Predicting mortality risk in pediatric severe pneumonia using a CNN-BiLSTM model with dynamic clinical indicators

Severe pneumonia is a leading cause of mortality in children under 5 years old, and there is currently a lack of reliable biomarkers for early warning of adverse outcomes in the ICU. This study aimed to explore effective prognostic biomarkers for severe pneumonia and apply a deep learning model to predict in-hospital mortality in children with severe pneumonia, thereby supporting clinical decision-making.

This retrospective prognostic study analyzed clinical data from children under 5 years of age with severe pneumonia from the Pediatric Intensive Care (PIC) database. Patients were categorized according to in-hospital mortality status. Dynamic biomarker screening was performed from longitudinal laboratory data, and a CNN-BiLSTM-based prediction model was subsequently constructed using the selected biomarkers.

Eleven key predictive indicators were identified from the laboratory parameters. The CNN-BiLSTM model achieved an area under the curve (AUC) of 0.956 on an independent test set, with a sensitivity of 85.7% and a specificity of 92.7%. Interpretability analysis revealed that lactate, partial pressure of carbon dioxide (pCO₂), and pH were the most influential predictors.

This study provides an effective tool for dynamic risk stratification in children with severe pneumonia, offering support for timely clinical decision-making in the ICU. Future multi-center studies are still needed to validate the effectiveness of this model.