A Real-Time Dynamic Warning Method for MODS in Trauma Sepsis Patients Based on a Pre-Trained Transfer Learning Algorithm

Background and objectives: Multiple organ dysfunction syndrome (MODS) is a common and severe complication associated with poor prognosis in trauma sepsis patients. Early prediction is essential for identifying high-risk patients and enabling timely intervention. This study aims to develop a high-performance, interpretable, and multicenter-validated prediction model to support individualized risk assessment and guide precise therapeutic strategies.Methods: Based on the MIMIC-IV and eICU databases, a MODS prediction model was developed using a pre-trained transfer learning algorithm, and a separation processing strategy was adopted to optimize the model performance. For validation, 700 patients from MIMIC-IV were used for internal validation, and 110 patients from eICU were used for multicenter external validation. Model interpretability was assessed using the SHAP algorithm to identify key predictive features.Results: Across the 6, 12, and 24-hour prediction windows, the best pre-trained transfer model achieved an average AUC of 0.906. Notably, when fine-tuned using only 100 cases (3.6% of the total training set) of trauma sepsis patients, the pre-trained model still reached an AUC of 0.846, which was 0.165 higher than that of the non-pre-trained model. SHAP analysis further revealed that platelet count was one of the most important variables contributing to MODS prediction.Conclusion: The MODS prediction model based on a pre-trained approach demonstrated robust performance in both internal and external validation among trauma sepsis patients. With its strong interpretability, generalization capability, and portability, the MODS prediction model demonstrated potential for clinical application.