Unintentional Injury Predicton Among Children aged 0-10 years old in Shenzhen, China: Based on Machine Learning Models

Abstract Background Unintentional injury is a leading cause of death and years of healthy life lost due to disability among children. This study aimed to develop machine learning (ML) models to predict the occurrence of unintentional injury in children aged 0–10 years in Shenzhen, China, and to identify the associated influencing factors. Methods This cross-sectional study recruited 15,796 children aged 0–10 years in Shenzhen during 2017–2018. Information on the children and guardians were collected, and unintentional injuries in the past year was examined by using two nested questions. The dataset was randomly partitioned into training and test sets at a 7:3 ratio. Class imbalance in the training set was addressed using data balancing techniques. Subsequently, seven machine learning algorithms were employed to establish a pediatric injury risk prediction model. The top five predictors for injury were ultimately identified based on the optimal performing model. Results Among 15,796 children surveyed, 915 (5.79%) experienced at least one unintentional injury within the past year. Falls constituted a leading cause of injuries, with 53.06% of such incidents occurring at home. On the training set, The Random Forest (RF) model integrated with SMOTE demonstrated optimal performance, achieving an AUC of 0.986, sensitivity of 0.955, and specificity of 0.988. On the test set, the prediction effect of the Extreme Gradient Boosting model combined with Random Under-Sampling (RUS) outperformed other algorithms, yielding an AUC of 0.593, sensitivity of 0.540, and specificity of 0.582. The children’s grade was the most important predictor of child injury. Conclusions Our research indicated that ML models, when coupled with data balancing techniques, prove to be potent instruments for predicting child injury in scenarios involving imbalanced datasets.