Children with strong impulsivity tend to have weak self-control and often act immediately on their thoughts. Examples of behaviors in the classroom that stem from impulsivity include interrupting others to suddenly express their own opinions, or being...
Children with strong impulsivity tend to have weak self-control and often act immediately on their thoughts. Examples of behaviors in the classroom that stem from impulsivity include interrupting others to suddenly express their own opinions, or being unable to endure listening to a lesson and instead engaging in unrelated activities. Children who exhibit such behaviors are at risk of developing secondary disorders such as depression and anxiety disorders. Therefore, early detection of impulsivity-induced behavior is crucial for finding appropriate ways to support child development. Methods for detecting impulsivity and hyperactivity in children with attention-deficit/hyperactivity disorder (ADHD) using electroencephalography (EEG) signals and skeletal information extracted from video cameras have previously been proposed. However, EEG devices are physically large and unsuitable for long-term use during children's daily activities. Video camera recording also fails to extract skeletal information accurately when hands are hidden by legs, and if the number of cameras is increased, the children's attention may shift to the cameras themselves. Furthermore, it's difficult to collect a sufficient amount of impulsivity-related behavior, and collecting data from ADHD subjects repeatedly in response to changes in environment or situation is impractical. To address these problems, we attempt to detect abnormal behavior using inertial measurement unit (IMU) sensors, which measure motion using accelerometers and gyroscopes. IMU sensors enable stable, continuous measurement regardless of a child's posture and are more suitable for long-term wear compared to EEG devices. To overcome the challenge of insufficient impulsive behavior data, we propose a method that detects anomalies using only normal data. Specifically, we adopt a Variational Autoencoder (VAE) model based on unsupervised learning. By incorporating networks capable of capturing temporal-dependent features into the encoder and decoder, we propose a model that can more accurately interpret children's behavior. In the experiment, we evaluate the detection performance using the area under the receiver operating characteristic curve (AUC) and compare the results with those of conventional methods such as VAE and LSTM-VAE, and we have demonstrated that the proposed method's model is useful as a highly sensitive anomaly detector with fewer false positives.