Pre-Imaging Clinical Factors Associated With Cardiac MR Image Quality Using Large Language Model-Enabled Data...

Area of Science:

• Medical imaging
• Artificial Intelligence
• Cardiology

Background:

• Poor cardiac magnetic resonance (MR) image quality necessitates repeat examinations, impacting patient care and clinical decisions.
• Improving cardiac MR image quality is crucial for accurate diagnosis and efficient healthcare delivery.

Purpose of the Study:

• To determine if pre-imaging clinical information, extracted by a large language model (LLM), is independently associated with cardiac MR image quality.
• To investigate the predictive value of patient clinical data for cardiac MR image quality outcomes.

Main Methods:

• Retrospective analysis of 1006 adult cardiac MR examinations across 1.5T and 3T scanners.
• Utilized a HIPAA-compliant LLM to extract clinical information and assign image quality labels from radiology reports.
• Employed multivariable logistic regression and chi-square tests to assess associations between clinical variables and image quality.

Main Results:

• LLM-derived image quality labels demonstrated substantial agreement with expert assessments (κ=0.689).
• Cognitive/communication impairment (OR 1.81) and respiratory issues (OR 1.57) were significantly associated with poor cardiac MR image quality.
• These associations remained significant after adjusting for repeat imaging (p<0.001 for cognitive impairment, p=0.027 for respiratory compromise).

Conclusions:

• Cognitive/communication impairment and respiratory compromise are independent predictors of poor cardiac MR image quality.
• Pre-imaging clinical data, extracted via LLM, can identify patients at higher risk for suboptimal cardiac MR imaging.
• This finding may inform strategies to optimize cardiac MR imaging protocols and reduce image quality-related failures.