From: Artificial intelligence and machine learning for hemorrhagic trauma care
Study topic | Number of papers | Range of year | Range of patients/data used | Best performing models (AUROC) | Mean performance ± SD |
---|---|---|---|---|---|
Different outcomes in trauma | 45 | 1995–2022 | 32–1,511,063 | Performance of max EDI after 24 h for mortality (0.98) [61] | 0.91 ± 0.06 (n = 36) |
Risk assessment | 18 | 2009–2021 | 73–2,007,485 | Performance of RF model trained using ISS, AIS chest, and cryoprecipitate given within first 24 h (0.97) [92] | 0.88 ± 0.07 (n = 13) |
Transfusion | 11 | 2015–2021 | 477–12,624 | Performance of RF model using age, gender, mechanism of injury, involvement in explosion, vital signs (0.98) [49] | 0.87 ± 0.09 (n = 8) |
Hemorrhage detection | 11 | 2007–2021 | 24–368,810 | Performance of Poisson Regression model using epidemiological data, GCS, SBP, DBP, HR, haemoglobin, amount of RBC packs, platelets and fresh frozen plasma transfused, transfusion timing, and coagulation tests results (0.92) [110] | 0.92 ± 0.07 (n = 6) |
Coagulopathy | 4 | 2014–2019 | 54–18,811 | Performance of BN model using HR, SBP, Temperature, Hemothorax, FAST result, GCS, Lactate, Base deficit, pH, mechanism of injury, pelvic fracture, long bone fracture (0.96) [57] | 0.89 ± 0.08 (n = 3) |