Table 1 Studies utilizing artificial intelligence in trauma surgery
Study | Objective | Algorithm type(s) | Input variables | Prediction output(s) | Results (testing set) |
|---|---|---|---|---|---|
Injury | |||||
Abdel-Aty et al. [7] | Predict driver injury severity after crash has occurred | MLP Fuzzy adaptive resonance theory | Alcohol involvement Area type Demographics (age, gender) Lighting Peak period Point of impact Seatbelt Speed ratio Trafficway character Vehicle type Weather | Disabling injury/fatality Evident injury Possible injury No injury | 73.5% accuracy for the MLP 70.6% accuracy for fuzzy adaptive resonance theory |
Al Mamlook et al. [8] | Predict severity of traffic accidents and contributing factors to Traffic Accident Severity | AdaBoost LR Naive Bayes RF | Alcohol or drug involvement Car manufacturing year Clear weather Demographics (age, gender) Driver hazard action Lighting Seatbelt Traffic volume | Fatal or severe crash Minor, possible, or property damage crashes | 75.5% accuracy for RF 74.5% accuracy for AdaBoost and LR 73.1% accuracy for Naïve Bayes |
Amiri et al. [9] | Predict severity of crashes where driver is > = 65-years-old and hits a fixed object | ANN Hybrid Intelligent Genetic Algorithm | Average annual daily traffic Cause of collision Demographics (age, gender) Facility access Left shoulder Lighting Median type Number of lanes Number of vehicles Right shoulder Road surface condition Surface type Time | Fatal Severe injury Visible injury Complaint of pain Property damage only | 0–94.6% accuracy for ANN depending on level of damage predicted 0–78.6% accuracy for hybrid model depending on level of damage predicted Both models best predicted property damage only |
Assi et al. [10] | Predict crash injury severity using attributes that can be quickly identified on crash sites | Feed-forward NN SVM Fuzzy C-means clustering based NN Fuzzy C-means based SVM | Area type Day of the week Junction control Junction type Lighting Number of casualties Number of vehicles involved Road class Road surface condition Road type Speed limit Vehicle type Weather | Severe crash Non-severe crash | ~ 74% accuracy for SVM-FCM ~ 73% accuracy for SVM ~ 71% accuracy for FNN-FCM ~ 69% accuracy for FNN |
Assi et al. [11] | Predict severity of traffic crashes using attributes that can be quickly identified on crash sites | Deep neural network SVM | Age, vehicle Crash type Day Gender, driver Geometry of roadway Lighting Number of persons involved Number of vehicles involved Roadway median separation Speed limit Surface condition Surface type Traffic control Type of vehicle Weather condition | Severe crash Non-severe crash | 95% accuracy for DNN ~ 80% accuracy for SVM |
Bao et al. [12] | Predict short-term crash risk at weekly, daily, and hourly levels | Spatiotemporal Convolutional Long Short-Term Memory Network | Arterial percentage Commercial area Crash risk Daily vehicle kilometers traveled Freeway percentage Intersections Local road percentage Population Precipitation Pressure Residential area Road density Snowfall Taxi trips Temperature Wind speed | Crash No crash | 88.78–99.21% accuracy on weekly level depending on level of spatial resolution used 75.35–96.46% accuracy on daily level depending on level of spatial resolution used 71.02–93.72% accuracy on hourly level depending on level of spatial resolution used |
Delen et al. [13] | Predict motor vehicle crash severity and factors that increase risk of severity during crashes | MLP | Age, vehicle Alcohol or drug involvement Demographics (age, sex) Highway Impact location Lighting Role in accident (striking vs struck) Rollover Seatbelt Surface conditions Vehicle orientation in collision Vehicle type Weekend evenings | Fatality Incapacitating injury Minor non-incapacitating injury Possible injury No injury | 70.11–89.34% overall accuracy depending on which of the 5 outcome measures were being predicted |
Elamrani Abou Elassad et al. [14] | Design a real-time crash prediction model | SVM MLP | Brake Drift angle Lateral gravity Longitudinal gravity RPM Speed Spin angle Throttle Vertical gravity Weather season Yaw angle | Crash occurrence Crash non-occurrence | 93.34% average accuracy for MLP 92.00% average accuracy for SVM |
