Machine Learning Approach for Analyzing Mixed Case Interval Censored Data with a Cured Subgroup.

We introduce a novel two-component framework for analyzing mixed case interval censored (MCIC) data featuring a cured subgroup. In such data, the time-to-event is known only within certain intervals determined by multiple random examination time points. Moreover, a portion of the subjects will never experience the event. The first component of our model focuses on estimating the likelihood of being cured (incidence), departing from the conventional generalized linear model to adopt a more adaptable support vector machine (SVM) approach capable of accommodating complex or non-linear covariate effects. The second component addresses the survival distribution of the uncured individuals (latency) and employs a Cox proportional hazards structure to maintain the straightforward interpretation of covariate effects. We develop an expectation maximization algorithm, incorporating the Platt scaling method, to estimate the probability of being cured. Our simulation study demonstrates that our model outperforms both logit-based and spline-based models in capturing complex classification boundaries, leading to more accurate estimates of cured/uncured probabilities and enhanced predictive accuracy for cure. We emphasize that enhancing the estimation accuracy regarding incidence subsequently improves the estimation outcomes concerning latency. Finally, we illustrate the efficacy of our methodology by applying it to the NASA's Hypobaric Decompression Sickness Data.