A1237
Title: Machine learning-based mixture model
Authors: Suvra Pal - University of Texas at Arlington (United States) [presenting]
Abstract: A two-component mixture cure model (MCM) is considered. The first component of the MCM that describes the cure rate, or the incidence, is modeled using a machine learning (ML) based approach. The second component of the MCM that describes the survival distribution of the uncured, or the latency, is modeled using the Cox's proportional hazards structure. For the estimation of model parameters, an expectation maximization algorithm is developed in conjunction with the Platt scaling method which is used to convert the ML outputs to posterior probabilities of cure. When the classification boundary, with respect to classifying the cured and uncured units, is non-linear it is shown that the ML-based modeling results in more accurate and precise estimates of the cured probability when compared to the logistic regression-based modeling. This further results in improved predictive accuracy for cure. Interestingly, it is found that improving the estimation results related to the incidence also improves the latency estimation results. Finally, an application of the ML-based MCM is illustrated using a data on the lifetime of Kevlar 49 wrapped pressure vessels subject to different stress levels.
