A Deep Neural Network-based ADHD, ADHD-RISK Screening Method using 3D Skeleton Data Acquired Through a Robot-assisted ADHD Screening Game

Abstract

In this paper, ADHD, ADHD-RISK, and Normal were classified using Skeleton data of children acquired through the ADHD screening game. During the game, the child's skeleton is acquired through 5 Kinect azure. The acquired skeleton data was divided into 11 stages and stored. The model used to classify the above three groups used deep learning based on RNN. GRU, RNN, and LSTM were used as the model types. The above models add a bi-directional layer to improve performance. The input data used in the above model was a total of 596 people, 66 people with ADHD, 181 people with ADHD-RISK, and 349 people with Normal. Since the number of data for each group is different, a weighted cross entropy loss function is used as the loss function to prevent overfitting and improve performance. Among the models to which the bi-directional layer was added, the LSTM model obtained an accuracy of 97.82%. To check which of the 11 stages used in this study is helpful for deep learning, a model was designed by adding a channel attention layer. When the channel attention layer is used, attention scores are generated for each 11 stages for each child after learning is completed. Attention scores of children in the ADHD, ADHD-RISK, and Normal groups were separately classified. As a result of the classification, the stages ranked in the top 3 among the 11 stages of each child were collected. As a result, the ADHD and Normal groups showed similar attention scores at all stages. However, in the case of ADHD-RISK, it was confirmed that the attention score of the top 3 increased from the 3rd waiting stage and game stage. From the above results, it was confirmed that the body movements of children in the ADHD-RISK group increased rapidly from the third game. In other words, it can be inferred that children with ADHD-RISK show similar patterns to ADHD because their concentration decreases toward the latter part of the game. When developing an ADHD screening game using the above results and analyzes in the future, if you develop a game that requires concentration even in the second half while repeatedly playing multiple games, an ADHD screening system that can better classify ADHD, ADHD-RISK, and Normal will be able to develop.