Iranitalab et al. [15] | Compare the ability of 4 algorithms to predict traffic crash severity | Multinomial logit Nearest neighbor classification SVM RF | Accident in traffic Alcohol involvement Animal in roadway Debris Demographics (Driver < 25, Driver 13–19, Female, Male) Double bottom trailer Farm equipment Glare Intersection involved Lighting Median type Non-highway work Number of lanes Obstruction in roadway Population group Public/private property Road characteristics Road classification Road surface condition Road surface type Rut, holes, bumps School bus Shoulders Total trucks/buses Traffic control device inoperative, missing, etc. Vehicles Vision obstruction Weather condition Work zone Worn, travel-polished surface | Disabling/fatal injury Visible injury Possible injury Property damage only | High variability by output level predicted; prediction most accurate for “property damage only” and decreased with increasing severity ~ 0–99% prediction accuracy for MNL ~ 5–80% accuracy for NNC ~ 1–95% accuracy for SVM ~ 5–90% accuracy for RF |
Mansoor et al. [16] | Predict crash severity based on easily available crash features | kNN DT AdaBoost SVM Feed-forward ANN | Area type Day of the week Intersection control Intersection type Lighting Number of vehicles involved Road class Road surface condition Road type Speed limit Vehicle type Weather condition | Severe crash Non-severe crash | 67.1% accuracy for kNN 69.2% accuracy for DTs 71.4% accuracy for AdaBoost 69.7% accuracy for ANN 68.8% accuracy for SVM 76.7% accuracy for a two-layer ensemble model |
Taamneh et al. [17] | Predict severity of road traffic injuries in real time | MLP | Accident Type Causality status Day Demographics (age, gender, nationality) Lighting Number of lanes Reason Road surface Seatbelt Speed limit Time Weather Year | Death Severe accident Moderate accidents Minor accidents | 65.1% accuracy |
Pre-hospital triage | |||||
DiRusso et al. [18] | Predict survival of trauma patients based on pre-hospital and emergency room admission data | Feed-forward ANN | Certification level of responder Demographics (age, race, sex) GCS Hct ICD-9-CM E-code Intubation status ISS Time to ED Vitals (sBP, HR, RR, temp) | Survival | 0.910–0.912 AUC for the ANN |
Kang et al. [19] | Predict need for critical care patients in emergency medical services | Deep learning | Chief complaint Demographics (Age, sex) Mental status Time from symptom onset to visit or EMS contact Trauma Vitals, initial | Need for critical care | 0.867 AUC for predicting need for critical care |
Kim et al. [20] | Triage patients by casualty likelihood for mass casualty incidents | LR RF NN | Age Consciousness score Vitals (sBP, HR, RR) | Death Survival | 0.71–0.88 AUC for LR depending on combination of input variables 0.89 AUC NN 0.87 AUC RF |
Liu et al. [21] | Predict injury severity in real time as defined by the need for life-saving intervention in the pre-hospital or emergency department settings | Hybrid system: Basic detection rules + MLP | Pulse pressure Shock index Vitals (SpO2, dBP, sBP, HR, RR, MAP) | Life-saving intervention | 69.5–89.8% accuracy depending on defined true positive rate |
Nederpelt et al. [22] | Design algorithm to support in-field triage decisions after gunshot wound | Information-aware Dirichlet deep neural network | Alcohol involvement BMI Comorbidity Demographics (age, ethnicity, race, sex) Drug screen GCS GSW anatomical site Time from dispatch to evaluation Transfer status Vitals (sBP, HR, SpO2, RR, temp) | Early massive transfusion Need for major hemorrhage control procedures Shock | 0.88–0.89 AUROC for shock depending on input variables used 0.86 AUROC for massive transfusion regardless of input variables used 0.80–0.82 AUROC for hemorrhage control depending on input variables used |
Emergency department | |||||
Dennis et al. [23] | Predict trauma admission volume, penetrating trauma admissions, and mean ISS | Feed-forward ANN | Center Daily high temperature Daily low temperature Day of week Day of year Precipitation Snow | Mean ISS score Number of penetrating traumas Number of traumas | R = 0.8732 correlation coefficient for all variables |
Menke et al. [24] | Predict emergency department patient volumes on a daily level | ANN | Air quality Days of the week Special events Weather | ED volume | 95% accuracy of volume within 20 visits of the true volume |
Rauch et al. [25] | Predict hourly emergency department volume based on traffic data | Seasonal autoregressive cross-validated models | Historical traffic data (direction and number of vehicles) | ED occupancy | 3.21–4.23 patient root-mean square error depending on prediction horizon 2.32–3.25 patient mean average error depending on prediction horizon |
