Hybrid Optimization-Based Machine Learning Framework for Predicting Mild Cognitive Impairment Progression

Abstract

Mild Cognitive Impairment (MCI) is a transitional state from normal aging to dementia. Early detection of its progression is critical for clinical intervention and patient care in a timely manner. In this study, a hybrid machine learning approach to predict the progression from MCI to dementia using OASIS dataset. The proposed methodology combines Recursive Feature Elimination (RFE), Particle Swarm Optimization (PSO), Sequential Model-Based Optimization (SMBO) and Gradient Boosting Classifier (GBC) to improve prediction performance and model robustness. Preprocessing techniques were utilized to rectify missing values, remove superfluous features and execute category encoding. RFE was utilized to discern the most relevant clinical and neuroimaging attributes associated with dementia progression, while PSO was applied to refine feature subsets and enhance model efficiency. The hyperparameter optimization of the GBC was accomplished using Optuna-based SMBO to determine the optimal parameter configuration. The optimized framework effectively found complex nonlinear relationships among cognitive, demographic and brain imaging factors. The experimental results demonstrated that the proposed model achieved a classification accuracy of 96%, showcasing good precision, recall and F1-score in distinguishing stable MCI from MCI that advanced to dementia.The model demonstrated reliable predictive performance with robust ROC-AUC and Precision-Recall AUC scores. SHAP-based interpretability analysis revealed that age, normalized whole-brain volume, Clinical Dementia Rating (CDR), Mini-Mental State Examination (MMSE)were the most significant predictors of dementia conversion. The combination of PSO with RFE and SMBO improved feature optimization and model generalization. The proposed method is an effective decision-support tool for early evaluation of dementia risk and personalized clinical care for neurodegenerative diseases.