MPR-MOPSO-ASFS: A Stable Multi-Objective Feature Selection Algorithm for Metabolomics

Background: Metabolomics data is inherently characterized by high dimensionality and small sample size. Existing multi-objective particle swarm optimization (MOPSO)-based feature selection algorithms suffer from two critical limitations: they generally neglect the stability of selected feature subsets and exhibit poor adaptability when processing high-dimensional data, which restricts their practical application in biomedical research. Methods: To address these challenges, this paper proposes a novel MPR-MOPSO-ASFS algorithm. Specifically, we first construct a stability-driven bi-objective optimization model. Then, we integrate a Maximum Pattern Recognition (MPR) filter to achieve rapid dimensionality reduction of high-dimensional features. Finally, an improved Adaptive Sparsity Feature Selection mechanism is designed to simultaneously optimize the compactness, classification accuracy, and stability of the selected feature subsets. Results: Extensive experiments were conducted on three metabolomics datasets and five public high-dimensional small-sample datasets. The results demonstrate that the proposed MPR-MOPSO-ASFS algorithm outperforms mainstream algorithms including CMDPSOFS and MOEAD-FS in core evaluation metrics such as Pareto front quality and classification accuracy. Additionally, we clarify the optimal configuration of the algorithm's core parameters and verify the collaborative effectiveness of its key design components. Conclusions: This study makes the first attempt to incorporate the harmonic mean of accuracy and stability into the multi-objective optimization framework. The two-stage combination of the proposed MPR strategy and MOPSO breaks through the performance bottleneck of traditional algorithms and significantly enhances their adaptability and robustness to high-dimensional small-sample data. The proposed algorithm provides an efficient new solution for feature selection in high-dimensional biomedical data and offers a technical reference for the application of swarm intelligence optimization algorithms in the biomedical field.