Modern methodological approaches Sample Clauses

Modern methodological approaches. Deep learning ANNs are frequently used for prediction of output features - especially in the context of image classification [28, 29]. Applying deep learning methodology to medical survival data, however, poses the risk of overfitting, as the available sample sizes are typically small. In 2019, Xxxxxx et al. predicted survival by using a deep neural network (DNN) with a hierarchical structure and FFANNs in the first layers of the model [30]. The DNN contains 2 sub-networks with fully connected layers to jointly optimize the C-index and Mean Absolute Error (MAE). For each sub-network, the optimization is performed separately. The C-index quantifies the probability that the predicted event times of two randomly selected individuals have the same order as their true event times. Due to the presence of censored data not all pairs can be compared; this implies that a pair of subjects are comparable if the earliest time is an event, or both are events. The C-index is a measure of probability of concordance between the observed and the predicted survival. The MAE is defined as the absolute difference between the observed survival time and the survival time predicted by the sub-network. The authors found that the DNN performance improved on inclusion of more clinical features (input variables). A drawback of XXXx is that they are frequently computationally intensive and can be too complex for clinical insights. In 2020, Xxxx Xxx et al. developed time-binned neural networks to predict recurrence-free survival of non-small- cell lung cancer after surgery, using 30 clinico-pathological features [31]. The authors present one supervised learning binned-time survival analysis model (called s-DeepBTS) and one semi-unsupervised learning model (called su-DeepBTS). Here, we focus only on the supervised learning model s-DeepBTS. This is a shallow network where the output layer provides the survival probability in each pre-defined time interval (recurrence-free survival in months). The output value, yj, is 1 when a patient is alive without relapse at the beginning of the jth time interval Ij, and 0 after relapse. For censored patients, yj is 1 until a patient is lost to follow-up and Q i=ti≤Ij 1−di after ni censoring occurs (Xxxxxx-Xxxxx survival probability), where ni is the total number of samples without recurrence at the beginning of the jth time interval, and di is the number of events. The activation function of the output layer is the sigmoid (logistic). The root mean ...