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

Area of Science:

• Biostatistics
• Survival Analysis
• Machine Learning

Background:

• Mixed case interval censored (MCIC) data presents unique challenges in survival analysis.
• Identifying a 'cured' subgroup, where subjects never experience the event, is crucial for accurate modeling.
• Existing methods may struggle with complex covariate effects or non-linear relationships.

Purpose of the Study:

• To introduce a novel two-component framework for analyzing MCIC data with a cured subgroup.
• To improve the estimation of cure probability (incidence) using a more flexible approach.
• To enhance the survival analysis of uncured individuals (latency) while maintaining interpretability.

Main Methods:

• A two-component model combining Support Vector Machines (SVM) for incidence and Cox proportional hazards for latency.
• Development of an expectation maximization algorithm with Platt scaling for cure probability estimation.
• Application to NASA's Hypobaric Decompression Sickness Data for validation.

Main Results:

• The proposed SVM-based incidence component effectively captures complex classification boundaries.
• The framework demonstrates superior performance compared to logit-based and spline-based models in simulations.
• Improved incidence estimation leads to enhanced latency estimation and predictive accuracy for cure.

Conclusions:

• The novel two-component framework offers a robust and accurate method for analyzing MCIC data with cure.
• The SVM approach provides flexibility in modeling incidence, outperforming traditional methods.
• Accurate incidence estimation is key to improving overall survival analysis outcomes in cured populations.