Enhanced stacking ensemble Model: A statistical ensemble pruning framework to classify anxiety severity for responsive emergency preparedness

Abstract

Early anxiety prediction is essential because it facilitates the delivery of timely and appropriate medical services. Despite the recent development of a variety of severity scores and machine-learning algorithms for early anxiety prediction, such prediction is still difficult regarding accuracy and a few other performance indicators. The enhanced stacking ensemble strategy is suggested in this research to forecast and classify the level of anxiety. This framework involves two layers where the first layer experiments with very popular and diverse machine learning classifiers in the literature; the second layer consists of a selective fusion of top-performing base learners who have performed well among the pool of base classifiers. The selective fusion of top learners (ensemble pruning) involved in this framework is purely based on a statistical procedure and the subsequent fusion of decisions is been done using logistic regression as a meta-classifier. This proposal gives a new dimension to handle the concept of ensemble pruning using a statistical approach where effective anxiety level classification is done. To the best of the authors' knowledge, such research has not been reported earlier in the literature. Thus, the proposed statistical approach-based pruned ensemble named Enhanced Stacking Ensemble (ESE) Model lifts the overall ensemble’s predictive performance. After analyzing the results, it is noticed that our proposed ESE model performed better than most of the conventional methods (bagging, boosting, and voting) based on an accuracy of 94.11 percent, precision of 97 percent, recall of 91 percent, and F1 score of 94 percent.