Stonko et al. [26] | Predict the volume and acuity of trauma volume in an emergency department | ANN | Daily high Date Day of week ED discharge disposition Injury type ISS Mechanism of injury Precipitation | Mean ISS score per day Number of OR cases per day Number of penetrating traumas per day Number of traumas per day | 0.8940 correlation coefficient for all outcome variables |
Workup | |||||
Batchinsky et al. [27] | Predict need for life-saving interventions based on EKG data | ANN | EKG findings | Life-saving intervention needed No life-saving intervention needed | ~ 0.86 AUC |
Bektas et al. [28] | Detect craniocervical junction injuries based on CT and patient/injury characteristics | LR ANN | Alcohol intoxication Demographics (age, sex) Falls GCS Motor vehicle accident Motorbike accident Pathology on head CT Pedestrian struck RTS Vitals (HR, MAP, RR) | Presence of craniocervical junction injury | 0.794 AUC for LR 0.912 AUC for ANN |
Bertsimas et al. [29] | Predict cervical spine injuries in patients < 3 to avoid imaging | OCT GB trees LR | Demographics (age, gender) GCS Mechanism of injury | Presence of cervical spine injury Absence of cervical spine injury | 90.43% AUC for OCT 96.69% AUC for GB trees 94.06% AUC for LR |
Cheng et al. [30] | Determine accuracy of AI-assisted free fluid detection in Morison’s pouch during FAST examination | Deep learning | Abdominal US | Negative/non-qualified view Negative/qualified view Positive/non-qualified view Positive/qualified view | 96.7% accuracy for detecting ascites 94.1% accuracy in classifying qualified and non-qualified images |
Dreizin et al. [31] | Use deep learning to segment the volume of pelvic hematomas | Recurrent Saliency Transformation Network | Chest, pelvis, abdominal CT scans | Pelvic hematoma volume | 0.81 AUC for predicted volumes (as compared to 0.80 as manually done by radiologists) |
Liu et al. [32] | Predict the need for life-saving interventions in trauma patients | LR MLP | GCS HR complexity HR variability Pulse pressure Shock index Vitals (dBP, sBP, HR, SpO2) | Life-saving intervention | 0.73–0.94 AUC for LR depending on variables included 0.99 AUC for MLP |
Paydar et al. [33] | Predict injury severity from clinical and paraclinical data on blunt trauma injury | SVM KNN Bagging AdaBoost NN | 67 features including vital signs, injury organs, and ISS (exact features not listed) | Critically ill Not critically ill | ~ 99.24% accuracy for SVM 63.84% accuracy for KNN 99.67% accuracy for Bagging ~ 75.81% accuracy for AdaBoost 51.60% accuracy for NN |
Intervention and outcome | |||||
Abujaber et al. [34] | Predict risk of prolonged mechanical ventilation with TBI | LR ANN SVM RF C.5 DT | AIS per body region Blood transfusions CT scan findings Date/time of injury Demographics (age, gender, race) GCS In-hospital complications Intubation status ISS Comorbidities Mechanism of injury Outcome and date of disposition Performed procedures Time of ED admission Vitals, on ED arrival | Prolonged mechanical ventilation (> 7, > 10, or > 14 days) No prolonged mechanical ventilation | 73–75% accuracy for LR 69–77% accuracy for ANN 74–79% accuracy for SVM 71–75% for RF 66–71% for C.5 DT |
Ahmed et al. [35] | Predict mortality in patients admitted to trauma surgery ICU | Deep-FLAIM Gaussian Naïve Bayes DT KNN Linear Discriminant Analysis | Acute Physiology Score III Angus Criteria of Sepsis Laboratories (albumin, anion gap, BUN, creatinine, glucose, INR, lactate, platelets, PT, PTT, serum electrolytes) Logistic Organ Dysfunction System Oxford Acute Severity of Illness Score qSOFA SAPS SAPS II Sepsis diagnosis using Martin Sepsis et al SIRS SOFA | Survival | 92.25% accuracy for Deep-FLAIM 80.07% accuracy for GNB 89.59% accuracy for DT 84.94% accuracy for KNN 81.84% accuracy for LDA |
Becalick et al. [36] | Compare ability of ANN to predict outcome after injury with UK TRISS | ANN | AIS for each body region Demographics (age, gender) GCS Injury type Vitals (sBP, HR, SpO2, RR) | Survival | 89.6% accuracy for both ANN and UK TRISS; however, accuracy higher for ANN for predicting survival while UK TRISS better predicts death |
