[Construction and validation of a multimodal risk prediction model for burn sepsis in infants and toddlers based on deep learning].

Abstract

Objective: To construct and validate a multimodal risk prediction model (hereinafter referred to as the multimodal prediction model) for burn sepsis in infants and toddlers based on deep learning. Methods: This retrospective cohort study included 1 709 burned infants and toddler patients (hereinafter referred to as pediatric patients) with total burn area ranged from 1.0%-85.0% total body surface area who were admitted to the Department of Burn and Plastic Surgery of Children's Hospital of Nanjing Medical University and met the eligibility criteria from January 2015 to December 2025. According to clinical outcomes, pediatric patients were divided into sepsis group (n=54, 34 males and 20 females, aged 500.0 (270.3, 695.0) days) and non-sepsis group (n=1 655, 998 males and 657 females, aged 515.0 (365.0, 790.0) days). Clinical data collected within 24 h after admission were analyzed and compared between the two groups of pediatric patients, including the total burn area, full-thickness burn area, the proportion of pediatric patients with concomitant inhalation injury, elevated C-reactive protein (CRP), and elevated procalcitonin, prothrombin time (PT), albumin level, and white blood cell count. A multimodal prediction model was constructed by integrating burn wound image features and paired clinical scale features. Its performance was compared with that of random forest, XGBoost, TabTransformer, ResNet, Vision Transformer, and Swin Transformer models (six control models), and its internal performance was evaluated using five-fold cross-validation. Gradient-based feature importance analysis was performed to interpret the multimodal prediction model and to quantify the contribution of clinical features. Gradient-weighted class activation mapping was applied for visualization of the multimodal prediction model. Results: Compared with those in non-sepsis group, pediatric patients in sepsis group had significantly larger total burn area and full-thickness burn area (with Z values of 8.266 and 7.945, respectively, P<0.05). Additionally, the proportions of pediatric patients with concomitant inhalation injury, elevated CRP, and elevated procalcitonin, and white blood cell count were significantly higher (with χ² values of 4.914, 7.803, and 8.859, respectively, Z=1.969, P<0.05), while albumin levels were significantly lower (Z=-2.132, P<0.05), and PT was significantly prolonged (Z=-4.247, P<0.05). The multimodal prediction model achieved an area under the receiver operating characteristic curve of 0.931±0.009 for predicting post-burn sepsis of pediatric patients, which was higher than that of random forest, XGBoost, TabTransformer, ResNet, Vision Transformer, and Swin Transformer models (with values of 0.861±0.031, 0.862±0.025, 0.876±0.019, 0.897±0.015, 0.900±0.015, and 0.916±0.012, respectively). The Dice similarity coefficient of the multimodal prediction model for burn wound image segmentation was higher than that of the ResNet, Vision Transformer, and Swin Transformer models. Across five-fold cross-validation, the multimodal prediction model exhibited superior overall predictive performance compared to the six control models. Gradient-based feature importance analysis revealed that PT had the highest normalized importance weight (1.000), followed by diastolic blood pressure (0.855). The wound regions delineated by the multimodal prediction model were largely consistent with those manually annotated by physicians. Conclusions: The constructed multimodal prediction model demonstrates good discriminative performance and clinical application value for assessing the risk of post-burn sepsis in pediatric patients.