Agentic AI for Severity Extraction in Clinical Notes: Enhancing Disease Diagnosis Beyond Rule-Based and Traditional ML Models

Abstract

The accurate interpretation of symptom severity in clinical information extraction is critical for timely and informed disease diagnosis and treatment planning. Traditional natural language processing (NLP) models often struggle with contextual understanding and domain-specific nuances in clinical text. Transformer-based pretrained models such as BioMedNER and BioClinicalBERT accurately extract entity mentions such as symptoms, disease names, treatments, and drug names efficiently. Despite growing advancements in clinical NLP, current models fail to capture contextual severity terms associated with the symptoms, which are crucial for accurate diagnosis. This paper proposes an Agentic AI framework designed to improve severity extraction by leveraging advanced large language models (LLMs) in a structured, multi-step reasoning process. The pipeline supports modular and interpretable architecture, which enables cross-model evaluation and fine-grained t esting, a nd a reusable workflow. <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$T$</tex> he <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$p$</tex> roposed <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$f$</tex> ramework <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$w$</tex> as <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$e$</tex> valuated on the MTSamples dataset, comparing the performance of Granite and Llama models against an established biomedical NER model. Experiments demonstrate that the agentic framework with Granite and Llama significantly outperforms existing models in accurately identifying and extracting disease severity cues from unstructured clinical narratives. The results highlight the potential of Agentic AI in clinical NLP, offering a more robust solution for real-world healthcare applications.

Ähnliche Arbeiten