Christie et al. [37] | Predict dynamic/up-to-date risk of complications and identify patient-specific modifiable factors to adjust patient trajectory after severe injury | Ensemble machine learning algorithm | Alcohol or drug involvement APACHE-II Coagulation marker Demographics Denver Post-Injury Multiple Organ Failure Score Fluid, colloid, blood, and medication administration GCS Inflammation markers Injury characteristics Input/output data ISS Past medical history Ventilator parameters Vitals | Acute Respiratory Distress Syndrome Blood transfusion Coagulopathy/coagulopathic trajectory Length of Mortality Organ failure Venous thromboembolic events | 0.76–0.98 AUC for predicting death depending on time since admission 0.87–0.96 AUC for multi-organ failure 0.82–0.86 AUC for venous thromboembolic events at 96-120 h 0.84–0.88 AUC for transfusion 0.71–0.83 AUC for acute respiratory distress syndrome 0.45–0.74 AUC for coagulopathic trajectory |
Demsar et al. [38] | Predict patient outcome after initial damage control surgery | Classification trees Naïve Bayes classifier | Bicarbonate excess in ICU Catecholamine administration Estimated blood loss Physician impression of coagulopathy during operation PT in ICU Type of closing Worst arterial carbon dioxide tension Worst mean blood pressure Worst partial active thromboplastin time Worst pH Worst pH value at ICU Worst sBP | Survival | 82.4% accuracy for classification trees 79.4–80.9% accuracy for Naïve Bayes |
DiRusso et al. [39] | Compare ability of ANN and LR to predict pediatric trauma death | ANN LR | Demographics (age, sex) GCS Intubation status ISS NISS Pediatric Trauma Score RTS Vitals (sBP, HR, RR) | Survival | 0.964 AUROC for LR 0.961–0.966 AUROC for ANN |
El Hechi et al. [40] | Predict 30-day outcomes in patients undergoing emergency operations | OCT | Comorbidities Demographics Laboratory values Wound characteristics | 30-day morbidity 30-day mortality Occurrence of 18 complications | 0.93 c-statistic for predicting mortality 0.83 c-statistic for predicting morbidity |
Gorczyca et al. [41] | Compare an algorithm they developed against established risk prediction models (BLISS, HARM, and TMPM) | Stacked generalization of 5 different ML algorithms (LR with Elastic Net Penalty, RF, GB Machine, NNs) | Comorbidities Demographics (age, gender) GCS ICD-9 codes Injury mechanism Injury type Intent of trauma | Risk of death | 96.8% accuracy when only ICD-9 codes used as input 97.6% accuracy when all inputs utilized Algorithm equaled or improved as compared to established risk prediction models |
Hale et al. [42] | Predict clinically relevant TBI in pediatric patients | ANN | Demographics (age, sex) GCS Injury mechanism Loss of consciousness Radiologist-interpreted CT scan with 17 variables identified Severity of injury mechanism | Clinically relevant TBI defined by needing neurosurgical procedure, intubation > 24 h, hospitalization > 48 h, or death | 0.9907 AUC for detecting CRTBI |
Ji et al. [43] | Predict final outcome and ICU length of stay for trauma patients | Classification and regression tree C4.5 AdaBoost SVM NN LR | AIS by body part Blunt vs penetrating injury Comorbidities Complications Demographics (age, gender) GCS Intubation ISS Method of injury Provided fluids Role in accident Safety measures used during injury Vitals (BP, HR, RR) | Discharge location ICU length of stay Survival | 69.4–72.9% accuracy for LR depending on input variables 70–73% accuracy for AdaBoost depending on input variables 68–75.2% accuracy for C4.5 depending on input variables 75.6–77.6% accuracy for CART depending on input variables 73–79% accuracy for SVM depending on input variables 67.2–79.04% accuracy for NN depending on input variables |
Matsuo et al. [44] | Predict morbidity and mortality after TBI using parameters that are quickly and easily available in emergency care | Ridge regression Least absolute shrinkage and selection operator RF GB Extra trees DT Gaussian naïve Bayes Multinomial naïve Bayes SVM | Abnormal pupillary response Age sBP CT findings GCS Laboratories (CRP, fibrin/fibrinogen degradation products, glucose) Major extracranial injury | Death Poor outcome based on Glasgow Outcome Score | 71.7% accuracy for Gaussian NB for morbidity 90.2% accuracy for GB for morbidity 91.7% accuracy for RF for morbidity 78.2% accuracy for SVM for morbidity 95.5% accuracy for RF for mortality 88.6% accuracy for ridge regression for mortality 88.5% accuracy for SVM for mortality Not all algorithms had test results shown in the paper |
