Automatic Gaze-Based Detection of Mind Wandering during Reading

We present a fully-automated person-independent approach to track mind wandering by monitoring eye gaze during reading. We tracked eye gaze of 84 students who engaged in an approximately 30-minute self-paced reading task on research methods. Mind wandering reports were collected by auditorily probing students in-between and after reading certain pages. Supervised classifiers trained on global and local features extracted from students’ gaze fixations 3, 5, 10, and 15 seconds before each probe were used to predict mind wandering with a leave-several-subjects-out cross validation procedure. The most accurate model tracked both global and local eye gaze in a 5-second window before a probe and yielded a kappa (accuracy after correcting for chance) of 0.23 on a downsampled corpus containing 50% yes and 50% no responses to probes. Implications of our findings for adaptive interventions that restore attention when mind wandering is detected are discussed.