Imputation of missing values in breast cancer data

• Main Author: Rajagopal, Tejas R.
• Other Authors: Fan Xiuyi
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Medicine, Health and Life Sciences; Imputation
• Online Access: https://hdl.handle.net/10356/176005
• Institution: Nanyang Technological University

The critical role of complete and accurate data in breast cancer research and breast cancer diagnosis is the impetus behind this study, which rigorously examines and compares the efficacy of various imputation methods, focusing on the potential superiority of autoencoders over established techniques. This comparative analysis initiates with the UCI Wisconsin Breast Cancer Dataset, where Multiple Imputation by Chained Equations (MICE) sets a commendable baseline for accuracy. Imputing missing data using autoencoders does not yield a performance as good as MICE within this dataset. The research then transitions to the SEER Breast Cancer Dataset, marked by a complex array of features, encompassing both categorical and numerical data. It is within this intricate dataset that autoencoders demonstrate remarkable proficiency, significantly outperforming the baseline MICE model. The dichotomy in results between the two datasets underscores the conditional nature of imputation method performance, heavily influenced by the dataset’s characteristics. Concluding the study is an exploration of the resilience of these imputation methods against datasets with incrementally introduced missing values. Even under heightened volumes of missing data, the autoencoder maintains a competitive edge on the SEER dataset, although the margin narrows.These findings suggest a nuanced approach to the imputation of missing breast cancer data, emphasizing the selection of the method contingent upon the dataset’s complexity and composition. Autoencoders emerge as a promising model, particularly adept at managing datasets of a sophisticated nature, potentially enabling better clinical decision-making and aiding in the conduct of breast cancer research. Although the study focuses on breast cancer data, the findings may be extended to other forms of medical data given the similarities within their data points. Overall, this study concluded that autoencoder’s imputation outperforms MICE on the SEER breast cancer dataset, whereas MICE outperforms autoencoder imputation on the Wisconsin breast cancer dataset.