Machine learning models for predicting cerebrospinal fluid bacterial culture outcomes in post-neurosurgical...

Post-neurosurgical bacterial meningitis (PNBM) is a serious complication associated with substantial morbidity and mortality. Bacterial culture presently serves as the gold standard for diagnosis, however, clinical reliance on third-tier reports to address empirical antibiotic concerns, with a turnaround time of 3-5 days, can lead to imprecise antibiotic administration in patients. Therefore, we develop and validate a branch of machine learning method to predict the result of cerebrospinal fluid (CSF) culture, which can shorten the time for precise diagnosis to within 2 h. A cohort study was conducted in nonmaternity inpatient units at Beijing Tiantan Hospital & Capital Medical University during Jan 2016 to Dec 2020. Six machine learning models including 4 decision tree models(Random Forest, Catboost, LightGBM, and XGboost), Support Vector Classification(SVC) and Multi-Layer Perceptron(MLP) were applied for two models construction by utilizing tenfold cross-validation.The performances of these models were assessed in terms of discrimination, calibration, and clinical application in another 3 years cohort. In this study, the derivation cohort included 4014 patients, of whom 995 had positive CSF cultures, including 431 Gram-negative and 564 Gram-positive cultures, and 3019 had negative cultures. The temporal validation cohort included 1059 patients. LightGBM achieved the highest the area under the receiver operating characteristic curve (AUC) for Model 1 and Model 2, at 0.898 and 0.875, respectively. SHapley Additive exPlanations (SHAP) analysis identified CSF leukocyte(C-Leu) count as the leading predictor of culture positivity and Gram-negative cultures, whereas platelet count was the principal contributor to Gram-positive culture classification. Machine learning, specifically leveraging LightGBM modeling showed good discrimination in predicting CSF bacterial culture outcomes after neurosurgery. The model may support early risk stratification while culture results are pending.