To improve occupant safety during building emergencies, evacuation simulations have been widely used for building safety design. Since occupant behavior is a determining factor for the outcome of building emergencies, accurately capturing how occupants make decisions and integrating occupants’ decision-making processes in evacuation simulations is important. In this study, based on the results of fire evacuation experiments in a virtual metro station, how different social (crowd flow) and environmental (visual access and vertical movement) factors would affect individuals’ wayfinding behavior was predicted using machine learning and discrete choice models. The trained models were further employed in agent-based evacuation simulations to examine crowd evacuation performance under different building design scenarios. Both the machine learning and discrete choice models could accurately predict individuals’ directional choices during emergency evacuations. Different building attributes could collectively influence occupant behavior, leading to distinct exit choices and evacuation times. While both the trained machine learning and discrete choice models generated similar results, the discrete choice model had better interpretability. Moreover, by comparing the trained models in this study with a model developed in a prior study, it was found that agents had significantly distinct responses to different building designs. Critical factors (e.g., type and size of buildings, occupants’ familiarity with the building) for the applicability of evacuation models were identified. Furthermore, recommendations were provided for future research that aims at employing evacuation simulations for building design evaluation and optimization.