Related Experiment Video
Updated: Jun 30, 2026
The Stroke Preclinical Assessment Network Multi-Laboratory Model of Thromboembolic Stroke with Thrombolysis: TE-MCAo
Published on: December 19, 2025
Development and external validation of an interpretable machine learning model for early prediction of
Mingyang Zhao1, Qianyu Zhou1, Qiang Zhang2
1College of Public Health, Zhengzhou University, Zhengzhou, Henan 450001, PR China.
Objective:
To develop and externally validate an interpretable machine learning model for predicting 7-day stroke-associated pneumonia (SAP) using routine clinical data collected within 24 h of admission.
Methods:
This multicenter study utilized a development cohort from the Henan Stroke Cohort and an independent external validation cohort from three Zhengzhou hospitals. Adult patients with imaging-confirmed ischemic or hemorrhagic stroke were eligible. Patients with infection at admission or who developed SAP within 24 h of admission were excluded. We evaluated 26 candidate predictors obtained within 24 h of admission. Nine machine learning algorithms were trained following recursive feature elimination. Model performance was evaluated based on discrimination, calibration, and clinical utility. The optimal model was interpreted using SHapley Additive exPlanations (SHAP) and deployed as an online calculator.
Results:
The development and external validation cohorts comprised 1201 and 645 patients, with 7-day SAP incidences of 20.6 % (n = 247) and 24.8 % (n = 160), respectively. Among the nine algorithms, stochastic gradient boosting (SGBT) demonstrated the most balanced overall performance. In internal validation, SGBT achieved an area under the receiver operating characteristic curve (AUC) of 0.947 in the training set and 0.905 in the test set. In external validation, the model yielded an AUC of 0.906, alongside an accuracy of 0.864, sensitivity of 0.712, specificity of 0.918, positive predictive value (PPV) of 0.756, negative predictive value (NPV) of 0.899, and F1-score of 0.733. The final model retained 10 predictors: stroke subtype, fibrinogen, D-dimer, C-reactive protein, uric acid, triglycerides, homocysteine, and clinical scores (ADL, GCS, NIHSS). SHAP analysis identified early neurological impairment and inflammatory burden as the primary contributors to SAP prediction.
Conclusion:
An interpretable SGBT model utilizing routine admission data accurately predicted 7-day SAP and remained robust during external validation. The accompanying online calculator facilitates individualized risk estimation to support early preventive decision-making in hospitalized patients with stroke.