A Unified Machine Learning Framework for Modelling and Feature Selection in Structured and Temporal Data with Biomedical Applications

Machine learning (ML) has been widely applied in biomedical and healthcare data analysis to support disease assessment and clinical decision-making. However, biomedical datasets often contain heterogeneous structured and temporal data, are limited in sample size, and require interpretable models for reliable deployment. This thesis develops interpretable and lightweight ML methodologies for modelling and feature selection in structured and temporal biomedical data.

Firstly, a NARX-based ML framework is proposed for classification tasks. By combining nonlinear feature construction with forward feature selection, the framework captures nonlinear relationships while maintaining interpretability through explicit feature interactions and compact feature subsets. Experimental results achieved 97.79% accuracy on an obesity dataset and 90.34% accuracy on an EEG eye-state dataset using only a small number of selected nonlinear features.

Secondly, a feature selection fusion framework is developed by integrating heterogeneous feature selection methods through a Feature Co-occurrence Network (FCN) and a PageRank-based ranking strategy. The framework identifies consistent and interpretable feature subsets by modelling feature co-selection relationships across multiple selection methods. Experimental results achieved 97.44% accuracy on the Parkinson dataset using only eight selected nonlinear features and 99.17% accuracy on the ECG dataset using twelve selected nonlinear features, outperforming conventional voting and rank-aggregation approaches.

Finally, the proposed methodology is extended from classification to regression tasks, demonstrating its applicability across different prediction objectives. The resulting unified learning pipeline supports preprocessing, nonlinear feature construction, feature selection, and model evaluation within a consistent framework.

Overall, the proposed methodologies demonstrate competitive predictive performance, improved feature efficiency, enhanced interpretability, and reduced computational complexity across multiple biomedical datasets involving structured clinical variables and temporal physiological signals.