Stress is a critical health issue with significant implications for both physical and mental well-being. Early detection and accurate classification of stress levels are crucial for timely intervention and effective management. This study proposes a deep-learning framework for early stress detection using physiological data from wrist-worn devices. The framework utilizes Electrodermal Activity (EDA) signals, which are processed with advanced preprocessing techniques. It uses both extracted features and the complete feature set of the EDA signal in a custom deep-learning model to effectively distinguish between stress and non-stress states. The study also explores multimodal signal fusion by combining EDA with electrocardiogram (ECG) and photoplethysmography (PPG) signals to evaluate its effectiveness in early stress detection. However, the results show that EDA signals alone outperform the combined signals for early stress detection. The methodology was trained and tested on six publicly available EDA-based stress datasets: VerBIO, UTD, WESAD, SWELL-KW, MSD Nurses, and DriveDB. Experimental results demonstrate that the framework achieves 95% accuracy and a 90% F1-score in early stress detection. These findings highlight the potential of the proposed framework for real-world applications in stress management, providing a reliable, continuous, and non-invasive tool for early stress monitoring through wearable technology.