Severity Detection of Anxiety, Depression, And Stress Using Ensemble Learning and Deep Neural Networks: A Multi-Class Approach

Abstract

 The study examines ordinary DASS-21 surveys and various models of machine and deep learning to identify the levels of anxiety, depression, and stress. Mental health data class imbalance greatly affects the upper class intensity of model training time. To optimize model performance, the dataset was balanced using the Synthetic Minority Over-sampling Technique (SMOTE). We assessed cross model performance by tuning accuracy, precision, recall, and f1 score and analyzing the various techniques employed. The enactment of the models was rated based on the predominance of upper intensity class response. Of the several models, the Random Forest algorithms surpassed the others/competitors across all levels of intensity. The model was able to identify all levels of intensity approximately 98.43% of the time. The apparent intra-class postulation of Random Forest where the average accuracy, recall, and F1-scores were high postulates perfect and equal intra-class distribution. The results showed that Random Forest did the best out of all the models, and this, along with the best performance as a whole, led to the conclusion that ensemble methods are a reliable way to classify different levels of mental health seriousness. Random Forest's success shows that it could be used to automatically find mental health problems in study and therapy settings. The data also show that automatic systems can help decision support systems find high-risk individuals early on and give mental health systems the tools they need to help people in the right way at the right time.