# Development of a predictive model for risk factors of multidrug-resistant bacterial pneumonia in critically ill post-neurosurgical patients

## TL;DR

This study develops a machine learning model to predict multidrug-resistant bacterial pneumonia risk in critically ill patients after neurosurgery, identifying key risk factors for better clinical decision-making.

## Contribution

The study introduces a Random Forest model with high accuracy for predicting MDR-BP risk and identifies critical risk factors using SHAP analysis.

## Key findings

- The Random Forest model achieved the highest mean accuracy (0.775) and AUC (0.860) in predicting MDR-BP risk.
- ICU length of stay, antibiotic treatment duration, and serum albumin level were identified as the most important predictors.
- SHAP analysis provided interpretable insights into feature importance for clinical decision-making.

## Abstract

Machine learning models have emerged as pivotal tools for enhancing the predictive accuracy of multidrug-resistant bacterial pneumonia (MDR-BP) risk in critically ill patients following neurosurgery procedures. By enabling early risk stratification, these models facilitate timely diagnosis and proactive therapeutic interventions. However, existing prediction frameworks exhibit limitations in elucidating the relative importance of risk factors, thereby impeding precise clinical decision-making and individualized patient management.

To evaluate the performance of six ensemble classification algorithms and three single classification algorithms in predicting MDR-BP risk factors among neurosurgical postoperative critically ill patients, identify the optimal predictive model, and determine key influential factors.

We conducted a retrospective study involving 750 neurosurgical patients admitted to a neurosurgery center at a tertiary hospital in Beijing between January 2020 and December 2023. Following rigorous data preprocessing, univariate analysis was performed to screen statistically significant variables. The Synthetic Minority Over-sampling Technique (SMOTE) was applied to address class imbalance. Predictive models for MDR-BP risk factors were constructed, and their performance was validated using 10-fold cross-validation to assess mean accuracy, recall, and specificity. The SHapley Additive exPlanations (SHAP) framework was employed to quantify feature importance.

The Random Forest model demonstrated superior performance, achieving the highest mean accuracy (0.775) and AUC value (0.860) compared to other models. SHAP interpretation revealed three critical predictors of MDR-BP: intensive care unit length of stay (ICU-LOS), antibiotic treatment duration, and serum albumin level.

The Random Forest algorithm demonstrates superior predictive accuracy for MDR-BP risk in critically ill post-neurosurgical patients. ICU-LOS, antibiotic treatment duration, and serum albumin level are significant predictors of MDR-BP.