A predictive model for secondary central nervous system infection after craniotomy based on machine learning

Abstract

To analyze the risk factors of secondary Central nervous system infections (CNSIs) after craniotomy, and to establish an individualized predictive model for CNSIs risk. The independent risk factors were screened by univariate and multivariate logistic regression analysis. Logistic regression, naive bayes, random forest, light GBM and adaboost algorithms were used to establish predictive models for secondary CNSIs after craniotomy. The predictive model based on the Adaboost algorithm demonstrated superior prediction performance compared to the other four models. Under 5-fold cross validation, the accuracy was 0.80, the precision was 0.69, the recall was 0.85, the F1-score was 0.76, the area under the ROC curve was 0.897,and the average precision was 0.880. The top 5 variables of importance in Adaboost model were operation time, indwelling time of lumbar drainage tube, indwelling lumbar drainage tube during operation, indwelling epidural drainage tube during operation, and GCS score. In addition, Adaboost model with the best prediction performance was used for clinical verification, and the prediction results were compared with the actual occurrence of CNSIs after surgery. The results showed that the accuracy of Adaboost model in predicting CNSIs was 60%, the accuracy of Adaboost model in predicting non-CNSIS was 92%, and the overall prediction accuracy was 76%.