The impact of imputation quality on machine learning classifiers for datasets with missing values
Tolou Shadbahr, Michael Roberts, Jan Stanczuk, Julian Gilbey, Philip Teare, Sören Dittmer, Matthew Thorpe, Ramón Viñas, Evis Sala, Píetro Lió, Mishal Patel, Jacobus Preller, Ian Selby, Anna Breger, Jonathan Weir‐McCall, Effrossyni Gkrania‐Klotsas, Anna Korhonen, Emily Jefferson, Georg Langs, Guang Yang, Helmut Prosch, Judith Babar, Lorena Escudero Sánchez, Marcel Wassin, Markus Holzer, Nicholas Walton, Píetro Lió, James H.F. Rudd, Tuomas Mirtti, Antti Rannikko, John A. D. Aston, Jing Tang, Carola‐Bibiane Schönlieb
Abstract
BACKGROUND: Classifying samples in incomplete datasets is a common aim for machine learning practitioners, but is non-trivial. Missing data is found in most real-world datasets and these missing values are typically imputed using established methods, followed by classification of the now complete samples. The focus of the machine learning researcher is to optimise the classifier's performance. METHODS: We utilise three simulated and three real-world clinical datasets with different feature types and missingness patterns. Initially, we evaluate how the downstream classifier performance depends on the choice of classifier and imputation methods. We employ ANOVA to quantitatively evaluate how the choice of missingness rate, imputation method, and classifier method influences the performance. Additionally, we compare commonly used methods for assessing imputation quality and introduce a class of discrepancy scores based on the sliced Wasserstein distance. We also assess the stability of the imputations and the interpretability of model built on the imputed data. RESULTS: The performance of the classifier is most affected by the percentage of missingness in the test data, with a considerable performance decline observed as the test missingness rate increases. We also show that the commonly used measures for assessing imputation quality tend to lead to imputed data which poorly matches the underlying data distribution, whereas our new class of discrepancy scores performs much better on this measure. Furthermore, we show that the interpretability of classifier models trained using poorly imputed data is compromised. CONCLUSIONS: It is imperative to consider the quality of the imputation when performing downstream classification as the effects on the classifier can be considerable.