Causal insights from clinical information in radiology: Enhancing future multimodal AI development.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Radiology report generation

Background:

• Radiology reports are crucial for patient care and AI model training.
• Annotation shifts in reports can introduce bias.
• Clinical information and prior imaging influence reporting.

Purpose of the Study:

• Investigate causal mechanisms of annotation shifts in radiology reports.
• Analyze the influence of clinical information and prior imaging on report generation.
• Quantify the impact of clinical context on reporting biases.

Main Methods:

• Retrospective analysis of 172,380 chest X-ray reports from the MIMIC-IV CXR database.
• Conditional effects analysis for diseases including pneumonia, pleurisy, heart failure, rib fracture, and COPD.
• Propensity score matching, logistic regression, neural networks, and risk difference calculations.

Main Results:

• Clinical questions significantly influence the reporting of findings like cardiomegaly (15% increase with rib fracture context).
• Support devices are also affected across multiple diseases.
• The impact of clinical information varies by disease, with significant effects on pneumonia mention in some cases.

Conclusions:

• Annotation bias in radiology reports stems from clinical context and prior imaging access.
• Understanding these causal mechanisms is vital for mitigating bias in dataset curation.
• This knowledge improves the reliability and generalizability of AI models in multimodal medical imaging.