OBJECTIVE: The aim of this study is to determine the efficiency of artificial intelligence in detecting craniocervical junction injuries by using an artificial neural network (ANN) that may be applicable in future studies of different traumatic injuries. MATERIALS AND METHODS:Major head trauma patients with Glasgow Coma Scale <or=8 of all age groups who presented to the Emergency Department were included in the study. All patients underwent brain computerized tomography (CT), craniocervical junction CT, and cervical plain radiography. A feedforward with back propagation ANN and a stepwise forward logistic regression were performed to test the performances of all models. RESULTS: A total of 127 patients fulfilling inclusion criteria were included in the study. The mean age of the study patients was 31+/-17.7, 77.2% (n=98) of them were male, 13.4% of the patients (n=17) had craniocervical junction pathologies. About 64.7% (n=11) of these pathologies were detected only by CT; 23.5% (n=4) of them by both craniocervical CT and cervical plain radiography; and 11.8% (n=2) of them only by cervical plain radiography. A logistic regression model had a sensitivity of 11.8% and specificity of 99.1%. Positive predictive value was 66.7% and negative predictive value was 87.9%. Area under the curve for logistic regression model was 0.794 (P=0.000). ANN had a sensitivity of 82.4% and specificity of 100%. Positive predictive value was 100% and negative predictive value was 97.3%. Area under the curve for ANN model was 0.912 (P=0.000). CONCLUSION: ANN as an artificial intelligence application seems appropriate for detecting and excluding craniocervical junction injury but it should not replace craniocervical junction CT. However, these findings should lead us to test the applicability of ANN on hard-to-diagnose trauma patients or in constructing clinical decision rules.
RCT Entities:
OBJECTIVE: The aim of this study is to determine the efficiency of artificial intelligence in detecting craniocervical junction injuries by using an artificial neural network (ANN) that may be applicable in future studies of different traumatic injuries. MATERIALS AND METHODS: Major head traumapatients with Glasgow Coma Scale <or=8 of all age groups who presented to the Emergency Department were included in the study. All patients underwent brain computerized tomography (CT), craniocervical junction CT, and cervical plain radiography. A feedforward with back propagation ANN and a stepwise forward logistic regression were performed to test the performances of all models. RESULTS: A total of 127 patients fulfilling inclusion criteria were included in the study. The mean age of the study patients was 31+/-17.7, 77.2% (n=98) of them were male, 13.4% of the patients (n=17) had craniocervical junction pathologies. About 64.7% (n=11) of these pathologies were detected only by CT; 23.5% (n=4) of them by both craniocervical CT and cervical plain radiography; and 11.8% (n=2) of them only by cervical plain radiography. A logistic regression model had a sensitivity of 11.8% and specificity of 99.1%. Positive predictive value was 66.7% and negative predictive value was 87.9%. Area under the curve for logistic regression model was 0.794 (P=0.000). ANN had a sensitivity of 82.4% and specificity of 100%. Positive predictive value was 100% and negative predictive value was 97.3%. Area under the curve for ANN model was 0.912 (P=0.000). CONCLUSION: ANN as an artificial intelligence application seems appropriate for detecting and excluding craniocervical junction injury but it should not replace craniocervical junction CT. However, these findings should lead us to test the applicability of ANN on hard-to-diagnose traumapatients or in constructing clinical decision rules.
Authors: Greta Falavigna; Giorgio Costantino; Raffaello Furlan; James V Quinn; Andrea Ungar; Roberto Ippoliti Journal: Intern Emerg Med Date: 2018-10-23 Impact factor: 3.397