关键词: 孤独症谱系障碍, 智力功能缺陷, 机器学习, 预测模型

Abstract: Objective To construct and evaluate a Gradient Boosting Machines(GBM) prediction model for children with autism spectrum disorder (ASD) and comorbid intellectual impairment, so as to provide a new perspective for early screening of this population. Method From January 2017 to December 2021, 241 children with a clear diagnosis of ASD in the Children's Hospital, Zhejiang University School of Medicine were included in the analysis. The prediction model of GBM was trained using sociodemographic and behavioral observation data and compared with traditional Logistic regression (LR) in this study. Hyperparameter adjustment was performed using grid search with ten-fold cross-validation, feature selection methods were performed using cross-validation LASSO, and the performance of the model was evaluated using discrimination and calibration. Explainability analysis was evaluated using SHapley Additive exPlanation (SHAP). Results The sample totaled 241 children with ASD, of whom 98 (40.66%) had intellectual impairments. Eight predictor variables were screened by the LASSO method, including language ability, mother's education attainment, age at the time of behavioral observation, stereotyped speech, pointing/gestures, social quality, unusual sensory interest and repetitive stereotyped behaviors. Both LR and GBM models before and after feature selection were better at distinguishing whether children with ASD had combined intellectual impairment. The area under curve (AUC) of the GBM model after feature selection (0.870, 95%CI: 0.749 - 0.989) was close to that of the conventional LR (0.851, 95%CI: 0.704 - 0.921). Regarding calibration, good calibration was observed for all models, except for the poor calibration of the full variable LR. In terms of feature importance, language ability contributed the most to the prediction of combined intellectual functioning deficits in children with ASD. Conclusion The prediction model for children with ASD complicated with intellectual impairment constructed by the feature selection method of LASSO and the GBM model has good performance and some clinical value.

Key words: autism spectrum disorder, intellectual impairment, machine learning, prediction model