Advances in deep learning allow to capture the information contained in histological images of cancer tissue. Histological images are larger than typical images processed by deep learning requiring tailored algorithms. Owkin has developed two such procedures, one relying only on information at the slide level and the other also leveraging annotations on the slide. These procedures were applied on digitized biopsies from mesothelioma (Courtiol et al. 2019) and resected HCC (Saillard et al. 2020) patients, and the resulting deep learning covariates have been validated as an independent predictor of overall survival (OS). 

Adjustment on prognostic covariates allows for improved precision and increased statistical power for treatment effect estimation in randomized controlled trials. In the specific setting of time-to-event outcomes, parametric and semi-synthetic simulations yield critical information on the factors impacting the reduction in sample size following covariate adjustment. We advocate for more systematic adjustment on prognostic covariates, which can lead to more efficient and more inclusive clinical trials.