Human operators play a critical role in various Cyber-Physical System (CPS) domains, for example, transportation, smart living, robotics, and medicine. The rapid advancement of automation technology is driving a trend towards deep human-automation cooperation in many safety-critical applications, making it important to explicitly consider user behaviors throughout the system development cycle. While past research has generated extensive knowledge and techniques for analyzing human-automation interaction, in many emerging applications, it remains an open challenge to develop quantitative models of user behaviors that can be directly incorporated into the system-level analysis. This dissertation describes methods for modeling different types of user behaviors in medical CPS and integrating the behavioral models into system analysis. We make three main contributions. First, we design a model-based analysis framework to evaluate, improve, and formally verify the robustness of generic (i.e., non-personalized) user behaviors that are typically driven by rule-based clinical protocols. We conceptualize a data-driven technique to predict safety-critical events at run-time in the presence of possible time-varying process disturbances. Second, we develop a methodology to systematically identify behavior variables and functional relationships in healthcare applications. We build personalized behavior models and analyze population-level behavioral patterns. Third, we propose a sequential decision filtering technique by leveraging a generic parameter-invariant test to validate behavior information that may be measured through unreliable channels, which is a practical challenge in many human-in-the-loop applications. A unique strength of this validation technique is that it achieves high inter-subject consistency despite uncertain parametric variances in the physiological processes, without needing any individual-level tuning. We validate the proposed approaches by applying them to several case studies.