Player decision modeling can provide useful guidance to understand player performance in serious games. However, current player modeling focuses on high-level abstraction of player behavior rather than decision-level player modeling, and is predominantly applied to entertainment games. In this paper, we describe an approach from game design to data mining and data analysis to determine detailed player decision patterns. We illustrate this approach with VistaLights, a supply chain game we developed based on a recent oil spill event in Houston. With this game, we set up a within-subjects experiment to study decision making under varying circumstances, specifically to consider whether/how a recommendation system can improve human decisions. Using a series of data analysis techniques we built a coarse-grained decision model as well as a fine-grained model to compare players' actions on the game outcomes. The results confirm the need for decision-level modeling and show an ability of our approach to both identify the good and bad decision patterns among players.