A robust neural network with random effects for subject-specific prediction of clustered count data

Area of Science:

• Biomedical data analysis
• Machine learning in healthcare
• Statistical modeling

Background:

• Traditional neural networks overlook dependencies in large-scale biomedical data due to high-cardinality categorical features.
• Subject-specific predictions are crucial for analyzing complex biomedical datasets.

Purpose of the Study:

• To propose a novel hierarchical likelihood learning framework for clustered biomedical count data.
• To capture both nonlinear overall effects and subject-specific effects using gamma random effects within Poisson neural networks.

Main Methods:

• Incorporation of gamma random effects into Poisson neural networks.
• Development of a robust end-to-end algorithm for clustered biomedical count data.
• Introduction of an adjustment procedure for random effects and variance component to enhance learning efficiency.

Main Results:

• The framework yields maximum likelihood estimators for fixed parameters and best unbiased predictors for random effects.
• Estimating equations remain unbiased even with misspecified random-effects distributions, ensuring robustness.
• Achieved competitive predictive performance using mean squared Pearson error and mean deviance.

Conclusions:

• The proposed hierarchical likelihood learning framework is practically effective for clustered biomedical count data.
• The method demonstrates robust and competitive predictive performance across diverse datasets and random-effects distributions.
• This approach enhances subject-specific predictions in biomedical research using neural networks.