Study: Medical Records Beat Complex Vascular Graphs for PE Risk
Graph neural networks on pulmonary vessels? Turns out simple tabular data works better.
A new study accepted at MICCAI 2026 challenges the assumption that detailed vascular graph representations improve AI-driven risk stratification for pulmonary embolism (PE). The team—Painchaud, Habémont, des Ligneris, and colleagues from multiple French hospitals and research labs—systematically compared tabular models using global features (medical records, cardiac biomarkers from CTPA) against graph neural networks (GNNs) applied to the vascular tree's intrinsic graph structure. Using a private dataset of 353 patients with uniquely complete data, they found that adding vascular biomarkers to tabular models did not boost accuracy, and GNNs on vascular graphs failed to outperform a strong tabular baseline.
The authors probe why: the lack of discriminative power may stem from the complexity of the vascular representation or limitations in the dataset size and label granularity. They note that blood test results, often missing in practice, were simulated, which might affect generalizability. The work serves as a cautionary tale for the medical AI community: sometimes simpler baselines with clinically established features—like cardiac measurements (e.g., right ventricle/left ventricle ratio) and patient history—are sufficient, and expensive graph-based pipelines may add little value for this specific task. The code is open-sourced on GitHub.
- Tabular models using medical records and cardiac biomarkers outperformed graph neural networks on vascular trees for PE risk stratification.
- Adding vascular biomarkers did not improve performance; dataset n=353 from French hospitals.
- Study accepted at MICCAI 2026; code available on GitHub for reproducibility.
Why It Matters
Saves medical AI teams from chasing costly graph architectures when simpler models already work.