A novel ensemble machine learning approach for interpretable modeling, feature extraction and selection with applications to medical and biomedical signals and data

Feature extraction and selection are crucial in biomedical data analysis to address high dimensionality, reduce computational complexity, and enhance model interpretability. However, traditional methods often focus on individual feature importance, overlooking complex inter-feature relationships, especially when processing and modeling dynamic and time-series data. In this study, we propose a novel framework that integrates Feature Co-occurrence Networks (FCN) with global importance scoring via the PageRank algorithm, which is built on a parametric Nonlinear AutoRegressive with eXogenous inputs (NARX) model structure to better capture temporal dependencies in sequential data. The proposed NARX-FCN-PageRank approach combines the strengths of multiple feature selection strategies while leveraging network analysis to identify stable and representative feature subsets. Extensive evaluations across diverse biomedical datasets, including both static and dynamic scenarios, demonstrate that our method effectively reduces feature dimensionality without compromising predictive performance. Moreover, the network visualizations provide valuable insights into the interdependencies and centrality of selected features, supporting model interpretability and enhancing trustworthiness. The NARX-FCN-PageRank framework thus offers a versatile and interpretable solution for feature selection in biomedical data analysis, with the potential to facilitate more efficient and reliable modeling in clinical and medical research applications.