An iterative feature selection framework for statistical classification: The LR-GMM OctoCore model

Logistic regression (LR) is widely used in medical data analysis for disease prediction and diagnosis due to its simplicity and interpretability. However, its linear assumption limits its effectiveness in high-dimensional and complex medical datasets. Gaussian Mixture Models (GMMs) can capture complex data distributions and identify patient subgroups but lack effective feature templates to support applications in medical research and related fields. To address these challenges, we propose an LR-GMM OctoCore model that leverages GMM’s probabilistic modeling with LR’s classification strength. The model starts with LR and passes low-confidence samples to parallel GMM chains, where multiple GMMs with different initializations explore the underlying distribution patterns to optimize classification decisions. At inference time, where ground truth is unavailable, the transition from LR to GMM is triggered by an LR posterior-probability confidence threshold τp, ensuring a training–inference consistent routing rule. Experiments on simulation studies show that the OctoCore model can effectively handle various data distributions and overcome the limitations of linearly non separable data, achieving superior classification results. In empirical analysis, we evaluate the model on three medical datasets and three classification tasks from different domains. The results show that OctoCore exhibits high predictive power in both binary and multi-class classification, outperforming traditional machine learning (ML) models. Although still in its early stages of refinement, OctoCore has significant potential for future applications.