Abstarct: Accurate and timely detection of stress through physiological signals is crucial, as stress has widespread negative effects on both mental and physical health. The aim of this study is to investigate the effectiveness of physiological signals, both individually and in combination with other modalities, to improve the accuracy of stress detection under personalized conditions. To this end, a personalized fine-tuning approach was implemented for binary stress classification using data from 15 subjects in the WESAD dataset across eleven physiological modalities. In this thesis, the Bi-GRU architecture and the Leave-One-Subject-Out (LOSO) validation method were employed to prevent data leakage and ensure model generalization. The training protocol included 50 baxse training epochs on general data and 25 fine-tuning epochs baxsed on entropy, where samples with the highest uncertainty were prioritized. The dataset was divided into 50% for testing, 30% for fine-tuning, and 20% for validation. Model performance was evaluated using accuracy, F1-score, and statistical analysis to assess the reliability and stability of the results. The findings showed that individual fine-tuning significantly improved system performance, and combined signals achieved the highest accuracy. Specifically, the chest composite signal reached 99.51% accuracy, while the combination of chest and wrist signals achieved 99.35% accuracy. Signals with low baxseline performance, such as Chest_ECG and Wrist_BVP, demonstrated substantial improvement after fine-tuning, rising from 43.06% to 98.52% and from 53.95% to 97.94%, respectively. Overall, the study demonstrated that integrating the Bi-GRU architecture with entropy-baxsed individual fine-tuning yields over 99% accuracy and high performance stability, making the developed system suitable for real-world stress detection applications.