JMIR Publications

ABSTRACT

Background:

Deaths related to physical trauma impose a heavy burden on society, and the Abbreviated Injury Scale (AIS) is an important tool for injury research. AIS covers injuries to various parts of the human body and scores them based on the severity of the injury. Due to the complexity of human body structure and the diversity of injury types, accurately identifying and classifying each type of injury based on its corresponding AIS code level is highly challenging. In practical applications, the complex coding rules of AIS require experts to classify injuries by reviewing patient medical records, which inevitably increases the difficulty and time cost of evaluation, and also puts higher demands on the workload of information collection and processing.

Objective:

we aimed to use advanced deep learning techniques to predict AIS codes based on easily accessible diagnostic information of patients, to improve the accuracy of trauma assessment.

Methods:

We used a trauma patient dataset (n=26810) collected by Chongqing Daping Hospital between October 2013 and June 2024. We mainly selected the patient's diagnostic information, place of injury, cause of injury, chief complaints, present illness history, injury locations, and injury types as the key feature inputs. We employed the robustly optimized BERT pre-training method to embed these features, thus constructing our BERT model. This model is designed to predict AIS codes and comprehensively evaluate their performance through 5-fold cross-validation. In addition, we compared the BERT model with previous investigation and current mainstream machine learning methods, to validate BERT model advantages in prediction tasks.

Results:

The BERT model proposed in this paper performs significantly better than the comparison model on independent test datasets, with an accuracy of 0.8971, which surpassed the previous study by 10 percentage points. In addition, the AUC value of the BERT model is 0.9970, and the F1 score is 0.8434. Among all the compared machine learning methods, the decision tree method achieved excellent performance in accuracy, AUC, and F1 metrics, with values of 0.8506, 0.9945, and 0.7049, respectively. These results indicate that our model has high generalization ability and prediction accuracy.

Conclusions:

The BERT model we proposed is mainly based on diagnostic information to predict AIS codes, and its prediction accuracy is superior to previous investigation and current mainstream machine learning methods. However, validation in external databases is needed.