Real-time AI Model for Prehospital Trauma Mortality Predcition

This project presents a real-time AI model for predicting emergency room (ER) mortality in trauma patients using only prehospital data.
Traditional triage tools rely on limited physiological data and subjective assessments, often leading to misclassification.
To address these limitations, we developed an interpretable AI model utilizing 21 prehospital variables, validated across multiple institutions and countries.

---

System requirements

Programming Language

• Python 3.8.19
• R 4.4.0 (for statistical analysis)

Python Libraries

• NumPy 1.23.5
• Pandas 2.0.3
• Matplotlib 3.7.5
• Scikit-learn 1.3.0
• BetaCal 1.1.0

R Packages

• Base R (v4.4.0), with standard statistical functions

Hardware Requirements

• Standard desktop or laptop computer
• Recommended: CPU with ≥ 4 cores, 16GB RAM
• Optional: NVIDIA GPU (e.g., RTX 3080, CUDA 11.3) for faster model training

---

Overview

We implemented a gradient boosting ensemble model (XGBoost + LightGBM + Random Forest) to predict trauma mortality before hospital arrival.
The model was designed for real-time application, achieving an inference speed of 1.06 seconds, making it feasible for use in emergency scenarios.

Compared to the conventional Shock Index (SI) triage tool, our AI model demonstrated superior predictive performance (AUROC: 0.9433) and was externally validated across multiple institutions and nations.

---

Study Setting & Dataset

• Development & Internal Validation: Trained on the Korean Trauma Data Bank (KTDB, 227,567 patients)
• External Validation:
  • South Korea: 4 regional trauma centers (8,867 patients)
  • Australia: 1 Level 1 trauma center (3,786 patients)
• Inclusion Criteria: Patients with available prehospital trauma care data
• Exclusion Criteria: Non-trauma-related deaths and patients transferred to another hospital

---

Trauma Data Prepreocessing

• This script anonymizes and processes trauma-related prehospital data into model-ready features.
• Main Steps:
  • Classification of time-of-injury as day, night, or missing
  • Age group binning (0–120 by 5-year intervals)
  • Selected categorical variables (e.g., gender, injury type, intentionality) are one-hot encoded.
    • Example categories include:
      • gender_male
      • intentionality_{accident, suicide, assault, others, unknown, missing}
      • injury_type_{blunt, penetrating, burn, others, unknown, missing}
  • Vital Signs Categorization
    • Six core vital signs are discretized into clinically meaningful bins:
      • SBP (systolic_bp): 0, 1–49, 50–75, 76–89, 90+
      • DBP (diastolic_bp): 0, 1–29, 30–45, 46–59, 60+
      • Pulse (pulse): 0, 1–29, 30–59, 60–100, 101–119, 120+
      • Respiration Rate (respiration): 0, 1–5, 6–9, 10–29, 30+
      • Body Temperature (body_temp, °C): 0, 0–24, 24–28, 28–32, 32–35, 35–37, 38+
      • Oxygen Saturation (spo2, %): 0, 1–80, 81–90, 91–95, 96+
    • Each bin is encoded as a binary variable (e.g., sbp_50_75 = 1).
    • Additionally, two missing value flags are generated for each vital sign:
      • {vital}_uncheckable: if the value is -1
      • {vital}_unchecked: if the value is -9

---

Features

Dynamic Weight Adjustment

• The ensemble model dynamically adjusts weight distribution, reducing model bias during training

Real-time Prediction & Interpretability

• Fast inference speed of 1.06 seconds for real-time emergency applications
• SHAP (Shapley Additive Explanations) analysis for improved model interpretability

Multi-Institutional & International Validation

• Model performance validated on independent datasets from South Korea and Australia

Superior to Conventional Triage Tools

• AI Model AUROC: 0.9433, Sensitivity: 87.4%, Specificity: 86.3%
• Shock Index (SI) AUROC: 0.7117

Key Predictors Identified by SHAP Analysis

• Oxygen Saturation
• AVPU Scale (Alert, Verbal, Pain, Unresponsive)
• Transport Mode
• Injury Mechanism

---

File Structure

• model.py : Implements the AI model using XGBoost, LightGBM, and Random Forest.
• data_loader.py : Handles data loading, filtering, and preprocessing for internal and external validation datasets.
• train.py : Trains the AI model using ensemble learning and optimizes class weights.
• Model loader.py : Script for loading the trained models.
• evaluate.py : Evaluates the trained model's performance and generates key metrics.
• README.md : Provides an overview of the project and usage instructions.
• Prehospital-AI model : Folder containing the trained model files. (.pkl)
• Sample_dataset.csv : Sample prehospital data for testing.
• Preprocess Trauma Data : Code to preprocess structured prehospital trauma data.

---

Installation Guide

1. Clone the repository

git clone https://github.com/ohnaeun111/Pre_hospital_mortality.git
cd Pre_hospital_mortality

2. Install required packages Use Python 3.8.19 and install the following packages:

pip install numpy==1.23.5 pandas==2.0.3 scikit-learn==1.3.0 \
            xgboost==1.6.2 lightgbm==3.3.5 betacal==1.1.0 matplotlib==3.7.5

Installation typically takes about 1–3 minutes on a normal desktop computer with a stable internet connection.

3. Run a test with the sample dataset Pre-trained ensemble model (Prehospital_AI_model.pkl) is provided. To run the model on your own data, replace Sample Data.csv with your own CSV file.
   Your file must follow the same format with 21 prehospital variables used in this study. You can run evaluation on each fold as follows:

python evaluate.py --input Sample_dataset.csv --model Prehospital-AI-model.pkl

The inference time for running the model on the provided sample dataset is approximately 1.06 seconds on a normal desktop computer.

Expected Output When running the evaluation on the provided sample dataset, the model produces the following results :

• Accuracy: 94.7%
• Sensitivity (Recall): 100%
• Specificity: 94.1%
• AUROC: 1.000
• Balanced Accuracy: 97.1% The results are consistent across runs, since the pre-trained models and sample dataset are fixed.

---

Web demo

• http://ai-wm.khu.ac.kr/Prehospital_Trauma_Mortality/