Application of the Bidirectional Encoder Representations from Transformers Model for Predicting the Abbreviated...

Area of Science:

• Medical informatics
• Artificial intelligence in healthcare
• Trauma research

Background:

• Deaths from physical trauma represent a significant societal burden.
• The Abbreviated Injury Scale (AIS) is crucial for injury severity assessment but requires expert coding from medical records, increasing time and cost.
• Current AIS coding faces challenges due to large patient volumes and limited access to detailed medical records.

Purpose of the Study:

• To develop an advanced deep learning model for predicting AIS codes using easily accessible patient diagnostic information.
• To enhance the accuracy and efficiency of trauma assessment in clinical practice.

Main Methods:

• Utilized a dataset of 26,810 trauma patients from Chongqing Daping Hospital (Oct 2013 - Jun 2024).
• Employed diagnostic information, injury descriptions, cause, region, types, and illness history as input features.
• Developed a prediction model using a Bidirectional Encoder Representations from Transformers (BERT) pretraining method, validated via 5-fold cross-validation and external data from Chongqing Emergency Center.

Main Results:

• The BERT model achieved an accuracy of 0.8971 on independent test datasets, surpassing previous studies by 10% points.
• Achieved an Area Under the Curve (AUC) of 0.9970 and an F1-score of 0.8434.
• External validation showed strong generalization with accuracy, AUC, and F1-scores of 0.7131, 0.8586, and 0.6801, respectively.

Conclusions:

• The proposed BERT model demonstrates superior accuracy in predicting AIS codes based on diagnostic information compared to existing methods.
• The model exhibits high generalization ability, proving effective on external datasets for trauma assessment.