COPD Comorbidity Risk Prediction

• The paper introduces an automated imaging biomarker pipeline (PHT-BOT) that extracts the PA/Ao ratio from CT scans with high clinical accuracy and reproducibility.
• It details a causal heterogeneous graph representation learning (CHGRL) framework that improves prediction accuracy by over 5% compared to conventional graph models.
• The integration of imaging, clinical, and laboratory data enables scalable risk stratification and personalized management for COPD patients.

Chronic Obstructive Pulmonary Disease (COPD) comorbidity risk prediction encompasses a spectrum of computational and statistical methods that assess an individual’s propensity to develop clinically significant comorbidities, notably secondary pulmonary hypertension (PHT), using multi-modal data-derived biomarkers and advanced machine learning algorithms. State-of-the-art methodologies include deep learning-based extraction of noninvasive imaging surrogates for vascular remodeling as well as causal representation learning on heterogeneous patient–disease graphs, aiming to uncover robust predictors and enable risk stratification at scale.

1. Clinical and Computational Background

COPD is a leading driver of morbidity and mortality globally, with comorbidities such as pulmonary hypertension (PHT) contributing substantially to adverse outcomes. Secondary PHT in COPD arises from vascular remodeling due to chronic hypoxia, resulting in elevated pulmonary arterial pressure and potential right heart failure. Invasive measurements (e.g., right-heart catheterization) are not suited for large-scale screening. Radiographically, the diameter of the pulmonary artery (PA) relative to the ascending aorta (Ao)—with a PA/Ao ratio > 1—serves as a validated noninvasive marker for PHT. This marker predicts not only PHT but also future COPD exacerbations, hospitalizations, and all-cause mortality. However, manual CT measurement is time-consuming and underutilized, motivating the need for automated, reproducible solutions (Chettrit et al., 2019).

Simultaneously, accurate risk prediction for COPD comorbidity requires leveraging multidimensional and incomplete clinical data, modeling heterogeneous relationships (e.g., patient–disease, patient–patient, and disease–disease), and inferring causality rather than simple association. This complexity gives rise to graph-based and causal-inference-driven architectures (Zhou et al., 22 Dec 2025).

2. Automated Imaging Biomarkers: Deep Learning for PA/Ao Ratio Extraction

PHT-BOT is a fully automated pipeline that quantifies the PA/Ao ratio from routine contrast-enhanced chest CTs as a noninvasive biomarker for PHT and COPD comorbidity risk (Chettrit et al., 2019). Its workflow comprises two principal convolutional neural network (CNN) stages: axial slice selection and vessel segmentation.

• Axial Slice Selection: Employs a CNN classification network to identify the axial slice at the main PA bifurcation. The architecture features three convolutional blocks, batch-normalization, ReLU activations, max-pooling, and two fully connected layers. Training uses manual reference labels and data augmentation.
• Vessel Segmentation: Utilizes a U-Net variant with a VGG16-style encoder for multi-windowed CT slice input, generating soft-max segmentation for Ao, PA, and background. The PA diameter is obtained by skeletonization and width measurement orthogonal to the trunk, and Ao diameter by ellipse fitting to the mask.

Measurement and Evaluation:

• The PA/Ao ratio is computed as ${\rm PA/Ao} = D_{\rm PA} / D_{\rm Ao}$.
• Test Pearson $r$ with radiologist measurements: 0.93 (Ao), 0.92 (PA).
• Average diameter bias: –0.94 mm (Ao), –0.86 mm (PA).
• Diagnostic threshold: PA/Ao > 1 for PHT.
• Confusion matrix and performance (test set): Sensitivity = 64.6%, Specificity = 97.0%, Positive Predictive Value = 80.3%, Accuracy = 91.9%.
• Runtime: 70 s per study (CPU), 22 s (GPU).
• Reviewer acceptance: 99.05% of automated measurements deemed clinically acceptable; 99.76% located at the correct slice.

This solution provides a scalable, reproducible imaging-based biomarker that can be integrated with other clinical and functional data for risk modeling.

