Mengqi Xie,1,* Zhiying Liu,1,* Fangfang Dai,1 Zhen Cao,2 Xiaobei Wang3,4 

1The Second Clinical Medical College of Xinjiang Medical University, Xinjiang Uygur Autonomous Region, People’s Republic of China; 2Department of Clinical Medicine, Xinjiang Medical University, Xinjiang Uygur Autonomous Region, People’s Republic of China; 3Department of Neurology, The Second Affiliated Hospital of Xinjiang Medical University, Xinjiang Uygur Autonomous Region, People’s Republic of China; 4Xinjiang Key Laboratory of Neurological Disorder Research, Xinjiang Uygur Autonomous Region, People’s Republic of China

*These authors contributed equally to this work

Correspondence: Xiaobei Wang, Department of Neurology, The Second Affiliated Hospital of Xinjiang Medical University, Xinjiang Uygur Autonomous Region, People’s Republic of China, Email wangxiaobei1203@163.com

Purpose: Stroke-associated pneumonia (SAP), a critical complication of ischemic stroke, significantly worsens outcomes. Our aim was to identify SAP risk factors and develop a machine learning (ML) model for early risk stratification.
Methods: This retrospective study analyzed 574 ischemic stroke patients, divided into training (75%) and testing (25%) sets. Nine ML models were trained using 10-fold cross-validation, with performance evaluated by accuracy, AUC-ROC, and F1-score. Key predictors were interpreted via SHAP analysis. An interactive web tool was developed using the optimal model.
Results: SAP incidence was 32.4%. LightGBM demonstrated superior predictive performance (ranking score=54) without overfitting, identifying Monocyte-to-lymphocyte ratio (MLR), systemic immune-inflammation index (SII), NIHSS score, age, aggregate index of systemic inflammation (AISI), and platelet-to-lymphocyte ratio (PLR) as the top predictors.
Conclusion: Our findings demonstrate that machine learning models exhibit strong predictive performance for SAP, with the LightGBM algorithm outperforming other approaches. The web-based prediction tool developed from this model provides clinicians with actionable insights to support real-time clinical decision-making.

Keywords: stroke-associated pneumonia, machine learning, ischemic stroke