Prediction of respiratory failure in severe pediatric pneumonia via CT and electronic medical records fusion

To construct a prediction model for respiratory failure in children with severe pneumonia by integrating CT images and electronic medical records, aiming to improve the accuracy of clinical risk screening and provide a quantitative basis for early intervention in high-risk children. The clinical data of 364 children with severe pneumonia admitted to the pediatric department of our hospital from January 2020 to February 2024 were retrospectively included. The patients were randomly divided into a training set ( n = 255) and a validation set ( n = 109) at a ratio of 7:3. 20 indicators were integrated, including the patients’ demographic data, clinical test indicators, imaging features, and electronic medical record text information. Relevant predictive factors were screened through univariate analysis. Feature dimension reduction was performed using LASSO regression, and independent risk factors were determined by multivariate Logistic regression. Three types of prediction models, namely Random Forest (RF), Support Vector Machine (SVM), and Gradient Boosting Machine (GBM), were constructed. The performance of the models was evaluated by area under the receiver operating characteristic curve (AUC), sensitivity, and specificity. The optimal model was selected, and the importance of key early-warning indicators was analyzed. There was no significant difference in the baseline data between the training set and the validation set ( P > 0.05). Multivariate Logistic regression analysis showed that FiO₂, C-reactive protein, serum lactate, the number of affected lung lobes, entropy of the gray-level co-occurrence matrix, contrast of the gray-level run-length matrix, and the absolute value of the first-order statistical skewness were independent risk factors for concurrent respiratory failure ( P < 0.05). In the training set, the RF model achieved the highest AUC of 0.893 (95% CI: 0.826–0.961), outperforming the SVM (0.836, 95% CI: 0.756–0.916) and GBM (0.854, 95% CI: 0.783–0.924) models. In the validation set, the RF model maintained the best discriminative performance. DeLong tests confirmed significant differences between RF and both SVM ( p = 0.018) and GBM ( p = 0.042). The RF prediction model constructed by integrating CT images and electronic medical record data has good discriminatory performance. The contrast of the gray-level run-length matrix, entropy of the gray-level co-occurrence matrix, the absolute value of the first-order statistical skewness, and the number of affected lung lobes are the core early-warning factors, which shows potential to assist in the early identification of respiratory failure in children with severe pneumonia.