Abstract
The timely stratification of trauma injury severity can enhance the quality of trauma care but it requires intense manual annotation from certified trauma coders. There is a need to establish an automated tool to identify the severity of trauma injuries across various body regions. We gather trauma registry data from a Level I Trauma Center at the University of Wisconsin-Madison (UW Health) between 2015 and 2019. Our study utilizes clinical documents and structured electronic health records (EHR) variables linked with the trauma registry data to create two machine learning models with different approaches to representing text. The first one fuses concept unique identifiers (CUIs) extracted from free text with structured EHR variables, while the second one integrates free text with structured EHR variables. Both models demonstrate impressive performance in categorizing leg injuries, achieving high accuracy with macro-F1 scores of around 0.8. Additionally, they show considerable accuracy, with macro-F1 scores exceeding 0.6, in assessing injuries in the areas of the chest and head. Temporal validation is conducted to ensure the models’ temporal generalizability. We show in our variable importance analysis that the most important features in the model have strong face validity in determining clinically relevant trauma injuries.
Introduction
In trauma centers, quantifying injury severity is critical for clinical decision-making and quality improvement. Currently, this is achieved using the Abbreviated Injury Scale (AIS), which requires certified trauma coders to manually review electronic health records (EHR) and clinical notes. This labor-intensive process is expensive and typically occurs post-discharge, limiting its utility for real-time clinical guidance. While previous automation attempts focused on binary outcomes or single body regions, this study addresses the complexity of predicting three levels of severity across four distinct body regions simultaneously.
Methodology
The researchers developed and compared two primary multi-modal architectures designed to integrate unstructured text with structured clinical data:
- CUIs + Structured EHR Model: This approach extracts Concept Unique Identifiers (CUIs)—standardized medical terms—from clinical notes, processing them via a 1D Convolutional Neural Network (CNN).
- Free Text + Structured EHR Model: This model employs ClinicalBERT, a large-scale transformer model, to process raw, unstructured clinical text directly.
Both architectures fuse text-derived features with 59 structured EHR variables (e.g., vitals and lab results) using a multitask neural network encoder. The resulting fused embeddings are then passed to separate prediction heads (Multi-Layer Perceptrons) for four body regions: Leg, Chest/Abdomen/Spine, Head/Face/Neck, and Arm/Extremities.
To ensure interpretability, the team utilized Integrated Gradients (IG) to identify the specific clinical terms and variables driving severity predictions.
Data and Experimental Setup
The study used data from the University of Wisconsin-Madison (UW Health) Level I Trauma Center, spanning 2015 to 2019 and encompassing 9,686 trauma admissions. The dataset was split temporally: 2015–2018 data (7,693 encounters) for training and tuning, and 2019 data (1,983 encounters) for temporal validation to assess the models' generalizability to new patient cohorts over time.
Results
The models demonstrated strong discriminative power, particularly in distinguishing "Serious" injuries from "Negative" (no injury) cases, with AUROC scores exceeding 0.90 across several categories. Performance varied by body region, as measured by the macro-F1 score:
- Leg Injuries: Highest performance, with scores approximately 0.82.
- Head/Face/Neck: Scores around 0.71.
- Chest/Abdomen/Spine: Scores around 0.66.
- Arm/Extremities: The most challenging region, likely due to data skewness and a lower frequency of "serious" cases.
Discussion
A key finding was the efficiency of the CUI-based model. Despite having only 3.1 million parameters—compared to ClinicalBERT's 108 million—the CUI model achieved comparable or even superior performance. This suggests that for specific clinical tasks, structured medical concept extraction can be more computationally efficient than full natural language processing.
Furthermore, structured EHR data proved most critical when clinical text was vague, particularly for extremity injuries. Interpretability analysis confirmed "face validity," showing the models prioritized clinically relevant terms like "pneumothorax" and "rib fractures" for serious chest injuries.
Conclusion and Future Work
While the study demonstrates the feasibility of automated multi-region trauma stratification, it is limited by its single-center scope. Future work will focus on external validation across different health systems and extending the model to calculate the aggregate Injury Severity Score (ISS). This research paves the way for real-time trauma decision support tools that do not require massive computational overhead.
Summary updated