Construction and effectiveness test of multimodal data fusion prediction model for intracranial infection after severe craniocerebral injury in children based on deep learning.

objectiveTo develop and validate a multimodal data fusion prediction model based on deep learning for the early postoperative identification of intracranial infection in pediatric patients with severe traumatic brain injury (TBI).

methodsA total of 203 pediatric TBI patients who underwent surgery at Children's Hospital, Zhejiang University School of Medicine between March 2022 and May 2025 were included as the internal validation cohort. These patients were stratified into infection group (46 cases) and non-infection group (157 cases) based on the occurrence of postoperative infection. General clinical data were compared between the two groups, and multivariate logistic regression analysis was performed to identify risk factors for postoperative infection. Radiomic features and deep learning features were extracted and screened. Additionally, 101 pediatric patients who underwent surgery during the same period were selected as the temporal validation cohort (25 infected cases and 76 non-infected cases). Four predictive models were constructed for the internal temporal validation cohort: a radiomics model, a clinical model, a deep learning model, and a combined model. The predictive performance of these models was evaluated using receiver operating characteristic (ROC) curves, clinical decision curve analysis, and calibration curves.

resultsStatistically significant differences (P < 0.05) were observed between the infection and non-infection groups in terms of operative duration, preoperative Glasgow Coma Scale (GCS) score, external ventricular drainage (EVD), incisional cerebrospinal fluid (CSF) leakage, incisional effusion, number of surgeries, hypersensitive C-reactive protein (hs-CRP), procalcitonin (PCT), and blood lactate levels. Logistic regression analysis identified EVD, incisional effusion, hs-CRP, incisional CSF leakage, and PCT as independent risk factors for postoperative intracranial infection in pediatric TBI patients (P < 0.05). A total of 9 radiomic features and 20 deep learning features were retained. ROC and DCA analyses demonstrated that the combined prediction model exhibited superior predictive performance and greater clinical net benefit compared to the clinical, radiomics, and deep learning models alone.

conclusionThe multimodal data fusion prediction model integrating deep learning, radiomics, and clinical data demonstrates excellent performance in the early postoperative monitoring and identification of intracranial infection in pediatric severe TBI.