TriML- XAI - Predictive Model For Pre-Hospital Triage With Transparency

Abstract

This study develops a Machine Learning model to predict pre-hospital triage levels with enhanced transparency, using clinical data such as vital signs and patients' chief complaints. Accurate triage is essential to prioritize high-risk patients in emergencies. The study evaluates multiple models, including Logistic Regression, Random Forest, Gradient Boosting, and an FCDNN with various embeddings. Voting models were also tested, combining Logistic Regression, Random Forest, and Gradient Boosting. Initial evaluation using a REST API-based test framework showed that each model performed best at specific triage levels. The FCDNN with GloVe embeddings was selected for triage level 1, achieving 90% accuracy, 1.00 recall, and a 0.99 F1-score in identifying high-risk cases. For triage levels 2 and 3, the Random Forest model performed well, with a recall of 0.83, an F1-score of 0.65, and an accuracy of 90%. For triage levels 4 and 5, Logistic Regression was chosen, with a recall of 0.74, an F1-score of 0.81, and an accuracy of 60%. These results led to a novel ensemble approach that enhances the accuracy. The proposed ensemble approach operates through three specialized APIs corresponding to the FCDNN model, Random Forest model, and Logistic Regression model. Upon receiving a user request, the Ensemble API invokes these three APIs to generate individual triage level predictions. These predictions are then combined using a Hard Voting mechanism, where the triage level with the majority vote is selected as the final prediction. If the models produce conflicting results (each predicting a different triage level), the system applies weighted voting, assigning weights to models based on their F1-scores to determine the most appropriate outcome. Transparency is enhanced using SHAP, providing healthcare professionals with insights into feature contributions for each prediction. SHAP values reveal the factors influencing a model’s decision, helping doctors and responders understand why a particular triage level was assigned. This added interpretability fosters trust, making the model a reliable tool in critical, time-sensitive medical settings.