Development and validation of radiology-clinical statistical and machine learning model for stroke-associated pneumonia after first intracerebral haemorrhage

Abstract

Abstract Background Society is burdened with stroke-associated pneumonia (SAP) after intracerebral haemorrhage (ICH). Cerebral small vessel disease (CSVD) complicates clinical manifestations of stroke. In this study, we redefined the CSVD burden score and incorporated it into a novel radiological-clinical prediction model for SAP. Materials and methods A total of 1278 patients admitted to a tertiary hospital between 1 January 2010 and 31 December 2019 were included. The participants were divided into training and testing groups using fivefold cross-validation method. Four models, two traditional statistical models (logistic regression and ISAN) and two machine learning models (random forest and support vector machine), were established and evaluated. The outcomes and baseline characteristics were compared between the SAP and non-SAP groups. Results Among the of 1278 patients, 281(22.0%) developed SAP after their first ICH. Multivariate analysis revealed that the logistic regression (LR) model was superior in predicting SAP in both the training and testing groups. Independent predictors of SAP after ICH included total CSVD burden score (OR, 1.29; 95% CI, 1.03–1.54), haematoma extension into ventricle (OR, 2.28; 95% CI, 1.87–3.31), haematoma with multilobar involvement (OR, 2.14; 95% CI, 1.44–3.18), transpharyngeal intubation operation (OR, 3.89; 95% CI, 2.7–5.62), admission NIHSS score ≥ 10 (OR, 2.06; 95% CI, 1.42–3.01), male sex (OR, 1.69; 95% CI, 1.16–2.52), and age ≥ 67 (OR, 2.24; 95% CI, 1.56–3.22). The patients in the SAP group had worse outcomes than those in the non-SAP group. Conclusion This study established a clinically combined imaging model for predicting stroke-associated pneumonia and demonstrated superior performance compared with the existing ISAN model. Given the poor outcomes observed in patients with SAP, the use of individualised predictive nomograms is vital in clinical practice.

Keywords

• Intracerebral haemorrhage
• Pneumonia
• Cerebral vascular disease
• Machine learning
• Risk factor