This study develops a method for predicting stress and anxiety levels using physiological signals from wearable devices, specifically accelerometer (ACC), skin temperature, and electrodermal activity (EDA). Machine learning models classify stress and anxiety states based on extracted features, enabling real-time monitoring. Evaluation results show that the Simple Decision Tree achieved the highest accuracy (93.77%) with a precision of 94.01% and recall of 94.29%. The K-Nearest Neighbors (KNN) model also performed well, attaining 92.48% accuracy, 89.41% precision, and 95.24% recall. Both models demonstrated strong predictive capabilities, with decision trees excelling in overall performance. The findings highlight the effectiveness of tested machine learning techniques for noninvasive stress detection, offering a reliable tool for early intervention and personalized mental health monitoring in daily life.