Estimating Cognitive Load via Process-Based Telemetry for Adaptive Serious Games

This study examines whether telemetry-derived interaction metrics can estimate perceived cognitive load during gameplay, without the interruptions that questionnaires and physiological sensors impose. Thirty university students played LUTLAB, a puzzle-based serious game teaching FPGA hardware concepts across ten levels, while the game logged sixteen process-based metrics spanning mouse movement, pause patterns, repeated game states, button and tool use, click behavior, help-seeking, and dwell time. Participants completed the Raw Task Load Index after each level, and Spearman correlations were computed between each metric and reported load across 150 level-observations.

The principal finding is that dwell time in task rooms was the only metric that correlated consistently with perceived load, showing moderate-to-strong correlations (ρ = 0.48 to 0.66, all p < 0.01) across all five analyzed levels. Active mouse movement speed produced stronger correlations in individual levels, reaching ρ = -0.67 in an extended analysis of a smaller subgroup, but it did not hold across the full set of levels. State duplicates and contemplation pause frequency were predictive primarily in Level 3, the first level allowing free interaction with game elements. Most other metrics showed weak or no relationship with reported load.

The central conclusion is that the predictive value of an interaction metric depends on the structure of the task in which it is measured rather than being a general property of the metric. The contrast between a structured tutorial level and a free-form level illustrates this, as different behavioral signals became informative under different task conditions. For adaptive serious games, this implies that load indicators must be selected and validated for each module rather than applied uniformly across an entire game.

The study is exploratory and carries clear limitations. Its sample of 30 self-selected students from a single course restricts statistical power in later levels and generalizability, and the extended subgroup analysis reflects a selection effect in which faster, lower-load players reached higher levels. Mechanic-specific metrics such as state duplicates are unlikely to transfer to other games, while dwell time and active mouse speed appear to be the most plausible cross-game candidates, pending replication. Dwell time may also conflate genuine effort with non-task causes of long room times, and load was measured only per level, which may mask variation between room types.