Purchase this article with an account.
Jean-Claude Mwanza, Joshua Warren, Donald Budenz; Logistic Regression Model for Principal Factor Analysis Combining Spectral Domain OCT Structural Parameters for Detection of Early Glaucoma. Invest. Ophthalmol. Vis. Sci. 2013;54(15):1463.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To create a multivariate predictive model for early glaucoma using a combination of optic nerve head (ONH), peripapillary retinal nerve fiber layer (RNFL), and macular ganglion cell-inner plexiform layer (GCIPL) parameters measured with spectral domain OCT, to determine its sensitivity, and to compare the results with single variable models.
A modeling set including 57 eyes with early glaucoma and 99 normal eyes underwent peripapillary and macular scanning using Cirrus OCT. Sixteen parameters were analyzed: rim and cup areas, and vertical cup-to-disc (VCDR) ratio for ONH; average and quadrant RNFL thicknesses; average, minimum, and sectoral GCIPL thicknesses. These parameters were submitted to a principal factor analysis (PFA), after which a stepwise logistic regression model was fitted with glaucoma as the outcome variable and the identified factors as candidate explanatory variables. From the estimated logistic regression coefficients, predicted probabilities were calculated and submitted to an ROC curve analysis. Similar logistic regression and ROC curve analyses were carried out for single variable models. The sensitivity of the models was tested in a separate set including 46 eyes with early glaucoma and 51 normal eyes for validation.
PFA identified five latent factors which accounted for 94.87% of the variability seen in the original set of variables: all GCIPL parameters (factor 1), all ONH parameters (factor 2), average, superior and inferior RNFL (factor 3), temporal RNFL (factor 4), and nasal RNFL (factor 5). The multivariate model had an area under the curve (AUC) of 0.995 for 98.2% sensitivity, 93.9% specificity, and an Akaike information criterion (AIC) value of 34.02. Single variable models (VCDR, minimum GCIPL, inferior RNFL) yielded AUCs of 0.943-0.974, sensitivities of 86.0%-96.5%, specificities of 89.9%-90.9%, and AICs of 64.87-88.36. The PFA logistic regression model correctly classified 90.72% of cases with a median 95% prediction interval length (MPIL) of 0.016 in the validation set. Single variable models correctly classified 74.22%-89.69% of cases with MPILs 5 to 7 times higher.
These five factors were successful in predicting early glaucoma status. The multivariate model outperformed single variable models both in terms of AUC, AIC, MPLIs, and classification rates.
This PDF is available to Subscribers Only