Pulmonary Embolism Risk Stratification from CTPA and Medical Records: Vascular Graphs Are Not All You Need

This study investigates whether incorporating detailed pulmonary vascular information—represented as patient-specific graphs—improves the accuracy of pulmonary embolism (PE) risk stratification. While clinical guidelines rely on medical records, blood tests, and cardiac biomarkers, blood tests are often missing in routine practice. The authors developed a pipeline to automatically extract vascular graphs and cardiac biomarkers from CTPA images and combined these with structured medical records. They benchmarked various models, including state-of-the-art tabular models (TabPFN, XGBoost) and several graph neural network (GNN) architectures, to determine if vascular topology adds predictive value.

The researchers found that medical records and cardiac biomarkers are the most significant predictors of PE risk. Surprisingly, incorporating vascular biomarkers or using GNNs to process the full vascular tree did not improve classification performance over strong tabular baselines. While GNNs successfully learned to regress local vascular features (such as thrombus volume), this local information did not translate into better patient-level risk stratification. The authors systematically investigated potential causes for this, such as model overfitting or poor label quality, but concluded that the vascular graphs themselves likely lack discriminative information for this specific clinical task.

This work challenges the clinical intuition that detailed mapping of thrombus burden within the pulmonary arterial tree is essential for risk stratification. By demonstrating that simpler, globally available clinical data is sufficient, the study suggests that complex, computationally expensive image-based graph analysis may not be necessary for this specific diagnostic goal. This finding helps streamline clinical workflows and directs future research toward more informative diagnostic markers.