A novel machine learning-assisted clinical diagnosis support model for early identification of pancreatic injuries in patients with blunt abdominal trauma: a cross-national study

Background The recognition of pancreatic injury in blunt abdominal trauma is often severely delayed in clinical practice. The aim of this study was to develop a machine learning model to support clinical diagnosis for early detection of abdominal trauma. Methods We retrospectively analyzed of a large intensive care unit database (Medical Information Mart for Intensive Care [MIMIC]-IV) for model development and internal validation of the model, and performed prospective outer validation based on a cross-national data set. Logistic regression was used to develop three models (PI-12, PI-12-2, and PI-24). Univariate and multivariate analyses were used to determine variables in each model. The primary outcome was early detection of a pancreatic injury of any grade in patients with blunt abdominal trauma in the first 24 hours after hospitalization. Results The incidence of pancreatic injuries was 5.56% (n = 18) and 6.06% (n = 6) in the development (n = 324) and internal validation (n = 99) cohorts, respectively. Internal validation cohort showed good discrimination with an area under the receiver operator characteristic curve (AUC) value of 0.84 (95% confidence interval [CI]: 0.71–0.96) for PI-24. PI-24 had the best AUC, specificity, and positive predictive value (PPV) of all models, and thus it was chosen as the final model to support clinical diagnosis. PI-24 performed well in the outer validation cohort with an AUC value of 0.82 (95% CI: 0.65–0.98), specificity of 0.97 (95% CI: 0.91–1.00), and PPV of 0.67 (95% CI: 0.00–1.00). Conclusion A novel machine learning-based model was developed to support clinical diagnosis to detect pancreatic injuries in patients with blunt abdominal trauma at an early stage.