Multiple Imputation with Neural Network Gaussian Process for High-dimensional Incomplete Data.

Area of Science:

• Computational Statistics
• Bioinformatics
• Machine Learning

Background:

• Missing data is a common challenge in data analysis, potentially causing information loss and biased results.
• High-dimensional incomplete datasets, like those in multi-omics, pose significant challenges for traditional imputation methods.
• Existing single imputation methods lack uncertainty quantification, while multiple imputation (MI) methods struggle with high dimensionality.

Purpose of the Study:

• To develop novel multiple imputation (MI) methods capable of handling high-dimensional incomplete data.
• To leverage advances in neural network Gaussian processes (NNGP) within a Bayesian framework for improved imputation.
• To address the limitations of current imputation techniques in complex, high-dimensional scenarios.

Main Methods:

• Proposed two novel MI methods based on neural network Gaussian processes (NNGP), termed MI-NNGP.
• Utilized a Bayesian approach to apply multiple imputations from a joint posterior predictive distribution.
• Evaluated performance across different missing data mechanisms: MCAR, MAR, and MNAR.

Main Results:

• MI-NNGP methods demonstrated superior performance compared to state-of-the-art methods on both synthetic and real-world datasets.
• Significant improvements were observed in imputation accuracy, statistical inference validity, and robustness to varying missing data rates.
• The proposed methods also showed competitive computation costs.

Conclusions:

• MI-NNGP offers a powerful and effective solution for handling missing data in high-dimensional applications, particularly in multi-omics.
• The Bayesian NNGP framework provides a robust approach for multiple imputation, enhancing analytical reliability.
• These methods represent a significant advancement in addressing methodological gaps in missing data imputation.