Predicting Pneumonia After Brain Tumor Surgery

Key Takeaways

• Across more than 1,000 patients undergoing brain tumor surgery at four Chinese centers, early postoperative pneumonia occurred within 7 days in part of the cohort.
• A logistic regression model using routine perioperative variables was selected after comparison with other algorithms and remained discriminative in prospective testing.
• Performance stayed consistent across three outside centers, with sensitivity ranging from 75.8% to 96.2% and specificity from 82.3% to 86.6%.
• SHAP analysis, a nomogram, and a web calculator were used to make the final risk model more interpretable and easier to apply.

Early postoperative pneumonia after brain tumor surgery can be difficult to recognize because neurologic deficits, dysphagia, sedation, and respiratory decline often overlap in the first postoperative week. Aspiration risk, impaired airway clearance, and fluctuating mental status can further blur the usual clinical signals of infection during immediate recovery. The unresolved question is whether routinely collected perioperative information can identify which patients are at higher risk before pneumonia becomes clinically obvious, prompting evaluation of an early warning model built from data already captured around surgery.

In China, a multicenter study combined one retrospective derivation cohort with four prospective validation cohorts, including one internal temporal cohort and three external cohorts, among patients undergoing craniotomy for intracranial brain tumor resection at tertiary hospitals. The analysis included 1,856 patients with 242 early postoperative pneumonia events. Early postoperative pneumonia was defined by Centers for Disease Control and Prevention (CDC)-based criteria within 7 days after surgery, and predictors were limited to information available from admission through postoperative day 3 to reduce information leakage. Among 41 candidate perioperative variables, 11 were retained, and 10 algorithms were compared before logistic regression was locked as the final model. The retained predictors were body mass index, diabetes mellitus, smoking, chronic lung disease, admission Karnofsky Performance Status, preoperative albumin, preoperative glucose, preoperative chemoradiotherapy, American Society of Anesthesiologists score, convexity tumor location, and postoperative glucose.

In prospective validation of the final logistic regression model, the internal temporal cohort showed an area under the curve (AUC) of 0.897 (95% confidence interval [CI] 0.842-0.952), with accuracy 0.853, sensitivity 0.784, and specificity 0.866. Across the three external centers, AUCs were 0.853, 0.928, and 0.847, with a mean external AUC of 0.876 ± 0.044; calibration remained good, sensitivity and specificity stayed within the reported ranges, and decision curve analysis favored logistic regression across a broad range of threshold probabilities.

All participating sites were Chinese tertiary hospitals, so applicability beyond similar neurosurgical settings remains uncertain, and North American applicability remains untested because no U.S. or Canadian centers were included. The derivation cohort was retrospective, outcome labeling could have varied because imaging frequency, radiographic interpretation, microbiologic testing, and diagnostic thresholds were not identical across centers, and several intraoperative or postoperative process-of-care variables were unavailable or inconsistently recorded. The authors framed the model as an early warning and triage tool rather than a standalone diagnostic test.

The authors concluded that early postoperative pneumonia risk after brain tumor surgery can be estimated with an interpretable 11-variable logistic regression model built from routinely available perioperative data and translated into Shapley additive explanations (SHAP), a nomogram, and a web-based calculator.

Clinician Questions

Which perioperative variables were included in the final model for early postoperative pneumonia after brain tumor surgery?

The final model used 11 routine perioperative variables for risk estimation: body mass index, diabetes mellitus, smoking, chronic lung disease, admission Karnofsky Performance Status, preoperative albumin, preoperative glucose, preoperative chemoradiotherapy, American Society of Anesthesiologists score, convexity tumor location, and postoperative glucose.

How was early postoperative pneumonia defined after brain tumor surgery in this model?

Early postoperative pneumonia was defined within 7 days of surgery using CDC-based criteria requiring a new or progressive pulmonary infiltrate on chest radiography or computed tomography plus compatible clinical findings such as fever, purulent sputum, leukocyte abnormality, or worsening oxygenation.

How broadly does this brain tumor surgery pneumonia model apply outside the study centers?

The model was tested across Chinese tertiary hospitals, including one internal temporal cohort and three external centers, so it was evaluated across multiple institutions; however, transportability beyond similar neurosurgical settings remains uncertain, and North American applicability remains untested because no U.S. or Canadian centers were included.