- X_initial is observed variables of each patients at the first visit. Targets include time and event. Time target (T) is a continuous variable representing follow-up months
- m is the time window or the interval time, which is a tunable hyper-parameter (In this study, the interval time was 10 month)
- ti is the last observed time during the time interval i
- Survival target (E) is a binary value representing event (alive: 0, death: 1). Ei is a binary event during the time interval i. The neural network is trained with the targets Y[Ei, ti] recurrently at each time point i using cosine distance as a loss function. (훂 is a hyperparameter representing weight of death.) For example, E1 = 0 at the first time window could be E2 = 0, 1, or censored at the second window, thus the neural network should adjust their parameter to the following targets. After their serial training, the network learned to perceive severity of the cancer patient
