Prediction Accuracy of a Dynamic Structure-Function (DSF) model for Glaucoma Progression using Contrast Sensitivity Perimetry (CSP) and Confocal Scanning Laser Ophthalmoscop

Abstract

Purpose: The DSF model was shown to have better prediction accuracy compared to linear regression in short follow-up series when the mean sensitivity (MS) of static automated perimetry (SAP) was paired with rim area (RA) (Hu et al, IOVS, 2014; In Press). CSP was shown to have lower test-retest variability than SAP in glaucomatous defects (Hot et al, IOVS, 2008; 49:3049−57). The current study assessed whether the prediction accuracy of the DSF model could be improved using CSP instead of SAP.

Methods: Longitudinal data from 36 eyes of 36 patients with open-angle glaucoma were analyzed. The first set of 18 patients had 6 pairs of structure-function data (each considered as a visit), and the second set of 28 patients had 4 pairs, with a maximum of one month between structural and functional measurements in each pair, and a minimum of 6 months between pairs. The structural parameter was RA (Heidelberg Retina Tomograph) and the functional parameter was MS obtained on CSP and SAP. RA and MS were expressed as percent of mean normal based on an independent dataset of 102 healthy eyes. The first 3 visits were used to predict the 4th visit, the first 4 visits to predict the 5th visit, and the first 5 visits to predict the 6^th visit. The median prediction error (PE) was compared to that of ordinary least squares linear regression (OLSLR) using the Wilcoxon signed-rank test.

Results: For CSP MS and RA in the first set of 18 patients, median PE with OLSLR for predicting visits 4, 5, and 6 was 5.2%, 5.2%, and 4.9% of mean normal, respectively. These values decreased by 2.0% of mean normal (p = .004), 1.5% (p = .006), and 1.0% (p = 0.85) with the DSF model. For SAP MS and RA, the median PE with OLSLR for predicting visits 4, 5, and 6 was 6.3%, 9.0%, 7.0%, respectively. These values decreased by 2.0% (p = .157) and 1.0% (p = .396) with the DSF model for visits 5 and 6, respectively. For visit 4, there was an increase of 1.4% (p = .184). This was confirmed on the second set of 28 patients: PE was 1.5 % lower for DSF with CSP (p = .002).

Conclusions: These results are in agreement with previous work (Hu et al, IOVS, 2014; In Press). The DSF model had lower prediction error than OLSLR for CSP MS compared to SAP MS, which might be partly explained by the reduced test-retest variability of CSP.