Missing Data in Machine Learning -- Tree-based generative imputation, uncertainty in explanation and data augmentation in healthcare application

Healthcare datasets are often affected by missing data and data scarcity, which require careful handling to ensure reliable analysis. Missing data is typically addressed with imputation methods, which fill the missing values with point estimates. The literature distinguishes between single imputation, which outputs one complete dataset and analyzes the dataset once, and multiple imputation, which outputs several imputed datasets, analyzes each dataset, and pools the results into a point estimate and a standard error. Novel imputation methods based on machine learning are usually only introduced for single imputation, which is convenient but does not account for imputation uncertainty as multiple imputation does. Data scarcity is usually addressed by data augmentation, for example, by synthesizing new data. Both missing data imputation and synthesizing data can be solved with generative models, which have recently seen great success in generating text and images. Most of the proposed methods, however, are deep-learning-based, which is not always the best approach for tabular data, which this work focuses on. Here, tree-based machine learning methods traditionally perform well, and there are also some generative models that offer promising solutions for data synthesis and density estimation, such as adversarial random forests (ARF).

This thesis discusses handling missing values in a machine learning pipeline, from prediction to explanation. I examine the role of missing data, how it is addressed in machine learning models, and the effects of missing data on model explanations. I focus on the application of generative modeling in this field and connect imputation and generative modeling. My contribution to the field is a novel imputation method missing value imputation with adversarial random forests (MissARF), based on ARF, which offers competitive imputation performance in both single and multiple imputation at no additional computational cost in multiple imputation. Second, I discuss imputation uncertainty in post-hoc global explanations and show that only multiple imputation achieves good coverage. Lastly, I examine a real data example from the field of geriatrics, where data is scarce, and the aim is to improve the predictive performance through data augmentation. I compare two approaches, synthesizing the data and combining the dataset with a related dataset with imputation.