Abstract
Purpose :
Covariate-adjusted clinical trial analysis can improve power and decision making. Here, we report the development, validation, and comparison of different models to predict GA progression and quantify their impact on power.
Methods :
This study was performed using data from patients in the lampalizumab program. The outcome was defined as growth rate (mm2/year) of GA lesion area (measured on fundus autofluorescence [FAF]). The data from the Proxima A (NCT02479386), Chroma (NCT02247479), and Spectri (NCT02247531) trials were split into development (n = 1279) and holdout (n = 443) sets. Three different classes of models were developed: (1) benchmark model: a baseline simple feature-based model using GA area, location, contiguity, and low-luminance deficit; (2) run-in model: uses a 6-month initial lesion growth rate with or without clinical features; and (3) imaging model: a baseline deep learning model using FAF images. These models were validated on 2 additional datasets that were not used for model development: MAHALO (NCT01229215; n = 122) and Proxima B (NCT02479386; n = 175). Performance was evaluated by calculating the squared Pearson correlation coefficient (r2). Additionally, the effective sample size increase (ESSI) that would result from using the respective model for adjustment in the clinical trial analysis was calculated. ESSI was calculated relative to no adjustment or adjustment with the benchmark model.
Results :
The FAF-based imaging model outperformed the feature-based benchmark and run-in models in all datasets, with an r2 of 0.48 on the holdout set, 0.63 on the MAHALO dataset, and 0.48 on the Proxima B dataset. This translates into ESSIs of 92%, 174%, and 91%, respectively, compared with an unadjusted analysis. Compared with the benchmark model, the imaging model provides an additional effective increase in sample size of 61%, 82%, and 43%, respectively.
Conclusions :
These findings suggest that the FAF-based imaging model, using a single image at baseline, is robust and shows a higher prognostic performance than models with simple features derived at baseline and during a run-in phase. These results suggest that the imaging models can significantly improve trial power and decision making.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.