Stress has become a significantly common mental condition in the current society. It is important for people to be aware of this condition effectively when it occurs. Due to the current improvements in sensor technology, have significantly improved the accuracy and efficiency of collecting data related to human physiological biomarkers. This study focuses on employing Deep learning techniques to accurately determine mental stress states of mental stress based on the WESAD public dataset, a collection of physiological data to detect stress and affect the state. The study has demonstrated the effectiveness of using Long Short-Term Memory (LSTM) and proposed a novel framework namely, Deep Reinforcement Active Learning to enhance the accuracy of LSTM for stress classification. The proposed framework has achieved 93% accuracy, which resulted 4.91% improvement from the original study. This improvement in accuracy was achieved only by employing respiration data, which clearly demonstrates the potential of the proposed framework.