Abstract
Purpose :
To develop an interpretable statistical model involving optimally selected spectral-domain optical coherence (SD-OCT) imaging biomarkers related to conversion time from non-neovascular to neovascular age-related macular degeneration (AMD)
Methods :
The HARBOR dataset, with 686 fellow eyes and time-varying SD-OCT images, was used for this study. Twenty-one drusen imaging features were extracted at each visit.
A multivariable joint model was implemented which consists of two parts fitted jointly- (1) nonparametric mixed effect models using natural cubic splines of ‘time’ and demographic factors to smoothen longitudinal trajectories of standardized imaging biomarkers, (2) a time-to-conversion model with 5 selected single nucleotide polymorphisms (fixed covariates) and smoothened trajectories of the longitudinal biomarkers (time-dependent covariates). For (2), 3 different models were considered: (a) a model with longitudinal biomarker values, (b) rate of changes in biomarker trajectories, (c) and area under trajectory (AUT).
A Bayesian regularized approach was used for selection among the 21 biomarkers in this multivariable model. A variable importance (VI) measure was defined as the probability of being over .25 in the absolute magnitude of the regularized coefficients of longitudinal variables in the standardized scale.
Results :
The top drusen biomarkers based on VI for 3 models were as follows (posterior mean, VI):
(a) value: area (.21, .33) and volume (.16, .26) within 3mm of fovea;
(b) rate of change: area within 3mm (-4.2, 1) and 5mm (-1.5, .54), volume within 3mm (5.8, 1) of fovea, total area within the field of view (FOV) (-.28, .5);
(c) AUT: area within 3mm of fovea (.12, .18), mean area in FOV (.08, .13).
Conclusions :
High values and high AUT of drusen area and volume within 3mm of the fovea expedite the conversion to neovascular AMD. However, variations in time for area and volume yield opposite interpretations. Low variations in drusen area over time expedite the conversion, whereas high variations in drusen volume do not. Additionally, as the distance from fovea increases, the impact of drusen area and volume dies out as we observe at 5mm distance, where only drusen area has an impact through its variations over time. Lastly, aggregate and mean area in FOV delay the conversion with higher variations and lower AUT, respectively.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.