Artificial intelligence (AI) and machine learning (ML) are increasingly transforming healthcare operations, particularly in emergency departments (EDs) where rapid, high-stakes decision-making is critical. Building on traditional supervised models, this study introduces a comparative framework evaluating large language model (LLM) + ML architectures against an extended hybrid system that integrates agentic AI (superagent) capabilities. Using structured medical records and unstructured triage notes, the agentic model coordinates data retrieval, contextual interpretation, and probabilistic inference. The agentic framework demonstrated the strongest overall performance, with the reasoning-enabled agent achieving the highest AUC (0.81) and the no-reasoning agent achieving the highest F1-score for admitted cases (0.55). In comparison, the baseline LLM reached an AUC of 0.535 and an F1-score of 0.33, while the XGBoost model achieved an AUC of 0.52 and an F1-score of 0.17. These results highlight the potential of agentic workflows for more balanced and adaptable ED admission prediction.
Read more about the project in this article about the Autumn 2025 Capstone Showcase.
Watch the team present this project at 46:24 in the session recording here.
Keywords: generative AI, agentic AI, large language models, machine learning, emergency department, hospital admission prediction, healthcare analytics, clinical decision support
