机器学习在儿童创伤后应激障碍识别及转归预测中的应用

创伤后应激障碍(PTSD)会给儿童发展带来负面效应,其影响甚至延续至成年期。然而传统诊断方式难以做到快速、客观、准确的识别和诊断儿童PTSD,机器学习作为一种处理大量变量和数据的新兴方法,逐渐被应用到儿童PTSD的早期预测、识别及辅助诊断等研究中。机器学习凭借其性能、原理等方面的优势,可被应用在儿童PTSD的识别与转归领域。相比自我报告式的诊断,通过机器学习辅助识别和诊断儿童PTSD的过程具有效率高、客观准确、节约资源等独特优势。然而,机器学习也在硬件成本、算法选择和预测准确度等方面存在局限性。未来研究人员需要进一步提高机器学习诊断识别儿童PTSD的准确率,并将机器学习算法同传统诊断方法结合进行更多的探索和应用。

Application of machine learning in prognosis and trajectory of post-traumatic stress disorder in children

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.