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

Abstract

Background/Objectives: Routine postoperative ICU admission following brain tumor surgery may not benefit selected patients. The objective of this study was to develop a risk prediction instrument for early (within 24 h) postoperative adverse events using machine learning techniques. Methods: Retrospective cohort of 1000 consecutive adult patients undergoing elective brain tumor resection. Nine events/interventions (CPR, reintubation, return to OR, mechanical ventilation, vasopressors, impaired consciousness, intracranial hypertension, swallowing disorders, and death) were chosen as target variables. Potential prognostic features (n = 27) from five categories were chosen and a gradient boosting algorithm (XGBoost) was trained and cross-validated in a 5 × 5 fashion. Prognostic performance, potential clinical impact, and relative feature importance were analyzed. Results: Adverse events requiring ICU intervention occurred in 9.2% of cases. Other events not requiring ICU treatment were more frequent (35% of cases). The boosted decision trees yielded a cross-validated ROC-AUC of 0.81 ± 0.02 (mean ± CI95) when using pre- and post-op data. Using only pre-op data (scheduling decisions), ROC-AUC was 0.76 ± 0.02. PR-AUC was 0.38 ± 0.04 and 0.27 ± 0.03 for pre- and post-op data, respectively, compared to a baseline value (random classifier) of 0.092. Targeting a NPV of at least 95% would require ICU admission in just 15% (pre- and post-op data) or 30% (only pre-op data) of cases when using the prediction algorithm. Conclusions: Adoption of a risk prediction instrument based on boosted trees can support decision-makers to optimize ICU resource utilization while maintaining adequate patient safety. This may lead to a relevant reduction in ICU admissions for surveillance purposes.

Keywords: ICU; complications; craniotomy; machine learning; postoperative surveillance.

Figure 1

Prognostic performance. Pre- and post-op features yield a ROC-AUC 0.81 ± 0.02 (mean ± CI₉₅) in 5-times repeated 5-fold cross-validation (p < 0.01). Using only pre-op features, ROC-AUC still scores at 0.76 ± 0.02 (p < 0.001, (A)). As the underlying class distribution is clearly skewed towards the negative class (1:9), the AUC of the precision-recall curve was expected to be lower than AUC-ROC. In the case of pre- and post-op data, AUC-PR was 0.38 ± 0.04, and 0.27 ± 0.03 for pre-op data only (both p < 0.001, (B)). Compared to the baseline value of a random classifier (0.09), these numbers represent a 3-fold increase in the baseline value and represent a good classification performance.

Figure 2

Relationship between negative predictive value (NPV) and ICU admission rate as a function of the selected threshold. With pre- and post-op data, targeting an NPV of at least 95% requires ICU admission in 15% of cases. Using only preoperative data, approximately 30% of cases were selected for surveillance at the ICU. Higher target NPVs lead to continuously rising ICU admission rates.

Figure 3

Relative contribution of features to the model output.