Pulmonary Embolism Risk Stratification from CTPA and Medical Records: Vascular Graphs Are Not All You Need
Nathan Painchaud, Tristan Habémont, Morgane des Ligneris, Allan Serva, Pierre Croisille, Laurent Bertoletti, Thomas Lampert, Johannes F. Lutzeyer, Odyssée Merveille
Read on arXiv →
Key claim
Vascular graphs do not improve PE risk stratification accuracy.
In plain English
Imagine you're trying to assess the risk of pulmonary embolism (PE) in patients. Traditionally, doctors rely on a mix of medical records, imaging tests, and blood tests to make these assessments. However, blood tests are often missing, which can lead to incomplete evaluations. This is a problem because without all the necessary information, doctors might misclassify a patient's risk, potentially leading to inadequate treatment or unnecessary anxiety. This situation is what's called incomplete data reliance.
To address this, the authors of the paper investigated whether advanced machine learning models could still accurately classify PE risk using just the available medical records and information extracted from imaging tests. They combined these data sources in various ways, including adding vascular biomarkers to traditional models and using graph neural networks (GNNs) to analyze the vascular structure represented as a graph.
What they found was quite surprising: while medical records and cardiac biomarkers were strong predictors of risk, the vascular biomarkers did not improve the risk stratification at all. Even more unexpectedly, the GNNs, which are typically powerful tools for analyzing complex data structures, failed to outperform simpler tabular models that used just the global features from the medical records. This suggests that, contrary to what one might expect, the vascular graphs may not provide any additional useful information for assessing PE risk. For anyone building systems in this area, it means that relying on complex models might not always yield better results, and simpler approaches could be more effective.
The paper explores a novel approach to risk stratification using medical records and imaging data, revealing unexpected limitations of vascular graphs.
The study uses a private dataset with complete data and benchmarks various methods, providing solid evidence for its claims.
Deep reliability assessment
The methodology supports a careful within-dataset comparison between tabular global-feature models and several GNN-based vascular graph fusion strategies for PE risk stratification on a uniquely annotated private cohort. The stronger claim that vascular graphs contain no discriminative information is more speculative, because it rests on one dataset, automated segmentation/graph extraction choices, and the absence of external validation.
Reproducibility
Code is open source at the GitHub URL mentioned in the abstract, but the dataset is private; public alternatives lack the same combination of thrombi annotations, vessel annotations, medical records, and PE risk labels.
Key figure
Figure 1 shows the full PE risk stratification pipeline: CTPA scans are segmented into ventricles, arteries, and thrombi; cardiac and vascular biomarkers plus vascular graphs are extracted; these are combined with structured medical records to classify patients into low, intermediate-low, or intermediate-high risk.