An explainable Naive Bayes model temporally validated for early risk stratification of postoperative pulmonary infection after aSAH clipping (validation AUC 0.811), and an exploratory analysis found a nonlinear association between enteral nutrition start time and PPI risk with an inflection near 24 h, which needs prospective validation.
The Naive Bayes model showed the best overall performance among nine machine-learning algorithms for predicting postoperative pulmonary infection in the temporal validation cohort.
Candidate predictors were selected with a combined LASSO regression and random-forest importance strategy.
SHAP analysis was used to identify the major contributors to the primary 24-h prediction model, making the model explainable.
A secondary exploratory analysis found a nonlinear association between continuous enteral nutrition initiation timing and subsequent PPI risk, with an inflection point near 24 hours.
The authors conclude the model may support early PPI risk stratification after aSAH clipping, but the EN timing finding is exploratory and requires prospective validation before clinical use.
This study looked at patients who had surgical clipping of a ruptured brain aneurysm (aneurysmal subarachnoid hemorrhage). Lung infection after surgery is a common complication in these patients and is linked to slower recovery and greater use of hospital resources. The researchers built and compared nine machine-learning models to predict who would develop a postoperative lung infection, and tested the best one on a later group of patients. The best model, based on Naive Bayes, correctly separated patients who did and did not develop infection reasonably well (AUC 0.811 in the validation group). The researchers used an explainability method (SHAP) to show which factors drove the predictions. This research suggests such a tool could help identify higher-risk patients early, so that clinicians can focus monitoring and preventive efforts. The researchers also explored when tube or enteral feeding was started. They saw a nonlinear pattern, with the risk pattern changing around 24 hours after the start of the observation period (estimated 24.0 hours, range 23.0-27.0 hours). This was a secondary, exploratory finding from a retrospective study, so it cannot show that starting feeding at a certain time prevents infection. The authors say prospective studies are needed before this is used in practice.
Li L, Li X, Shi S. (2026). Explainable machine learning-based prediction of postoperative pulmonary infection after aneurysmal subarachnoid hemorrhage clipping and exploratory analysis of enteral nutrition timing: a dual-cohort retrospective study.. Frontiers in neurology. https://doi.org/10.3389/fneur.2026.1873566