Behavioral motion patterns to predict child mobile device use: Evaluating model performance across a standardized screen time protocol

Purpose Screen time is an increasingly prevalent behavior among children, and accurately measuring children’s screen time is critical for understanding its implications for health. When screens are shared among families with young children, it is challenging to tell who is using the device and attribute this screen time to the correct user. This proof-of-concept study evaluates whether accelerometer and gyroscope data can identify a child user, independent of game, by applying models to data collected during a standardized screen time protocol. Methods Thirty-seven parent-child dyads (children: 9.4 ±1.8 years, 57% female; parents: 41.8 ±6.7 years, 70% female) engaged in a ∼30-min standardized protocol, where they played a counterbalanced series of games (i.e., Like a Dino, Block Blast, Toy Blast, Melon Maker) and watched YouTube videos on a Pixel tablet, while a sensor logging application captured tablet accelerometer and gyroscope data. Random Forest models were trained to identify each target child user from their parent at the second level. We applied a split-half train/test approach, where the first half of the protocol (two unique games, video watching) was used for model training and the second half of the protocol (two unique games, video watching) was used for model testing. Model performance was evaluated across seen (model training included the target child’s parent and the target child) and unseen (model training only included the target child) conditions. Results/findings Model performance was high (F1 scores > 0.80), when applying models to predict a target child when games were standardized across participants. Performance metrics were higher in the seen condition (F1 score: 0.91) compared to the unseen condition (F1 score: 0.83). Conclusions Models successfully identified unique movement patterns (i.e., user-specific movement patterns) of children even when playing the same games as their parents. The study also demonstrated that children can be distinguished from their parents, and including parents in the training data improves model performance. Future research should explore if researchers can reliably identify users over time.