A Novel Approach with Deep Learning to Address Multilabel Imbalance in Predictive Modeling: ML-RUSMOTE
E Erjiang, Tingyan Wang, Ming Yu, Mary Dempsey, Attracta Brennan, JOHN J. CAREY · 2024
In medical settings, a significant disparity often exists between the rates of positive and negative health outcomes, exemplified by diseases such as cancer and instances of bone fractures. Predictive modeling for these outcomes can be adversely affected by data imbalance. Sample ratio imbalance is still a challenging and critical issue. While many existing studies have explored the imbalance between majority and minority class samples, the internal imbalance within minority class samples has not yet been well addressed. This study proposed and developed a novel algorithm, called Multi-label Random Undersampling and Synthetic Minority Oversampling Technique (ML-RUSMOTE), to address such complex multilabel sample imbalances. We then developed multi-label predictive models by combining the proposed ML-RUSMOTE algorithm with three representative classifiers: Binary Relevance, Multi-label k-Nearest Neighbors, and Multi-label Deep Neural Networks, respectively. The approach was benchmarked against various traditional methods for handling multi-label imbalances. We utilized a substantial cohort of 6,374 patients from three hospital centers in Western Ireland for model validation. Our findings demonstrate that the proposed ML-RUSMOTE algorithm, particularly when integrated with deep learning techniques, significantly outperforms conventional methods in managing multi-label imbalances. Promisingly, the proposed approach can help address common imbalance issues in disease risk predictions, particularly for those patient subgroups whose numbers or outcomes are underrepresented.