Abstract:
|
This paper introduces a prognostic method called lights to deal with the problem of joint modeling of longitudinal data and censored durations, where a large number of longitudinal features are available. Yet there is no standard model so far to learn from such high-dimensional multivariate longitudinal data in a survival analysis setting. Features are extracted from the longitudinal processes and included as potential risk factor in a group-specific Cox model with high-dimensional shared associations. Appropriate penalties are then used during inference to allow flexibility in modeling the dependency between the longitudinal features and the event time. The statistical performance of the method is examined on a simulation study, and finally illustrated on publicly available datasets. We compare state-of-the-art survival models regarding risk prediction in terms of C-index. Our method provides powerful interpretability by automatically pinpointing significant features being relevant from a clinical perspective. Thus, we propose a powerful tool for personalized medicine, with the ability of automatically determining significant prognostic longitudinal markers.
|