In the busy world of today, early stress detection in people has emerged as a crucial study topic. Using two distinct stress datasets, this work suggests a machine learning (ML)-based stress detection algorithm. Designing threshold- and split-based decision tree classifiers to analyze stress using physiological, behavioral, and environmental data is the goal of the suggested methodology. In order to identify post-traumatic stress disorder (PTSD), the study also investigates sentiment analysis. It talks about issues including feature selection, contextual factors, individual variability, data quality, and model generalization that arise in ML-based stress detection. There are three case studies in which case 1 uses sentiment and PTSD data while Case 2 employs motion-based stress detection data and human activity. The outcomes show how successful decision tree-based classifiers are in both situations. Despite the unbalanced dataset, Case 1 achieves an accuracy of 85%., In Case 2, decision tree classifier produces an accuracy of 99.75% with precision, recall, and F1-scores at 0.99. In Case 3, we test how class imbalance affects model performance. We first change categorical features into numbers, use the method on several datasets, and then create an imbalanced dataset. The proposed Decision Tree method, which works by splitting data based on thresholds, performs better than advanced methods for stress detection. SVM based classifier also gets better results for all three classes.