关键词: 流感, 人；肺炎；个体化诊断预测模型

Abstract: Objective  To build an individualized prediction model for children with influenza A complicated with severe pneumoniausing machine learning technology ,and to fully explore high-risk factors, in order to improve the prognosis of children by early implementation of accurate preventive measures and treatment methods.
Methods  The clinical data of children with influenza A infected pneumonia were collected. Two sets of diagnosis and prediction models were constructed based on full clinical indexes and commonly used clinical indexes by using machine learning XGBoost algorithm, and the diagnostic value of the two models and the importance ranking of the corresponding indexes for prediction were compared.
Results  According tothe accuracy of the full index prediction model,the model diagnosis results and consistency of the observation results,Kappa value, area under curve(AUC), sensitivity, positive predictive value and negative predictive value of full index prediction modelwere better than those of commonly used clinical indexes. Based on full index prediction model,the main predictors of the model according to importance rankingwere respiratory failure, severity of cough and asthma(severe wheezing), total B cells, monocytes,duration offever(>15 d) andhelper T cells(CD4). The main predictors of the diagnostic prediction model based on commonly used clinical indexesaccording to ranking importance were respiratory failure, chest X-ray(pneumonia), mixed bacteria or virus. The gain value(X axis) of each index of the full index prediction model was significantly better than that of indexes inthe commonly used clinicalindexprediction model.
Conclusion  Based on the clinical data of children with influenza A complicated with pneumonia, using machine learning technology, children with influenza A complicated with critical pneumonia serveas variables, and all indexessuch asbasic clinical data and biochemical, immune,and imaging indexes serveas predictors. The individualized diagnostic prediction model has a higher diagnostic value, which could predict the probability of individuals complicated with critically ill pneumonia, and target high-risk factors for individuals, thereby preventing and treating children with influenza A complicated with critically ill pneumonia early and improving the prognosis.

Key words: influenza, human, pneumonia, individualized diagnosis prediction models