Article: watching people decide: decision prediction using heatmaps of reading a decision-support document.

Introduction: Previous studies show that reading behavior varies with the readers’ Levels of Expertise (LoE) in a task area. Except for LoE, other factors like acquired information plays a role in this process. In the area of health policymaking, people read supporting documents to inform their decisions. This leads to a natural question: could it be possible to predict the decisions based on the reading pattern of the supporting document on top of their LoE? Method: We collected eye tracker data from a group of people with various LoE. We used the heatmaps as the primary pattern of reading. These were prepared using the average fixation duration of the individuals. First, we performed a hierarchical cluster analysis with the pairwise correlation matrix between the heatmaps, to see whether heatmaps as A single feature were effective to reach our goal. In the second step, we made an ensemble of the features of the reading patterns from the heatmaps and pupillometric features, and LoE, with the decision made by the participants as an outcome, using AdaBoost regressor. In this decision-making task, one could choose one among expensive, prudent, and midway. Result: The first analysis reveals to us that there are a minority number of individuals who read less than the majority group. This minority group tends to make decisions in the extremities. The result of AdaBoost-regressor, shows us 1. the LoE is a stronger feature than the patterns of reading to predict the decision to be taken. 2. the pupillometric features are weaker feature than the reading patterns from the heatmaps for our task. Conclusion: Reading patterns could be useful for forecasting a decision, given the LoE of the individual. Heatmaps can be used as both qualitative and quantitative measures for reading patterns.