automated testing, all were familiar with manual
perimetry. The Statpac 1 software of the Humphrey field analyzer provides a total deviation and a pattern deviation plot.16'17 The total devi¬ ation plot displays the difference between the observed threshold and that expected
subject of the same age at each location, except for two in the region of the blind spot. The pattern deviation plot is similar to the total deviation plot, except for a normal
that the difference between the observed and expected threshold values is adjusted for the sensitivity level in the most normal region of the visual field. Probability maps are printed below each deviation plot. The values associated with each location for which the deviation is "outside normal lim¬ its" are indicated on these plots. A series of "global" indexes are also printed on the Statpac output. The mean deviation (MD) is a variance-weighted average difference of observed threshold sensitivity from the age-specific normal sensitivity. A large neg¬ ative MD is a measure of overall abnormali¬ ty of the visual field. The pattern standard deviation (PSD) is a weighted SD of the differences of observed threshold from the age-specific normal thresholds. A large PSD suggests the presence of localized de¬ fects because the variance of retinal sensi¬ tivity at different locations is greater than expected for a normal subject of the same age. The short-term fluctuation represents
test-retest variability as a weighted mean of the SD of repeated measurements, as estimated at 10 test locations where thresh¬ olds are determined twice. A large shortterm fluctuation indicates a high degree of test-retest variability. The corrected pat¬ tern standard deviation (CPSD) estimates the extent of localized visual field defects, adjusted to reflect only irregularities in the field that are greater than those generated by measurement error alone. Subjects who constituted the manufacturer's normal da¬ tabase were experienced in automated peri¬ metry and "reliable" (
