Building a SHAP interpretable prediction model for stroke-associated pneumonia [stroke-associated pneumonia (SAP)] in patients with aneurysmal subarachnoid hemorrhage (aSAH).

Background: Aneurysmal subarachnoid hemorrhage (aSAH) is a neurological emergency characterized by intracranial aneurysm rupture, leading to blood influx into the subarachnoid space and imposing high mortality and disability rates. The aim of this study was combining the epidemiological characteristics of aSAH and the high incidence of stroke-associated pneumonia (SAP), we aimed to develop a SHAP-explainable prediction model that may provide clinical reference value, providing a new approach to reduce the burden of complications. We take the random forest (RF) model combined with SHAP interpretation as the primary predictive model, and the logistic regression-based nomogram as a supplementary transparent tool for clinical bedside application. Methods: A retrospective analysis was conducted in aSAH patients at Huizhou Central People's Hospital. The patients were randomly split into training and validation sets for methodological purposes to generate and validate a SHAP interpretable machine learning model. Results: A total of 375 patients were initially enrolled, with 85 excluded due to exclusion criteria, leaving 290 patients for retrospective analysis. Univariate logistic regression, LASSO regression, and Boruta algorithm were used for multivariable feature selection and the common factors across the three methods were lactate dehydrogenase (LDH, SIRI, BMI, Age, AST, D-Dimer, and Hunt-Hess score). We established a nomogram model with good predictive performance, low error rate, and significant clinical benefits. The RF model showed stable performance and good efficacy in both training and validation sets. Based on the RF model, SHAP analysis was conducted to evaluate risk factor importance and individual impacts. The RF model with SHAP interpretation was identified as the primary predictive model, while the nomogram served as a supplementary transparent tool. Conclusion: This study identifies postoperative stroke-associated pneumonia [stroke-associated pneumonia (SAP)] within 14 days after endovascular embolization BMI, Age, SIRI, Hunt-Hess score, D-Dimer, AST, and lactate dehydrogenase LDH as key predictors of postoperative stroke-associated pneumonia (SAP) in aSAH, and demonstrates the efficient performance of the RF random forest model (with SHAP interpretation) in prediction.