Overview
This research investigates the accuracy of risk stratification for pulmonary embolism (PE) using models that integrate data from medical records and biomarkers extracted from Computed Tomography Pulmonary Angiography (CTPA) images. The study specifically addresses the common clinical challenge of missing blood test data in routine PE risk stratification. It benchmarks different approaches to combine medical records and cardiac biomarkers with detailed pulmonary vascular information, evaluating both the addition of vascular biomarkers to tabular models and the application of graph neural networks (GNNs) to the intrinsic graph representation of the vascular tree.
Research Context
Risk stratification for pulmonary embolism is a critical component of clinical decision-making. Current guidelines for PE stratification rely on data from patient medical records, parameters derived from CTPA, and blood test results. However, the absence of blood test results in routine clinical practice presents a challenge for comprehensive risk assessment. This work aims to determine if state-of-the-art models can accurately classify PE risk stratification when relying solely on medical records and CTPA-derived biomarkers, circumventing the need for potentially unavailable blood test data.
Approach
The study utilized a private dataset comprising 353 cases, characterized by uniquely complete data suitable for PE risk stratification. Researchers investigated various methodologies for combining different data types:
- Augmenting tabular models with vascular biomarkers.
- Applying Graph Neural Networks (GNNs) directly on the intrinsic graph representation of the vascular tree.
These approaches were benchmarked to assess their effectiveness in classifying PE risk stratification. The investigation also considered hypotheses regarding both the models and the data to explain observed performance.
Findings
The analysis yielded several key findings:
- Among global features, medical records and cardiac biomarkers were identified as the most significant predictors for PE risk stratification.
- The inclusion of vascular biomarkers did not further enhance the accuracy of stratification when added to models incorporating medical records and cardiac biomarkers.
- Graph Neural Networks (GNNs) applied to vascular graphs did not outperform strong tabular baselines that relied on global features.
- The investigation suggests that, contrary to some intuitions, vascular graphs might not contain discriminative information for PE risk stratification.
Why This Matters
These findings are relevant for the development of PE risk stratification models, particularly in scenarios where blood tests are unavailable. The research highlights the predictive power of medical records and cardiac biomarkers, while questioning the added value of complex vascular graph analysis for this specific clinical task. This could inform the design of more efficient and robust risk stratification tools.