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

The real contribution is negative: graph neural networks on vascular trees underperform when tabular medical data is excluded, meaning the imaging biomarker alone cannot carry the diagnostic load. This inverts the implicit assumption in much medical imaging ML that richer visual representations automatically yield better predictions.

This echoes a pattern visible in recent work on hierarchical classification (the marine species taxonomy paper from June) and federated learning under data imbalance (FedReLa, also June). Both showed that encoding domain structure and handling heterogeneous data distributions outweigh raw model sophistication. Here, the lesson is similar: multimodal fusion is not optional in clinical ML, and imaging-first architectures that treat tabular records as an afterthought will fail. The work also connects to the broader June wave on coupling neural methods with real-world constraints (formal syntax in code generation, compression budgets in model deployment), suggesting the field is maturing beyond single-modality optimization.

If this team or others publish ablations showing which specific tabular features (lab values, vitals, comorbidities) matter most for PE risk, and whether missing-data imputation strategy changes the GNN contribution, that will clarify whether the finding is about data fusion architecture or just incomplete feature engineering. Watch for deployment studies on actual hospital datasets within 12 months; if the multimodal model fails to outperform simpler baselines in production, the lab result may not generalize.