Linear regression modeling of rim area to discriminate between normal and glaucomatous optic nerve heads - The Bridlington Eye Assessment Project

Purpose: To create a diagnostic algorithm by modeling the normal variability of rim area (measured by Heidelberg Retina Tomography) using linear regression in normal elderly subjects. Methods: Multiple linear regression was performed between log rim area and disc area, age, and sex in 712 normal elderly subjects. The relationship between log rim area and disc area was not linear and showed significant heteroscedascity (increasing variability of rim area with increasing disc area). These factors violate the assumptions of linear regression as performed by the Moorfields Regression Analysis (MRA), and were overcome by conducting linear regression separately for each disc area quartile. Actual rim area in each disc sector was compared with the 95% lower limit predicted by the analysis in the normal subjects and 58 glaucoma patients. Results: Specificity and sensitivity of the New Regression Analysis (NRA) was 83% and 81%, respectively (84.4% and 82.7%, respectively for MRA). NRA specificity was unaffected by disc size (81.6% and 86.3% in the smallest and largest disc area quartiles, respectively, P 0.36). MRA specificity was reduced in larger discs (91.1% and 73.1% in smallest and largest disc area quartiles, P < 0.001), with reduced sensitivity in smaller discs (58.3% and 85.0% in smallest and largest quartiles, P = 0.05). Conclusions: Nonlinearity and heteroscedascity in the relationship between log rim area and disc area explain reduced specificity of the MRA in bigger discs as a statistical error. Although overall not offering better diagnostic performance, the NRA performed consistently across different disc sizes, offering better performance than the MRA in large discs.