Stress can be categorized into acute and chronic types, with acute stress having short-term positive effects in managing hazardous situations, while chronic stress can adversely impact mental health. In a biological context, stress elicits a physiological response indicative of the fight-or-flight mechanism, accompanied by measurable changes in physiological signals such as blood volume pulse (BVP), galvanic skin response (GSR), and skin temperature (TEMP). While clinical-grade devices have traditionally been used to measure these signals, recent advancements in sensor technology enable their capture using consumer-grade wearable devices, providing opportunities for research in acute stress detection. Despite these advancements, there has been limited focus on utilizing low-resolution data obtained from sensor technology for early stress detection and evaluating stress detection models under real-world conditions. Moreover, the potential of physiological signals to infer mental stress information remains largely unexplored in lifelog retrieval systems. This thesis addresses these gaps through empirical investigations and explores the potential of utilizing physiological signals for stress detection and their integration within the state-of-the-art (SOTA) lifelog retrieval system. The main contributions of this thesis are as follows. Firstly, statistical analyses are conducted to investigate the feasibility of using low-resolution data for stress detection and emphasize the superiority of subject-dependent models over subject-independent models, thereby proposing the optimal approach to training stress detection models with low-resolution data. Secondly, longitudinal stress lifelog data is collected to evaluate stress detection models in real-world settings. It is proposed that training lifelog models on physiological signals in real-world settings is crucial to avoid detection inaccuracies caused by differences between laboratory and free-living conditions. Finally, a state-of-the-art lifelog interactive retrieval system called \lifeseeker is developed, incorporating the stress-moment filter function. Experimental results demonstrate that integrating this function improves the overall performance of the system in both interactive and non-interactive modes. In summary, this thesis contributes to the understanding of stress detection applied in real-world settings and showcases the potential of integrating stress information for enhancing personalized lifelog retrieval system performance.