Enhancing Alzheimer's disease prediction using random forest: A novel framework combining backward feature elimination and ant colony optimization.

Background: Alzheimer's disease (AD) represents a significant global health challenge due to its increasing prevalence and the limitations of current diagnostic approaches. Early detection is crucial as pathological changes occur 10-15 years before clinical symptoms manifest, yet current diagnostic methods typically identify the disease at moderate to advanced stages. Machine learning techniques offer promising solutions for early prediction, but face challenges related to feature selection and hyperparameter optimization.

Objective: To develop an enhanced predictive model for Alzheimer's disease by integrating advanced feature selection techniques with nature-inspired hyperparameter optimization for Random Forest classifiers while ensuring robust validation and statistical significance testing.

Methods: This study employed three feature selection techniques (Whale Optimization Algorithm, Artificial Bee Colony, and Backward Elimination Feature Selection) and two hyperparameter optimization algorithms (Artificial Ant Colony Optimization and Bald Eagle Search) to improve Random Forest model performance. A dataset comprising 2,149 instances with 34 features was preprocessed using MinMax normalization and Synthetic Minority Oversampling Technique (SMOTE) applied only to training data to prevent data leakage. Statistical significance testing using McNemar's test was conducted to compare model performances. Model performance was evaluated using accuracy, precision, recall, F1-score, and AUC with confidence intervals calculated using bootstrap sampling.

Results: The combination of Backward Elimination Feature Selection with Artificial Ant Colony Optimization achieved the highest performance (95% accuracy ± 1.2%, 95% precision ± 1.1%, 94% recall ± 1.3%, 95% F1-score ± 1.0%, 98% AUC ± 0.8%), outperforming other methodological combinations and conventional machine learning algorithms with statistically significant improvements (p < 0.001). This approach identified 26 significant features associated with Alzheimer's disease. Additionally, nature-inspired optimization algorithms demonstrated substantial computational efficiency advantages over empirical approaches (18 minutes versus 133 minutes).

Conclusion: The integration of advanced feature selection with nature-inspired hyperparameter optimization enhances Alzheimer's disease prediction accuracy while improving computational efficiency. However, external validation on independent datasets and prospective clinical studies are needed to establish real-world utility. This methodological framework offers promising applications for early diagnosis and intervention planning, with potential extensions to other complex medical prediction tasks.