An inflammatory-nutritional machine learning model for risk stratification of hospital-acquired pneumonia in traumatic brain injury: a multicenter study.

BACKGROUND: Hospital-acquired pneumonia (HAP) is a frequent and serious complication following traumatic brain injury (TBI), leading to prolonged hospitalization and poor functional outcomes. Early identification of patients at high risk of HAP remains challenging. Systemic inflammation and nutritional status are recognized contributors to post-TBI infection susceptibility; however, these factors are not adequately incorporated into existing predictive models. This study aimed to develop and validate an inflammatory-nutritional machine learning model for predicting HAP after TBI and to evaluate its prognostic stratification performance using an independent testing cohort from a second center.

METHODS: A total of 567 adult patients with TBI were included in this retrospective multicenter cohort study conducted at two hospitals. Patients were divided into a training set (n = 396) and an independent testing set (n = 171). Baseline laboratory data obtained within 24 h of admission were used to calculate the Pan-Immune-Inflammation Value (PIV) and Prognostic Nutritional Index (PNI) scores. HAP, defined as occurring ≥48 h after admission, was designated as the primary outcome, and functional prognosis at discharge was assessed using the modified Rankin Scale (mRS). Six machine learning models were constructed and compared. Model performance was evaluated using discrimination, calibration, decision curve analysis, and 10-fold cross-validation. Model interpretability was assessed with Shapley Additive Explanations (SHAP), and Kaplan-Meier analyses were conducted for prognostic stratification.

RESULTS: The Light Gradient Boosting Machine exhibited the best overall performance, achieving an area under the receiver operating characteristic curve of 0.815 in the testing cohort, with good calibration and superior clinical net benefit. Cross-validation confirmed stable predictive capability. SHAP analysis identified PIV as the most influential predictor, followed by PNI, demonstrating consistent feature importance across cohorts. Model-derived risk stratification was significantly associated with functional outcomes, with high-risk patients exhibiting a markedly lower likelihood of favorable prognosis (mRS 0-2) in both cohorts.

CONCLUSION: The inflammatory-nutritional machine learning model integrating PIV and PNI provides accurate and interpretable prediction of HAP after TBI and effectively stratifies functional prognosis, supporting its potential value for early risk assessment and future individualized decision-support in patients with TBI, pending prospective validation.