Maurer et al. [45] | Design and validate a smartphone-based risk calculator for trauma patients | OCT | AIS by body region Comorbidities Demographics (age, ethnicity, sex, race) GCS Mechanism of injury Vitals (sBP, HR, SpO2, RR, temp) | Acute kidney injury Acute respiratory distress syndrome Cardiac arrest requiring CPR Deep surgical site infection Deep vein thrombosis In-hospital morality Organ space surgical site injury Overall morbidity Pulmonary embolism Severe sepsis Unplanned intubation | 0.941 c-statistic for predicting mortality in penetrating injury 0.884 c-statistic for predicting mortality in blunt injury 0.777 c-statistic for predicting morbidity in penetrating injury 0.753 c-statistic for predicting morbidity in blunt injury 0.689–0.835 c-statistics for predicting individual complications |
Nourelahi et al. [46] | Predict “favorable” or “unfavorable” outcome after 6 months in severe TBI | LR RF SVM | Demographics (age, sex) GCS motor response Laboratories (glucose, PT-INR) Pupil reactivity Rotterdam index | GOSE = < 4 (“Unfavorable”) GOSE > 4 (“Favorable”) | 78% accuracy for all three model types |
Pang et al. [47] | Predict outcomes of severe TBI patients | LR NN DT Bayesian network Discriminant analysis | Coagulopathy Demographics (age, ethnicity, gender) Mechanism of injury Pre- and post-resuscitation GCS Pre- and post-resuscitation pupillary anomaly Traumatic subarachnoid hemorrhage Vitals (hypotension, hypoxia) | Glasgow Outcome Scale | 73.1% overall accuracy for DTs 70.51% overall accuracy for LR 66.39% overall accuracy for discriminant analysis 65.67% overall accuracy for Bayesian network 63.38% overall accuracy for NN |
Rashidi et al. [48] | Determine if a burn-trained algorithm could be generalized to a non-burned trauma surgery population to predict acute kidney injury | LR KNN RF SVM MLP | Central venous pressure Demographics MAP Laboratories (creatinine, NGAL, NT-proBNP) Urine output | Acute kidney injury | ~ 70–75% accuracy for all algorithms when all input variables included Variation of all algorithms ~ 20–90% depending on which inputs included |
Rau et al. [49] | Predict survival probability of trauma patients by the addition of a large number of input variables | LR SVM NN | AIS in different body regions Comorbidities Demographics (age, sex) GCS ISS Laboratory results (WBC, RBC, Hgb, Hct, platelets, neutrophils, INR, glucose, Na + , K + , BUN, Creatinine, aspartate, AST, ALT) RTS TRISS Vitals (dBP, sBP, HR, RR, temp) | Survival | 97.9% accuracy for LR 98.0% accuracy for SVM 98.3% accuracy for NN |
Schetinin et al. [50] | Predict trauma severity in trauma surgery patients | Bayesian averaging over DTs | Demographics (age, gender) GCS Injury severity (head, face, neck, thorax, abdomen, spine, upper extremity, lower extremity, and external) Injury type Vitals (BP, RR) | Death | 87.5–98.7% accuracy depending on the number of injuries present |
Shahi et al. [51] | Predict outcomes in pediatric patients with blunt solid organ injury | Deep Learning | Blood transfusion Clinical events (e.g., intubation, CPR) CT grade of injury Demographics (age, gender) ED TEG values FAST exam findings Fluid administered GCS Laboratory values (Hgb, INR, base deficit, lactate) Multiple solid organ injuries Presence of head injury SIPA scores Weight Vitals (BP, HR) | Mortality Failure of non-operative management Massive transfusion Successful non-operative management without intervention | 90.0–90.5% accuracy for massive transfusion depending on prediction horizon 82.4–83.8% accuracy for failure of non-operative management depending on prediction horizon 91.9% accuracy for mortality across prediction horizons 86.9–90.3% accuracy for successful non-operative management without intervention Note: only validation data shown |
Staziaki et al. [52] | Predict extended length of stay and ICU admission in trauma of the torso | SVM ANN | AAST grading CT imaging findings Demographics (age, sex) GCS Laboratories (Hct, Hgb, lactate) RTS Vitals | Extended length of stay ICU admission | 77–82% accuracy for SVM for ICU admission depending on inputs used 77–83% accuracy for ANN for ICU admission depending on inputs used 58–73% accuracy for SVM for extended length of stay depending on inputs used 65–77% accuracy for ANN for length of stay depending on inputs used |
Tsiklidis et al. [53] | Predict trauma patient survival and identify patient warning signs | GB | Demographics (age, gender) GCS Vitals (sBP, HR, SpO2, RR, temp) | Deceased Survived | 0.924 AUC for predicting death |