Objective: To develop a grading prediction model of traumatic hemorrhage volume based on deep learning and assist in predicting traumatic hemorrhage volume.
Methods: A retrospective observational study was conducted based on the experimental data of pig gunshot wounds in the time-effect assessment database for experiments on war-traumatized animals constructed by the General Hospital of the Chinese People's Liberation Army. The hemorrhage volume data of the study population were extracted, and the animals were divided into 0-300 mL, 301-600 mL, and > 600 mL groups according to the hemorrhage volume. Using vital signs indexes as the predictive variables and hemorrhage volume grading as the outcome variable, trauma hemorrhage volume grading prediction models were developed based on four traditional machine learning and ten deep learning methods. Using laboratory test indexes as predictive variables and hemorrhage volume grading as outcome variables, trauma hemorrhage volume grading prediction models were developed based on the above fourteen methods. The effect of the two groups of models was evaluated by accuracy and area under the receiver operator characteristic curve (AUC), and the optimal models in the two groups were mixed to obtain hybrid model 1. Feature selection was conducted according to the genetic algorithm, and hybrid model 2 was constructed according to the best feature combination. Finally, hybrid model 2 was deployed in the animal experiment database system.
Results: Ninety-six traumatic animals in the database were enrolled, including 27 pigs in the 0-300 mL group, 40 in the 301-600 mL group, and 29 in the > 600 mL group. Among the fourteen models based on vital signs indexes, fully convolutional network (FCN) model was the best [accuracy: 60.0%, AUC and 95% confidence interval (95%CI) was 0.699 (0.671-0.727)]. Among the fourteen models based on laboratory test indexes, recurrent neural network (RNN) model was the best [accuracy: 68.9%, AUC (95%CI) was 0.845 (0.829-0.860)]. After mixing the FCN and RNN models, the hybrid model 1, namely RNN-FCN model was obtained, and the performance of the model was improved [accuracy: 74.2%, AUC (95%CI) was 0.847 (0.833-0.862)]. Feature selection was carried out by genetic algorithm, and the hybrid model 2, namely RNN-FCN∗ model, was constructed according to the selected feature combination, which further improved the model performance [accuracy: 80.5%, AUC (95%CI) was 0.880 (0.868-0.893)]. The hybrid model 2 contained ten indexes, including mean arterial pressure (MAP), hematocrit (HCT), platelet count (PLT), lactic acid, arterial partial pressure of carbon dioxide (PaCO2), Total CO2, blood sodium, anion gap (AG), fibrinogen (FIB), international normalized ratio (INR). Finally, the RNN-FCN∗ model was deployed in the database system, which realized automatic, continuous, efficient, intelligent, and grading prediction of hemorrhage volume in traumatic animals.
Conclusions: Based on deep learning, a grading prediction model of traumatic hemorrhage volume was developed and deployed in the information system to realize the intelligent grading prediction of traumatic animal hemorrhage volume.