Abstract
Purpose :
Train an algorithm predicting treatment response in patients with neovascular AMD (nAMD), treated with either 0.5mg monthly (SoC) or pro re nata (PRN) ranibizumab injections on features extracted from spectral domain optical coherence tomography (SD-OCT) images; Apply the resulting model to data from a different study; Identify the importance of baseline features.
Methods :
Baseline data from up to 274 patients with nAMD treated on SoC, as well as from up to 273 patients treated on a PRN basis from the HARBOR study (NCT00891735) were used. Intraretinal fluid (IRF), subretinal fluid (SRF), subretinal hypereflective material (SHRM), and pigment epithelial detachment (PED) were automatically segmented from a 6x6mm SD-OCT volume scan. Features included age, gender, baseline SD-OCT scans, and baseline Best Corrected Visual Acuity (BCVA), retinal layers, and fluidic features. Response to therapy was assessed as BCVA ≥20/40 in patients with baseline BCVA <20/40, and absolute BCVA. Five-fold crossvalidation was used to tune the algorithm, separately for SoC and PRN. A model using the best parameter combination was used to predict the outcome of 66 eyes under SoC from the AVENUE study (NCT02484690). Baseline feature values were assessed for impact on patient-level predictions.
Results :
Using HARBOR data, the best tuned performance for SoC was AUROC = 0.83 (95% CI 0.77 – 0.89) for 20/40 vision, and R-squared = 0.52 (95% CI 0.43 - 0.61) for absolute BCVA, both at month 12. For PRN, the respective numbers were significantly lower. Validation on AVENUE yielded comparable results: 0.79 (95% CI 0.67 – 0.91), and 0.45 (95% CI 0.23 - 0.64), respectively, at month 9. In HARBOR, occult CNV and baseline presence of SRF were associated with higher BCVA at month 12, while classic CNV and baseline presence of IRF and SHRM were associated with lower BCVA.
Conclusions :
The models were capable of predicting the probability of achieving BCVA >= 20/40, and absolute BCVA. Predictions for SoC were superior to PRN, likely due to the subjective criteria used in the latter. The SoC model showed robust performance on external study data, acquired with a different SD-OCT device. Occult CNV, absence of IRF, SHRM, and PED, and presence of SRF were positive baseline predictors for absolute BCVA. SRF presence was also a positive baseline indicator for relative gain.
This is a 2020 Imaging in the Eye Conference abstract.