关键词: 机器学习, 创伤后应激障碍, 转归预测, 儿童

Abstract:

Post-traumatic stress disorder (PTSD) could have negative effects on the development of children, and its impact can even last into adulthood. The traditional method to identify and diagnose PTSD in children is for the clinician or researcher to compares the symptoms with the criteria in the diagnostic manual. Therefore, the child who meets the symptom criteria could be diagnosed with PTSD. In addition, risk factors for PTSD in children are identified by traditional multiple regression methods using hypotheses based on previous literature or experience. However, these methods rely on clinicians or researchers' personal experience greatly. Moreover, identifying child PTSD is subjective, and the selected statistical method could impact the predicted risk factors. Generally, researchers use prediction method based on regression models. However, the identification of risk factors is not comprehensive enough, which needs a lot of empirical data to discuss. Thus, machine learning, as an emerging method to deal with big data, is a data-driven method to summarize rules and features based on existing data. Through continuous data training, the program could make its own judgment on whether children in new data have PTSD, which is more objective, faster and more efficient than human diagnosis. When predicting risk factors, machine learning models have also developed from traditional decision trees and regression to the field of deep learning, with greatly improves the accuracy of diagnosis and simultaneous processing of multi-dimensional variables. Therefore, using machine learning to predict children's PTSD could make up for the disadvantages of traditional prognosis investigation, which is difficult to follow for a long time and has large missing values, etc. Machine learning may also better solve other related problems, such as failing to detect PTSD symptoms in time and missing the optimal healing period due to the late-onset of PTSD in children. The application of machine learning in predicting the outcome of children’s PTSD results could be divided into two methods. One is the "classification" of supervised learning, which is the possible classification result of the artificially set training data sets. The other is "clustering", that is, the data in the training set would be automatically divided into several groups based on characteristics or some potential concepts. Each group is called a cluster, and then the commonalities of these clusters are artificially summarized through unsupervised learning. Although machine learning has some advantages in the diagnosis and recognition of PTSD in children, its application is still in the initial stage, with opportunities and challenges coexisting. It is worth noting that machine learning also has limitations such as a single algorithm, limited accuracy of prediction, different prediction results based on different models, relatively insufficient research on treatment methods, and difficulty in collecting children's PTSD indicators. In the future, researchers need to further improve the accuracy of machine learning diagnosis and children’s PTSD recognition, and explore more combinations of machine learning and traditional diagnosis methods. With the development of the Chinese medical industry, machine learning has shown great potential in the field of psychiatry. It is believed that the practical applications of machine learning in children’s PTSD would be developed rapidly in the future, which could provide guidance and suggestions for the early prevention or treatment of children’s PTSD.

Key words: machine learning, post-traumatic stress disorder, prognosis, children