Stress plays a major role in physical and emotional well-being, and is associated with several illnesses including depression, diabetes, and other chronic diseases. College student stress as a construct is important to detect in order to equip students in a timely manner with stress coping strategies. However, the lack of a passive sensing measure accepted as a gold standard impedes real time detection and treatment of stress. Many researchers are studying passive sensing of stress using wrist-worn sensors; however, this effort focuses on understanding the essential features of wrist-worn sensors in detecting stress, and how to best induce stress in a lab setting. Applying machine learning methods increasingly is making it feasible to validly infer in real time through passive sensing of physical features psychological states, such as stress. Given strong participant adherence to wrist-worn sensors, this paper focuses on analyzing the effect of replacing other body sensing platforms (e.g. chest-based heart rate) with their wrist-worn equivalent on stress prediction accuracy. Nine participants were equipped with multiple body sensors and were asked to wear a commercially available Android smartwatch, a custom-designed smartwatch equipped with a Galvanic Skin Response (GSR) sensor, a chest-based heart rate sensor, and a finger-based commercial GSR sensor. Based on participant self-reports, singing experiment showed greatest stress levels across participants. This paper further analyzes features for prediction from all sensors compared to wrist-worn only sensors. Using statistical features on one-minute fixed-time sub-divisions and correlation-based feature subset selection and a Random Forest model, the system is capable of detecting stress with 88.8% F-measure.