Evaluating Children’s Engagement Through Wrist-worn Biosensors and LSTM Network

Children’s engagement in life events plays an important role in their academic and social performances. In children’s learning and daily life, video watching is becoming a common life activity to obtain new information and interact with computers. Assessing children’s engagement during video watching through computational models is essential for providing smart feedback and better interaction. In our study, we recruited forty-five children and instructed them to watch selected videos to elicit distinct engagement states. A commercial smartwatch was used to record children's physiological responses, including motion (IMU) and photoplethysmography (PPG) signals. After dividing data at one-minute time resolution and extracting features, significant differences were reported between the engaged and non-engaged conditions. Moreover, by employing a long short-term memory (LSTM) model, our binary classification analysis of the engaged and the non-engaged states revealed an accuracy of 91.24±4.25% for the IMU features, 81.32±4.19% for the PPG features, and 91.83±3.02% for the combined features at the cross-participant level. These results demonstrated the feasibility of evaluating children’s engagement states using commercial physiological devices and promoted the application of automatic engagement evaluation.