Estimating Calibrated Individualized Survival Curves with Deep Learning

In survival analysis, deep learning approaches have been proposed for estimating an individual's probability of over some time horizon. Such can capture complex non-linear relationships, without relying on restrictive assumptions regarding the relationship between characteristics and their underlying process. To date, however, these methods focused primarily optimizing discriminative performance ignored model calibration. Well-calibrated curves present realistic meaningful probabilistic estimates true process individual. However, due to lack ground-truth stochastic individual, measuring calibration in analysis is inherently difficult task. this work, we i) highlight shortcomings existing terms ii) propose a new training scheme models that maximizes performance, subject good Compared state-of-the-art across two publicly available datasets, our leads significant improvements calibration, while maintaining performance.