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

Area of Science:

• Pediatric critical care medicine
• Biomarker discovery
• Artificial intelligence in healthcare

Background:

• Severe pneumonia is a major cause of mortality in children under 5.
• Lack of reliable early warning biomarkers for adverse outcomes in pediatric ICUs.
• Need for improved clinical decision-support tools for severe pediatric pneumonia.

Purpose of the Study:

• To identify prognostic biomarkers for severe pneumonia in children.
• To develop a deep learning model for predicting in-hospital mortality.
• To support clinical decision-making in pediatric intensive care units.

Main Methods:

• Retrospective analysis of clinical data from a pediatric intensive care database.
• Screening of longitudinal laboratory data for dynamic biomarker identification.
• Construction of a CNN-BiLSTM deep learning model using selected biomarkers.

Main Results:

• Eleven key predictive laboratory indicators were identified.
• The CNN-BiLSTM model demonstrated high predictive performance (AUC=0.956, sensitivity=85.7%, specificity=92.7%).
• Lactate, pCO2, and pH were identified as the most influential predictors.

Conclusions:

• The study presents an effective tool for dynamic risk stratification in pediatric severe pneumonia.
• The model supports timely clinical decision-making for critically ill children.
• Further multi-center validation is recommended to confirm the model's effectiveness.