This thesis presents a comprehensive study of the Maxima Nomination Sampling (MaxNS) method to construct rank-based classifiers for binary classification and their application. Utilizing the bone mineral density (BMD) dataset provided by the Manitoba BMD program, this research contrasts the efficacy of MaxNS methods with that of Simple Random Sampling (SRS) in handling data imbalance, a common challenge in medical data analysis.

The study explores the process of data sampling using expert knowledge that can be represented in terms of rank based on the likelihood of a sampling unit coming from the underlying class of interest, as well as data pre-processing including feature scaling, to optimize the dataset for machine learning models. A significant emphasis is placed on the use of Deep Neural Networks (DNNs), specifically in processing hip Dual-energy X-ray Absorptiometry (DXA) images, to extract ranking information for MaxNS sampling. The findings demonstrate that MaxNS methods significantly outperform SRS, especially in terms of recall metrics, showcasing their robustness and reliability in predicting the minority class.

This research contributes to the growing field of medical data analytics by providing insights into advanced sampling methods and their potential to improve the accuracy of predictive models in healthcare. The implications of this study extend to the development of more effective tools for medical diagnostics, ultimately aiming to enhance patient care through better predictive analytics.