Rolf Gedeborg1, Bodil Svennblad2, Liisa Byberg3, Karl Michaëlsson3, Ingemar Thiblin4. 1. Dept. of Surgical Sciences, Anesthesiology and Intensive Care, Uppsala University, Uppsala, Sweden; Uppsala Clinical Research Center, Uppsala University, Uppsala, Sweden. 2. Uppsala Clinical Research Center, Uppsala University, Uppsala, Sweden. 3. Dept. of Surgical Sciences, Orthopedics, Uppsala University, Uppsala, Sweden. 4. Dept. of Surgical Sciences, Forensic Medicine, Uppsala University, Uppsala, Sweden. Electronic address: ingemar.thiblin@surgsci.uu.se.
Abstract
BACKGROUND: To predict mortality risk in victims of violent crimes based on individual injury diagnoses and other information available in health care registries. METHODS: Data from the Swedish hospital discharge registry and the cause of death registry were combined to identify 15,000 hospitalisations or prehospital deaths related to violent crimes. The ability of patient characteristics, injury type and severity, and cause of injury to predict death was modelled using conventional, Lasso, or Bayesian logistic regression in a development dataset and evaluated in a validation dataset. RESULTS: Of 14,470 injury events severe enough to cause death or hospitalization 3.7% (556) died before hospital admission and 0.5% (71) during the hospital stay. The majority (76%) of hospital survivors had minor injury severity and most (67%) were discharged from hospital within 1day. A multivariable model with age, sex, the ICD-10 based injury severity score (ICISS), cause of injury, and major injury region provided predictions with very good discrimination (C-index=0.99) and calibration. Adding information on major injury interactions further improved model performance. Modeling individual injury diagnoses did not improve predictions over the combined ICISS score. CONCLUSIONS: Mortality risk after violent crimes can be accurately estimated using administrative data. The use of Bayesian regression models provides meaningful risk assessment with more straightforward interpretation of uncertainty of the prediction, potentially also on the individual level. This can aid estimation of incidence trends over time and comparisons of outcome of violent crimes for injury surveillance and in forensic medicine.
BACKGROUND: To predict mortality risk in victims of violent crimes based on individual injury diagnoses and other information available in health care registries. METHODS: Data from the Swedish hospital discharge registry and the cause of death registry were combined to identify 15,000 hospitalisations or prehospital deaths related to violent crimes. The ability of patient characteristics, injury type and severity, and cause of injury to predict death was modelled using conventional, Lasso, or Bayesian logistic regression in a development dataset and evaluated in a validation dataset. RESULTS: Of 14,470 injury events severe enough to cause death or hospitalization 3.7% (556) died before hospital admission and 0.5% (71) during the hospital stay. The majority (76%) of hospital survivors had minor injury severity and most (67%) were discharged from hospital within 1day. A multivariable model with age, sex, the ICD-10 based injury severity score (ICISS), cause of injury, and major injury region provided predictions with very good discrimination (C-index=0.99) and calibration. Adding information on major injury interactions further improved model performance. Modeling individual injury diagnoses did not improve predictions over the combined ICISS score. CONCLUSIONS: Mortality risk after violent crimes can be accurately estimated using administrative data. The use of Bayesian regression models provides meaningful risk assessment with more straightforward interpretation of uncertainty of the prediction, potentially also on the individual level. This can aid estimation of incidence trends over time and comparisons of outcome of violent crimes for injury surveillance and in forensic medicine.