Interpretable Deep Regression Models With Interval-Censored Failure Time Data

Area of Science:

• Statistics
• Machine Learning
• Biostatistics

Background:

• Deep neural networks (DNNs) excel at modeling complex data but are underexplored for interval-censored survival data.
• Existing deep learning methods for survival analysis primarily address right-censored data, limiting their application to interval-censored scenarios.

Purpose of the Study:

• To propose a general regression framework for interval-censored data using partially linear transformation models.
• To integrate deep neural networks (DNNs) for modeling nonlinear effects of nuisance covariates, balancing interpretability and flexibility.

Main Methods:

• Utilized sieve maximum likelihood estimation with monotone splines to approximate the cumulative baseline hazard function.
• Developed an Expectation-Maximization (EM) algorithm incorporating stochastic gradient descent for reliable and tractable estimation.
• Established asymptotic properties of parameter estimators, demonstrating minimax-optimal convergence for the DNN estimator.

Main Results:

• The proposed DNN estimator achieves superior estimation and prediction accuracy compared to state-of-the-art methods in extensive simulations.
• The framework successfully applied to the Alzheimer's Disease Neuroimaging Initiative dataset, yielding novel insights.
• Demonstrated improved predictive performance on real-world data over traditional survival analysis techniques.

Conclusions:

• The developed deep learning framework offers a flexible and powerful approach for analyzing interval-censored survival data.
• This method enhances understanding and prediction in complex datasets, particularly in biomedical research.
• The study highlights the potential of DNNs in advancing survival analysis for challenging data types.