APSevLM: Acute Pancreatitis Severity Language Model

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Clinical Decision Support

Background:

• Severe acute pancreatitis (AP) has high mortality rates, necessitating early severity prediction.
• Current prediction methods have limitations in accuracy and timeliness.
• Integrating diverse data sources is key for improved AP severity assessment.

Purpose of the Study:

• To develop and validate APSevLM, a novel large language model (LLM) for early prediction of AP severity.
• To compare APSevLM performance against established scoring systems and machine learning models.
• To identify key clinical factors contributing to severe AP prediction.

Main Methods:

• Developed APSevLM by integrating clinical data, imaging reports, and expert knowledge.
• Evaluated APSevLM on a dataset of over 500 patients.
• Compared APSevLM against BISAP, MCTSI, conventional ML, and deep learning models.
• Utilized attention visualizations and feature importance analysis for model interpretability.

Main Results:

• APSevLM achieved an Area Under the Curve (AUC) of 0.857, outperforming existing methods.
• Hematological parameters and cardiac markers were identified as significant predictive factors.
• Attention visualizations revealed dynamic weighting of data modalities based on severity.

Conclusions:

• APSevLM demonstrates high accuracy in early severe AP prediction.
• The model offers valuable insights into predictive factors for clinical practice.
• APSevLM represents a significant advancement in AI-driven diagnostic tools for AP.