Purchase this article with an account.
Wolf-Dieter Vogl, Sebastian M Waldstein, Bianca S. Gerendas, Jing Wu, Alessio Montuoro, Ana-Maria Glodan, Dominika Podkowinski, Christian Simader, Ursula Schmidt-Erfurth, Georg Langs; Population-wide Disease Modeling to Predict Macular Thickness and Treatment Response in Longitudinal OCT Data. Invest. Ophthalmol. Vis. Sci. 2015;56(7 ):2767.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
The ability to predict the future development of a disease is an extremely important goal both in clinical and scientific practice, as well as in the general health care setting. We propose a novel method which predicts the longitudinal disease course as well as treatment response patterns based on inter-patient aligned longitudinal spectral domain optical coherence tomography (SD-OCT) images within a disease population.
Baseline OCT scans and 12 monthly follow-up scans of 44 patients with central (CRVO) were included. All patients received initial ranibizumab injections for three months followed by a PRN regimen. All scans were transformed into a joint coordinate system by aligning the optical disk center and the foveal center. Total retinal thickness maps were computed using the Iowa reference algorithm. In a 5-fold cross-validation setting, several sparse and non-linear regression models were learned from the pixel-wise thickness values using the common loading phase (i.e. baseline - month 2) in the training set. On the test set retinal thickness map was predicted for month 3. Furthermore, we predicted the recurrence/non-recurrence of macular edema within the 12 month follow-up period, where 6 of 44 patients showed no recurrence.
From the predicted thickness maps the mean absolute error (MAE) in μm was computed. As regression methods elastic net, Lasso, ridge regression and random forest regression were evaluated. Elastic net showed the lowest median [mean/std] MAE of 14.32 [19.21/16.54] and a R² of 0.48, followed by Lasso with 15.87 [20.50/16.67] R² of 0.45, and ridge regression with a MAE of 15.82 [21.33/16.96] and a R² of 0.43. Finally, random forest had the lowest score with an MAE of 18.77 [23.95/16.93] and R² of 0.32.<br /> Logistic regression with elastic net regularization is the most sensitive method for the classification of recurrence/non-recurrence with a non-recurrence sensitivity/specificity of 1.00/0.90, followed by random forest with 0.86/0.98, and lastly sparse cox regression with 0.83/0.92.
This study demonstrates that precise prediction of future disease development is possible using longitudinal OCT data transformed to a joint coordinate space and sparse machine learning methods.
This PDF is available to Subscribers Only