3. Heterogeneous Graph-Based Causal Learning for COPD Comorbidity Risk

The Causal Heterogeneous Graph Representation Learning (CHGRL) framework addresses COPD comorbidity risk prediction by modeling interactions within a heterogeneous graph comprising patients, diseases, and their relationships (Zhou et al., 22 Dec 2025). The methodology includes the following components:

• Graph Construction: Nodes represent 516 patients and 386 diseases. Relations include patient–patient similarity, patient–disease associations, and inferred disease–disease similarity. Node features for patients are 68-dimensional laboratory vectors; missing data imputed by a causal-regularized latent factor model (CSINLF).
• Architecture:
  • Layer-wise heterogeneous message passing aggregates information from neighbors under all relation types using relation-specific projections.
  • Causal-strength estimation assigns an edge-specific, learned scalar ($CS_{ij}$) predicting causal influence via a dedicated MLP.
  • Causal attention computes an attention coefficient ($a_{ij}$) integrating node features and $CS_{ij}$, which gates the update.
  • Counterfactual embeddings are obtained by hypothetical interventions on node features, using a two-layer MLP.
  • The final node update is the sum of conventional GNN propagation and the causal message.
• Objective:

$$L_{\rm total} = L_{\rm CE} + \lambda_1 L_{\rm CF} + \lambda_2 L_{\rm CR}$$

$L_{\rm CE}$: cross-entropy loss; $L_{\rm CF}$: counterfactual loss; $L_{\rm CR}$: causal regularization.

• Implementation: Trained with Adam optimizer, full graph minibatching, $d=64$ dim, dropout 0.5, PyTorch and DGL.

Evaluation and Performance (mean $\pm$ std over five runs; n=516):

Model	AUC	ACC	F1
RGCN	0.7900±0.0356	0.7777±0.0491	0.7729±0.0472
CHGRL	0.8339±0.0455	0.8165±0.0290	0.8090±0.0331

CHGRL demonstrates a +5.39% increase in AUC, +3.88% in accuracy, and +4.56% in weighted F1-score compared to the strongest heterogeneous GNN baseline. Ablation studies show 4–6% AUC reduction if counterfactual or causal modules are removed.

4. Integration of Multi-modal Features and Risk Stratification

Both PHT-BOT and CHGRL facilitate integration with diverse patient data modalities to enable comprehensive comorbidity risk prediction in COPD. Imaging-derived biomarkers (e.g., PA/Ao ratio, absolute PA diameter) can be fused with:

• Demographics: age, sex, BMI, smoking history, exacerbation frequency.
• Pulmonary function tests: FEV1, FEV1/FVC.
• Laboratory biomarkers: NT-proBNP, CRP, IL-6, troponin.
• Echocardiographic measures: right ventricular function.

Editor's term: "Multimodal fusion"—the integration of imaging, clinical, laboratory, and physiologic data—enables multivariable models (gradient boosting, deep neural networks, or graph-based architectures) for predicting endpoints such as 12-month exacerbation risk, hospital admission, or mortality (Chettrit et al., 2019).

5. Model Validation, Limitations, and Interpretability

Validation:

• PHT-BOT reports high agreement with radiologist-derived measures (Pearson r > 0.9, diameter bias < 1 mm).
• Clinical reviewer studies endorse the acceptability of automated imaging-based risk biomarkers.
• CHGRL uses t-SNE visualization to demonstrate clear separation of case/control embeddings.

Limitations:

• PHT-BOT currently requires contrast-enhanced CT and has not been externally validated on multi-center data or noncontrast protocols.
• CHGRL does not explicitly model demographic variability or temporal dynamics; external validation with longitudinal EHR and multi-modal integration is not provided.
• The causal regularizer in CHGRL uses a generic penalty; domain-integrated priors or back-door adjustments are needed for robust causal inference in deployment contexts.
• No prospective clinical trials have yet demonstrated improved patient outcomes via these automated risk-prediction workflows (Chettrit et al., 2019, Zhou et al., 22 Dec 2025).

A plausible implication is that future work should address generalization across heterogeneous data sources and real-world cohorts while advancing causal interpretability.

6. Impact and Future Directions

Automated COPD comorbidity risk prediction methods provide scalable tools for early intervention and personalized management in high-burden populations. PHT-BOT introduces a reproducible imaging biomarker pipeline, while CHGRL extends predictive modeling to the causal, relational structure of patient-disease interactions.

Key directions for further research include:

• Adaptation of imaging pipelines to noncontrast and other acquisition protocols.
• Extension of graph-based models to handle temporal evolution (time-series graphs), incorporate domain knowledge for causal regularization, and support multi-site, prospective validation.
• Clinical outcome trials to determine the real-world impact of integrated multimodal risk scores on COPD management and healthcare resource utilization.

These advances collectively position automated and causally robust comorbidity risk stratification as central components in future COPD clinical decision support systems (Chettrit et al., 2019, Zhou et al., 22 Dec 2025).