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

Technical Summary: Pulmonary Embolism Risk Stratification from CTPA and Medical Records

Problem Statement
Risk stratification for Pulmonary Embolism (PE) is essential for determining clinical management, such as the need for intensive care or thrombolytic therapy. Current guidelines rely on a combination of medical records, blood tests (troponin, NT-proBNP), and visual evaluation of Computed Tomography Pulmonary Angiography (CTPA). However, blood tests are frequently unavailable during initial diagnosis, creating a gap between guidelines and routine practice. While CTPA is the standard diagnostic tool, its use for risk stratification is currently limited to assessing right ventricular (RV) dysfunction rather than thrombus burden. Existing clinical scores for thrombus burden (e.g., Qanadli, Mastora) require manual localization of thrombi, which is too time-intensive for routine use. Although deep learning has automated segmentation and detection, standard vision models struggle with the sparse nature of thrombi and arteries for patient-level risk prediction. This work investigates whether state-of-the-art models can accurately stratify PE risk using only medical records and biomarkers extracted from CTPA, specifically exploring if rich pulmonary vascular graph representations offer advantages over simpler global features.

Methodology
The authors propose a pipeline to categorize PE risk into three classes (Low, Intermediate-low, Intermediate-high) using a private dataset of 353 patients with uniquely complete data. The methodology involves three main stages:

1. Data Preprocessing and Feature Extraction:

   • CTPA Processing: 3D CTPA images are processed to segment the heart ventricles (LV/RV) and the pulmonary arterial tree (including thrombi) using TotalSegmentator and a custom nnUNet pipeline.
   • Graph Construction: The segmented arterial tree is converted into a patient-specific graph where junctions/terminals are nodes and vessels are edges. Edge features include vessel radius, length, volume, depth, and thrombus-related metrics (volume, obstruction).
   • Biomarker Computation:
     • Cardiac Biomarkers: Derived from ventricle masks (RV/LV volumes, RV/LV ratio).
     • Vascular Biomarkers: Six patient-level metrics summarizing thrombus burden (e.g., total thrombus volume, Qanadli score, Mastora scores) computed from the graph.
     • Medical Records: Structured data including sPESI score factors, age, heart rate, oxygen saturation, blood pressure, and history of cancer or cardiopulmonary disease.

2. Model Architectures and Fusion Strategies:
   The study benchmarks two primary approaches for combining medical records (global features) with vascular information:

   • Tabular Models: Global features (medical records, cardiac biomarkers, and vascular biomarkers) are concatenated and fed into tree-based models (XGBoost) and foundation models (TabPFN).
   • Graph Neural Networks (GNNs): GNNs are applied directly to the vascular graphs. The study evaluates:
     • Monomodal GNNs: Using only the graph.
     • Multimodal Fusion: Integrating global features with the graph using four strategies:
       • Early Fusion: Concatenating global features to each node.
       • Late Fusion: Concatenating global features to the final graph embedding.
       • Virtual Node (Proposed): Initializing a virtual node connected to all graph nodes using global features.
       • Cross-Attention (Proposed): Using a Graph Transformer (GPS backbone) where global features are tokenized and fused via bidirectional cross-attention with node tokens.

3. Experimental Setup:

   • Dataset: 353 PE patients from CHU Saint-Étienne (2014–2020). Risk labels are derived from ESC guidelines using blood test results (not provided to the model).
   • Evaluation: 10-fold cross-validation (80/10/10 split). Performance is measured using F1-score.
   • Hyperparameter Tuning: Exhaustive search using Optuna for GNN backbones (GCN, GATv2, GIN, GPS) and fusion configurations.

Key Contributions

1. Pipeline Development: Introduction of an automated pipeline for PE risk stratification that integrates structured medical records with CTPA-derived vascular graphs.
2. Comprehensive Benchmarking: A systematic comparison of tabular models against GNNs, testing two specific intermediate fusion methods (Virtual Node initialization and Cross-Attention) to combine global features with vascular graphs.
3. Empirical Investigation of Graph Utility: An analysis of why vascular graphs may fail to improve risk stratification, ruling out overfitting and label bias as primary causes.

Results

• Global Features Dominance: Medical records and cardiac biomarkers are the most significant predictors. TabPFN achieves an F1 score of 86.3% using medical records and cardiac biomarkers.
• Limited Value of Vascular Biomarkers: Adding vascular biomarkers (e.g., Qanadli score) to the tabular model does not significantly improve performance (86.3% vs. 86.8% with cardiac only; the addition of vascular biomarkers resulted in a negligible change).
• GNN Performance: GNNs applied to vascular graphs without global features perform poorly (F1 ~50-55%). When global features are added, GNNs improve but fail to outperform the strong tabular baseline. The best GNN configuration (GPS with Cross-Attention) achieves 83.8% F1, which is statistically indistinguishable from the TabPFN baseline (86.3%) and does not represent a significant improvement.
• Hypothesis Testing:
  • Overfitting: Ruled out; generalization gaps were consistent with larger datasets.
  • Learning Failure: Ruled out; GNNs successfully regressed vascular biomarkers (Total Embolism Volume, Qanadli score) better than tabular models, proving they can learn graph patterns.
  • Label Bias: Ruled out; model errors did not correlate with ambiguous cases defined by guideline version changes.

Significance and Conclusion
The paper presents the first study evaluating the utility of pulmonary vascular graphs for PE risk stratification. Contrary to the intuition that detailed vascular topology and thrombus burden would enhance risk prediction, the results suggest that vascular graphs do not provide additional discriminative information beyond medical records and cardiac biomarkers for this specific task.

The authors conclude that while GNNs can learn complex patterns within the vascular graph (as evidenced by biomarker regression), these patterns do not translate to improved patient-level risk stratification when compared to simpler global features. This finding challenges the assumption that graph-based representations of the pulmonary vasculature are inherently superior for this clinical application, suggesting that current clinical biomarkers and medical history may already capture the necessary risk information. The code is made available to support further investigation.