Purchase this article with an account.
Theodore Leng, Luis de Sisternes, Qiang Chen, Jeffrey Ma, Vibha Mahendra, Daniel Rubin; Automated prediction of AMD progression from quantified SD-OCT images. Invest. Ophthalmol. Vis. Sci. 2013;54(15):4150.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Identifying which patients with non-exudative (dry) age-related macular degeneration (AMD) will progress to the exudative (wet) form is crucial, since prompt treatment can greatly improve visual outcomes. Currently, no robust predictive models of progression exist based on spectral domain optical coherence tomography (SD-OCT). We developed a predictive model using a series of quantitative features automatically extracted from SD-OCT images, as well as historical and demographic information.
2146 longitudinal SD-OCT scans obtained over 5 years from 350 eyes of 178 patients with AMD were included in this retrospective study. A predictive model for AMD progression using a fully automated algorithm to segment the retinal pigment epithelium and drusen and to extract quantitative image features from them was generated using generalized linear regression with a Poisson distribution. Based on this model, a predictive score for AMD progression at 6 and 12 months was obtained at each scan time point. This predictive score considered quantitative features characterizing AMD extracted from all available previous SD-OCT scans, status of the fellow eye (dry or wet), age and gender. Using these scores, a threshold was identified that differentiated patients in two subgroups of risk of AMD progression at future times. The predictive value of developing wet AMD at 6 and 12 months were compared to current practice standards.
Of the 350 eyes, 222 remained dry, 106 were wet, and 22 showed progression from dry to wet during the study period. Receiver operating characteristic curves generated using the model showed increased prediction sensitivity (rate of progressing cases) as the specificity decreased (rate of non-progressing cases that were correctly identified). Selecting the point in which more than 95% of cases progressed (0.95 sensitivity), the model identified a higher percentage of patients who progressed in less than a year (24.7%) than using current AREDS classification, while maintaining 0% of missed cases of progression in the group of patients predicted to be low risk.
Considering that only 2.5% and 0.9% of the total cases evaluated progressed within 12 and 6 months, the model was able to identify ten times as many patients (24.7%) within one year and more than 25 times as many patients within 6 months who progressed than if patients were evaluated according to current practices.
This PDF is available to Subscribers Only