Uncertainty Quantification and Enhancing Panic Disorder Detection using Ensemble and Resampling Techniques

Panic disorder is one of the leading mental health problems that entail serials of extreme fear that can highly hinder an individual's activities of daily life. Sometimes, intervention is very important, understanding the problem is highly crucial. Unfortunately, some methods assist with diagnostics do not provide accuracy in identification. Using machine learning approaches can resolve this problem as accuracy can be improved with data-driven models. This paper used ensemble machine learning models, Random Forest, Bagging Classifier, and Balanced Bagging Classifier to identify panic disorder. Medical datasets usually have a class imbalance problem. Therefore, we performed resampling SMOTE, ADASYN, and Tomek Links. We evaluated these models by accuracy, precision, recall, F1 score, ROC AUC, Cohen's Kappa, uncertainty measures, aleatoric uncertainty, epistemic uncertainty, and predictive entropy. In our results, The Bagging Classifier was out performance with the highest accuracy (99.97%), recall (99.66%), F1-score (99.60%), and Cohen’s Kappa (99.58%), minimal uncertainty metrics (aleatoric: 0.00062, entropy: 0.002003), establishing itself as the optimal model for panic disorder diagnosis. This study proves the effectiveness of ensemble learning and resampling methods for early panic disorder diagnosis and future mental health technologies.