Enhanced Dempster-Shafer classifier with metaheuristics and feature selection for predicting temporomandibular osteoarthritis progression

Abstract

Accurately predicting temporomandibular joint osteoarthritis (TMJ-OA) progression represents a significant clinical challenge. An enhanced version of the Dempster-Shafer classifier, optimized using the Reptile Search Algorithm (RSA), a metaheuristic optimization technique is introduced to minimize the Binary Cross-Entropy (BCE) loss function and frame TMJ-OA progression as a binary classification problem. A feature selection strategy using the Permutation Feature Importance (PFI) technique is implemented, guided by relevance scores determined by the Dempster-Shafer model. This methodology effectively reduced the dimensionality from 141 to 10 relevant features, derived from clinical and imaging data of 66 patients. The selected features were used to train eight classical machine learning models: Logistic Regression, Random Forest, Gradient Boosting, Support Vector Machine, K-Nearest Neighbors, Gaussian Naive Bayes (GNB), Decision Tree, and AdaBoost. The optimized classifier achieved competitive performance (accuracy = 0.9095 ± 0.0744; ROC AUC = 0.9467 ± 0.0705; F1-score = 0.8590 ± 0.1280), comparable to Random Forest and GNB, while providing superior probabilistic calibration. The integration of metaheuristic optimization and model-specific feature selection enhances both the accuracy and interpretability of the system, offering an efficient and explainable alternative for predicting TMJ-OA progression.