Performance Evaluation of SXI++ and Comparative Triage Machine Learning Models for Enhanced Medical Decision-Making

Abstract

Effective triage is vital to prioritize patient care and optimize resource utilization. This study compares traditional machine learning, represented by an XGBoost classifier, with SXI++, a proprietary deep neural network and ensemble learning framework, for predicting Emergency Severity Index (ESI) levels. An emergency department dataset of 560,486 records and 972 features was reduced to 180 through correlation filtering, Least Absolute Shrinkage and Selection Operator (LASSO) regression, and principal component analysis. A 100,000-record sample was imputed, normalized, and split into training (70%), validation (10%), and test (20%) sets. On the large dataset test set (10,000 records), SXI++ achieved 99.8% accuracy, 99.8% precision, 99.8% sensitivity, 99.8% specificity, and an AUC of 0.999, outperforming XGBoost (82.9% accuracy, 85.5% precision, 73.7% sensitivity, 91.1% specificity, AUC 0.779) and SXI (90.76% accuracy, 80.97% precision, 90.7% sensitivity, 92.0% specificity, AUC 0.967). Validation on a smaller dataset of 1,267 adult ED cases confirmed adaptability, with SXI++ achieving 100% accuracy, precision, sensitivity, specificity, and AUC. These findings highlight SXI++’s potential for accurate and scalable triage automation. However, near-perfect metrics warrant further validation on diverse, real-world datasets to assess generalizability, computational efficiency, and overfitting risk before clinical deployment.