A custom build multidimensional medical combined imputation application for a transplantation dataset

Abstract

Background and Objectives: Data science methods have grown to solve complex medical problems. Data records utilized are often incomplete. Within this study we developed and validated a novel multidimensional medical combined imputation (MMCI) application to analyse multifaceted and segmented datasets as found in liver transplantation registries. Methods: The multidimensional medical combined imputation (MMCI) application is a pipeline of interconnected methods to impute segmented clinical data with the highest accuracy. Two different complete datasets were used in the testing procedure. A transplantation dataset (TxData) and a multivariate Wisconsin breast cancer (diagnostic) dataset (BcData). For both datasets, the most common imputation methods were tested, and their accuracy (ACC) compared to the novel MMCI (RF and LR). Results: In the TxData the MMCI RF and MMCI LR outperformed the other imputation algorithms regarding ACC. In the BcData the overall performance was good. The MMCI LR was the most superior algorithm for up to 10% of missing values with ACC = 91.9 (at 5% missing) to 90.6 (at 10% missing). The MMCI RF was the most accurate from 89.9 at 20% missing to 89.4 at 30% missing. All other established imputation algorithm showed inferior ACC, with MF and MICE showing results close to ACC of 90. Conclusion: This study presents the MMCI as a novel imputation pipeline to handle segmented and multifaceted clinical data. The MMCI proved to be more accurate than the established imputation methods when analysing 5–30% missing data. This study warrants future studies to investigate the value of the MMCI in predicting missing values in different datasets.

Keywords

• Statistical innovation
• Surgery
• Transplant
• Clinical data
• Imputation
• Machine learning