Prediction Error as a Marker of Unexpected and Preventable Trauma Deaths: Nationwide Register-Based Cohort Study (Preprint)

BACKGROUND Trauma is a leading cause of death and disease burden worldwide. Multidisciplinary review of patient deaths is a standard method to identify preventable deaths and improve trauma care, but selecting which deaths to review is challenging. Whether the prediction error of a mortality prediction model can identify unexpected deaths that are judged preventable on review remains unknown. OBJECTIVE To develop prediction models for 30-day mortality in trauma patients, and to evaluate the association between model prediction error and preventable death. METHODS We conducted a register-based cohort study using the Swedish trauma registry SweTrau. We trained three LightGBM models with increasingly comprehensive predictor sets (static, physiological and interventions models) on data from all hospitals in SweTrau except the Karolinska University Hospital in Solna. Data from the excluded hospital, where a multidisciplinary board had reviewed deaths for preventability, served as evaluation data. The prediction error was defined as observed 30-day mortality minus predicted probability. We estimated the association between prediction error and preventable death using logistic regression among reviewed deaths that occurred within 30 days, and assessed the discrimination of the prediction error for preventable death using the area under the receiver operating characteristic curve (AUC). The study is reported according to TRIPOD+AI. RESULTS The training data included 78,566 patients, of whom 3,586 died within 30 days. The evaluation data included 14,913 patients; of the 864 who died within 30 days, 709 had been reviewed and 42 were classified as preventable deaths. The internal validation AUCs for 30-day mortality were 0.938 (95% CI 0.930 to 0.947) for the static model, 0.966 (0.960 to 0.972) for the physiological model and 0.967 (0.961 to 0.974) for the interventions model. Prediction error was positively associated with preventable death, with an odds ratio per unit of prediction error of 4.44 (95% CI 1.48 to 14.29) for the static model, 10.48 (4.10 to 28.15) for the physiological model and 10.14 (3.97 to 27.10) for the interventions model. The prediction error discriminated preventable from non-preventable deaths with an AUC of 0.62 (95% CI 0.54 to 0.69), 0.73 (0.65 to 0.80) and 0.71 (0.63 to 0.79), respectively. CONCLUSIONS Deaths that were unexpected according to a mortality prediction model were more likely to be judged preventable on expert review. Prediction error may be useful for selecting trauma deaths for peer review.