Research on learning analytics (LA) has focused mostly at the university level. LA research in the K-12 setting is needed. This study aimed to understand 4,115 middle school students' learning paths based on their behavioural patterns and the relationship with performance levels when they used a digital learning game as their science curriculum. The findings showed significant positive relationships between various tool uses and performance measures and varied tool use patterns at different problem-solving phases by high- and low-performing students. The results indicated that students who used tools appropriately and wisely, given the phase they were at, were more likely to succeed. The findings offered an insightful glimpse of learners' navigation patterns in relation to their performance and provided much-needed empirical evidence to support using analytics for game-based learning in K-12 education. The findings also revealed that log data cannot explain all learners' actions. Implications for both research and practice are discussed.