Imputing Biomarker Status from RWE Datasets-A Comparative Study.

Area of Science:

• Biostatistics
• Health Informatics
• Machine Learning

Background:

• Missing data is a pervasive challenge in Real-World Evidence (RWE) analysis.
• Identifying features correlating with clinical outcomes requires robust handling of incomplete datasets.
• Biomarker data often presents missing values, necessitating effective imputation strategies.

Purpose of the Study:

• To evaluate and compare the performance of six distinct imputation methods for handling missing data in RWE.
• To assess imputation techniques on a large-scale dataset (Flatiron NSCLC) and through synthetic data experiments.
• To provide guidance on selecting appropriate imputation algorithms based on data characteristics and missingness patterns.

Main Methods:

• Comparison of six imputation algorithms: predictive mean matching, expectation-maximisation, factorial analysis, random forest, generative adversarial networks, and multivariate imputations with tabular networks.
• Utilisation of the Flatiron Non-Small Cell Lung Cancer (NSCLC) dataset (over 35,000 subjects).
• Extensive synthetic data experiments employing structural causal models to simulate missingness.

Main Results:

• Multivariate imputations with tabular networks demonstrated superior performance in synthetic data experiments.
• Neural network-based methods show promise for complex datasets with non-linear relationships.
• Predictive mean matching proved effective for the specific biomarker data within the Flatiron NSCLC dataset.

Conclusions:

• The choice of imputation algorithm should consider the nature of missingness and potential distribution shifts.
• Neural networks offer powerful solutions for complex RWE datasets.
• Traditional methods like predictive mean matching remain valuable for certain RWE biomarker analyses.