Auxiliary diagnosis model of children with autism spectrum disorder based on random forest

Abstract

Objective The random forest algorithm was used to construct a rapid screening diagnostic prediction model for children with autism spectrum disorder, to provide the references for early detection, early diagnosis of ASD children, and to reduce the pressure of ASD clinical diagnosis and assessment. Methods The random forest algorithm of machine learning was applied to build the auxiliary diagnosis model. Totally 346 ASD children and 90 normal children were evaluated by Social Responsiveness Scale and Vineland Adaptive Behavior Scales. ROC curve, and accuracy was used to evaluate the models. Results Among the models, the accuracy of 13 feature factors and 7 feature factors were above 0.9, the sensitivity was up to 0.927, the specificity was up to 0.936 and the AUC was up to 0.979. The accuracy, sensitivity, specificity and AUC of the model were 0.943, 0.959, 0.931 and 0.978 respectively. The fitting and generalization effects of the three models were all satisfactory. Conclusion A random forest model based on the SRS Scales and Vineland Adaptive Behavior Scales can be used to diagnose ASD accurately and provide scientific basis for the development of rapid screening and diagnosis tools. 【摘要】 目的 利用随机森林算法构建孤独症谱系障碍(autism spectrum disorder, ASD)儿童快速辅助诊断模型, 有助于 ASD儿童的早期发现、早期诊断, 减轻临床诊断及评估压力。 方法 采用机器学习中随机森林算法, 应用社交反应量表 (SRS)及文兰适应行为量表(VABS)对黑龙江省 346 名 ASD 儿童和 90 名健康儿童进行评估, 并基于量表数据以及儿童基础信息构建预测模型, 运用 ROC 曲线及准确率等指标评价模型拟合效果。 结果 得到的随机森林预测模型中, 13 个特征因素模型以及 7 个特征因素的预测模型准确率均达到 0.9 以上、灵敏度最高达到 0.927, 特异度最高达到 0.936, AUC 值为 0.979; 以年龄为筛选条件的模型准确率达到 0.943, 灵敏度达到 0.959, 特异度达到 0.931, AWC 值为 0.978。3 个模型的拟合和泛化效果都较为理想。 结论 采用社交及适应能力水平指标构建的随机森林模型可以较为精确辅助开展 ASD 的诊断, 为开发快速筛查和诊断的辅助工具提供了科学依据。