Purpose
To present a pilot evaluation of the prediction accuracy of a novel dynamic structure-function (DSF) model for monitoring the progression of glaucoma.
Methods
Longitudinal series of paired mean sensitivity (MS) and neuroretinal rim area (RA) from 164 eyes of 114 patients with suspected or diagnosed primary open-angle glaucoma enrolled in the Diagnostics Innovations in Glaucoma Study or the African Descent and Glaucoma Evaluation Study were included. MS and RA were expressed as percent normal based on mean values derived from an independent dataset of 91 eyes from 91 healthy controls (Racette et al, J Glaucoma, 2007; 16:676-84). The DSF model uses centroids and velocity vectors to assess glaucomatous progression. The centroids are the central location for the cloud of paired structural and functional data points and estimate the current state of the disease in each patient. The velocity vectors represent the trend at which structure and function are jointly changing. The fourth, fifth, and sixth longitudinal pairs of MS and RA were predicted from the first three, four, and five pairs, respectively. The prediction accuracy of the DSF model was compared to that of the ordinary least squares linear regression (OLSLR) using Wilcoxon signed-rank test on root-mean-square prediction error (RMSPE).
Results
For the prediction of the fourth and fifth pairs, the DSF model was on average more accurate than OLSLR (median RMSPE was lower for the DSF model by 3.2% and 2.0%, respectively, p < 0.0001). For the prediction of the sixth pair, median RMSPE was only slightly lower for the DSF model than OLSLR (by 0.5%, p = 0.33). The DSF model was more accurate than OLSLR for 74%, 64%, and 55% of the subjects on the predictions for the fourth, fifth, and sixth pairs, respectively. The absolute prediction differences between the two models were lower than 10% for about 80% of the subjects.
Conclusions
Overall, the two models have the similar prediction accuracy and the DSF model performs better than OLSLR in shorter time series. The results of this work are in agreement with those of Russell et al (Invest Ophthalmol Vis Sci, 2012;53:2760-9) in that OLSLR can be improved upon, particularly, when limited follow-up is available.
Keywords: 473 computational modeling