Routine ICU Surveillance after Brain Tumor Surgery: Patient Selection Using Machine Learning

Area of Science:

• Neurosurgery
• Machine Learning in Medicine
• Critical Care Medicine

Background:

• Routine intensive care unit (ICU) admission post-brain tumor surgery may not be necessary for all patients.
• Developing predictive tools can help identify patients who truly require ICU monitoring.

Purpose of the Study:

• To create a machine learning-based risk prediction instrument for early postoperative adverse events (within 24 hours) following brain tumor resection.
• To optimize ICU resource allocation by identifying patients who can safely avoid routine ICU admission.

Main Methods:

• Retrospective analysis of 1000 adult patients undergoing elective brain tumor resection.
• Utilized a gradient boosting algorithm (XGBoost) trained on 27 potential prognostic features.
• Cross-validation (5x5) was performed to assess model performance, including ROC-AUC and PR-AUC.

Main Results:

• The prediction model achieved a cross-validated ROC-AUC of 0.81 ± 0.02 using pre- and postoperative data.
• Using only preoperative data, the ROC-AUC was 0.76 ± 0.02.
• The model demonstrated potential to reduce ICU admissions to 15% while maintaining a negative predictive value (NPV) of at least 95%.

Conclusions:

• A risk prediction instrument based on boosted decision trees can effectively support clinical decision-making.
• Optimizing ICU resource utilization is achievable without compromising patient safety.
• This approach can lead to a significant reduction in surveillance-based ICU admissions.