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

Before the innovations presented in this paper, cardiac MR imaging faced significant challenges. The primary issue was the variability in image quality, which often led to repeat scans and inefficient use of imaging resources. Traditional methods for improving image quality relied heavily on reactive measures, addressing problems after they occurred rather than preventing them. The state of the art involved manual assessments and interventions based on radiologists' expertise, which were time-consuming and prone to subjective bias. This situation was unsatisfying as it limited the potential for streamlined workflows and consistent diagnostic outcomes.

The specific problem motivating this research was the inconsistency in cardiac MR image quality, which was not adequately addressed by existing pre-imaging assessments. Factors such as patient cognitive and respiratory conditions were known to impact image quality but were not systematically evaluated before imaging. This failure mode led to unnecessary repeat scans, increased costs, and delayed diagnoses, highlighting a critical gap in the imaging process. The reliance on post-imaging corrections and the lack of predictive tools for assessing potential image quality issues before scanning were major pain points in the current workflows.

The core insight of this research was the potential to leverage large language models (LLMs) to identify clinical factors from electronic health records (EHRs) that could predict cardiac MR image quality. Imagine if we could automatically extract relevant patient data before imaging, allowing for tailored interventions that preemptively address issues affecting image quality. This insight reframed the problem from one of reactive correction to proactive prevention. By recognizing the untapped potential of unstructured data in EHRs, the authors saw a way to integrate advanced AI techniques into clinical workflows, paving the way for significant improvements in imaging outcomes.

The system architecture centers around the integration of large language models to extract clinical factors from EHRs, categorizing cardiac MR image quality, and correlating these factors with image outcomes. The process begins with LLM-enabled data extraction, transforming unstructured clinical data into actionable insights. This is followed by image quality categorization, where images are labeled as 'Good' or 'Poor', validated through radiologist agreement. The system's design reflects a seamless flow from data extraction to quality assessment, ultimately leading to actionable interventions that can be implemented before imaging. The architecture is built to support a continuous feedback loop, where imaging outcomes inform future data extraction and intervention strategies.

Clinical Factor Extraction is a key component of this research, focusing on identifying relevant pre-imaging conditions from EHRs using LLMs. This process involves parsing unstructured data to extract variables like cognitive and respiratory conditions, known to impact imaging quality. The extraction process is designed to be comprehensive, capturing a wide range of clinical factors that could influence imaging outcomes. The use of LLMs marks a significant departure from traditional extraction methods, enabling a more nuanced understanding of patient conditions and their potential effects on imaging accuracy.

LLM-Enabled Data Extraction is a foundational method in this study, transforming unstructured EHR data into structured insights. The approach utilizes the advanced capabilities of large language models to parse complex medical records, identifying key clinical factors predictive of image quality. This method represents a technological breakthrough, allowing for the integration of rich, detailed clinical data into imaging workflows, significantly enhancing the pre-imaging assessment process. By leveraging LLMs, the study demonstrates how AI can bring a new level of precision and efficiency to data extraction, setting a new standard for clinical data integration.

Image Quality Categorization is the process by which cardiac MR images are classified into 'Good' or 'Poor' categories based on radiology reports. This classification is critical for correlating clinical factors with imaging outcomes. The research employs a systematic approach to ensure that categorization is consistent and reliable, validated by substantial radiologist agreement (κ = 0.689). This step is pivotal in the workflow, ensuring that the extracted clinical factors can be accurately assessed for their impact on imaging quality, reinforcing the study's findings with robust empirical backing.

Radiologist Agreement serves as a critical validation measure in this study, ensuring that the categorizations made by the system align with expert human assessments. The kappa statistic (κ = 0.689) indicates substantial agreement, underscoring the reliability of the image quality categorizations. This agreement is essential for building confidence in the system's ability to accurately classify image quality based on extracted clinical factors, providing a strong foundation for the study's conclusions and recommendations.

The training and data preparation involved in this study focus on optimizing the large language model for effective data extraction and image quality assessment. The study leverages a diverse dataset of 1006 adults undergoing cardiac MR exams, ensuring a comprehensive representation of patient demographics and scanner technologies. This diversity is crucial for training the model to accurately identify clinical factors across varied conditions, reinforcing the robustness and applicability of the study's findings.

The study's key results highlight the significant impact of cognitive and respiratory conditions on cardiac MR image quality. Cognitive impairments were associated with an odds ratio of 1.81 for poor image quality, while respiratory issues had an odds ratio of 1.57. These findings underscore the importance of considering these factors in pre-imaging assessments, providing a measurable basis for targeted interventions. The results confirm that pre-existing conditions play a critical role in imaging outcomes, offering actionable insights for clinical practice.

Ablation studies were conducted to evaluate the importance of different components in the system. The sensitivity analysis confirmed that key factors like cognitive and respiratory impairments remained significant under varied conditions, highlighting their robust impact on imaging quality. These studies provide a deeper understanding of which elements are most critical for accurate assessments, guiding future improvements and refinements in the system's design and application.

The innovations presented in this study have the potential to transform how clinical assessments are integrated into cardiac MR imaging workflows. By enabling pre-imaging interventions based on extracted clinical factors, the research promotes more efficient use of imaging resources and improved patient outcomes. The integration of predictive analytics into imaging processes marks a significant shift towards proactive healthcare, reducing the need for repeat scans and enhancing diagnostic reliability.

Despite its advancements, the study has limitations that warrant further exploration. The reliance on existing EHR data means that the model's accuracy is contingent on the quality and completeness of these records. Additionally, the study's focus on cognitive and respiratory factors may overlook other potential influences on image quality. Future research could explore a broader range of conditions and refine the model's ability to handle incomplete or inconsistent data.

For those in the healthcare technology industry, the implications of this study are profound. By integrating LLMs into imaging workflows, companies can develop pre-imaging assessment tools that predict and prevent poor imaging outcomes. This innovation has the potential to enhance imaging solutions, improve patient care, and optimize resource utilization, presenting a valuable opportunity for forward-thinking companies to lead in the evolving landscape of medical imaging technology.