Explainable Cross-Disease Reasoning for Cardiovascular Risk Assessment from Low-Dose Computed Tomography

Technical Summary: Explainable Cross-Disease Reasoning for Cardiovascular Risk Assessment from Low-Dose Computed Tomography

1. Problem Statement

Low-dose chest computed tomography (LDCT) is the standard of care for lung cancer screening but inherently captures both pulmonary and cardiac structures. While epidemiological evidence establishes a bidirectional link between lung cancer and cardiovascular disease (CVD)—driven by shared mechanisms like systemic inflammation and oxidative stress—current computational approaches treat these domains in isolation. Predominant methods model lung cancer detection and CVD risk stratification as separate tasks, analyzing pulmonary nodules and cardiac structures independently. This "radiological isolation" discards critical cross-organ information. Furthermore, existing cross-disease or foundation models often function as black boxes, offering only coarse post-hoc attribution rather than explicit, interpretable reasoning pathways that link pulmonary findings to cardiac outcomes. The challenge lies in leveraging the latent, multi-step physiological relationship between pulmonary abnormalities and cardiovascular risk without relying solely on unstructured feature concatenation or opaque deep learning models.

2. Methodology

The authors propose an Explainable Cross-Disease Reasoning Framework that treats a single LDCT scan as a knowledge-bearing source for cardiopulmonary risk assessment. The architecture decomposes the prediction task into four clinically grounded components organized into three interconnected pathways:

• Dual-Pathway Visual Perception:

  • Lung-Risk Prior: A frozen Lung Cancer Risk Predictor (based on Sybil) processes the full lung volume to extract a temporal malignancy risk vector. This serves as a high-level proxy for chronic systemic inflammation and overall disease burden, acting as a contextual prior rather than a direct clinical output.
  • Cardiac Feature Extraction: A Cardiac Localization module isolates the cardiac subvolume (ROI) to remove thoracic background noise. A trainable 3D CNN (MedicalNet backbone) then encodes fine-grained morphological biomarkers (e.g., coronary artery calcification, chamber geometry) into a dense representation.

• Agentic Pulmonary-to-Cardiac Reasoning:

  • Pulmonary Perception Module: A deterministic interface (using CT-CLIP) translates the LDCT volume into a structured set of normalized pulmonary findings (concept, description, confidence). This creates a discrete bottleneck, preventing the reasoning agent from accessing raw voxels or patient metadata.
  • Knowledge Reasoning Agent: Given the structured findings and a curated cardiopulmonary knowledge corpus (retrieved via text similarity), this agent constructs a mechanism graph. It links findings to intermediate mechanisms (e.g., hypoxemia, systemic inflammation) and cardiopulmonary effects (e.g., pulmonary hypertension), ensuring every link is supported by retrieved evidence.
  • Diagnostic Reasoning Agent: This agent synthesizes the mechanism graph into a bounded natural-language rationale. The output is a structured state containing mechanisms, effects, and a concise rationale, strictly constrained to avoid hallucinating patient history or making direct diagnostic verdicts.

• Knowledge-Aware Multimodal Fusion:

  • A semantic encoder projects the textual rationale into a latent embedding.
  • The final predictor fuses three normalized representations: the cardiac visual features, the frozen lung-risk prior, and the reasoning-derived rationale embedding.
  • A task-specific head predicts CVD screening status and CVD mortality using binary cross-entropy loss. Only the cardiac encoder and fusion heads are trainable; the perception, reasoning, and lung-risk modules remain frozen or externally instantiated to ensure the reasoning pathway is auditable.

3. Key Contributions

• Formulation of Cross-Disease Reasoning: The paper formulates LDCT-based CVD prediction as an explicit reasoning problem, converting pulmonary findings into mechanism-level cardiovascular hypotheses rather than treating them as unstructured auxiliary features.
• Auditable Multimodal Framework: The authors develop a framework that anchors predictions in localized cardiac visual evidence while exposing a transparent, step-by-step pulmonary-to-cardiac reasoning pathway. This shifts the Large Language Model (LLM) from a generic text aggregator to a constrained, knowledge-reasoning component.
• Targeted Evaluation Suite: The study provides a rigorous evaluation protocol including cardiac-specific baselines, vision-only fusion controls, structured semantic controls, fixed-rule propagation, and reasoning-backend ablations to isolate the source of performance gains.

4. Experimental Results

Experiments were conducted on the National Lung Screening Trial (NLST) cohort (32,136 LDCT scans from 10,395 subjects).

• Performance: The proposed framework achieved state-of-the-art performance with an AUC of 0.919 for CVD screening and 0.838 for CVD mortality prediction (using a 3D ResNet-200 backbone). This outperformed strong baselines including foundation models (M3FM, Sybil) and specialized cardiac models (DeepCAC, KAMP-Net).
• Ablation Studies:
  • Cardiac Region Only: Achieved moderate performance (AUC ~0.886 for screening), highlighting the value of localized cardiac morphology.
  • Lung Region Only: Performed poorly (AUC ~0.73), indicating pulmonary visual evidence alone is insufficient.
  • Fusion: Adding the lung-risk prior and the reasoning branch provided incremental gains, demonstrating that systemic context and mechanism-level reasoning complement local cardiac features.
• Control Variants: The performance gain was not explained by simply adding more thoracic visual features (vision-only fusion), fixed rule propagation, or a single proprietary reasoning backend. The structured reasoning approach outperformed fixed rules and alternative LLM backends (Qwen, Baichuan), suggesting the gain stems from the specific cross-disease formulation.
• Robustness: The model showed resilience to perturbations in pulmonary finding extraction (e.g., varying confidence thresholds or random dropout of findings) and maintained performance across different reasoning backends, supporting the feasibility of local deployment.

5. Significance and Claims

The paper claims that its primary significance lies in providing an auditable approach to cross-disease cardiovascular risk assessment. By explicitly modeling the physiological interplay between lung and heart through a constrained reasoning pathway, the framework moves beyond the "black box" nature of current deep learning models.

• Interpretability: The system generates natural-language rationales grounded in medical knowledge and linked to specific pulmonary findings, allowing clinicians to inspect the logic (e.g., how fibrosis leads to hypoxemia and subsequently heart failure risk).
• Clinical Utility: The framework utilizes LDCT scans already acquired for lung screening, offering an opportunistic window for CVD assessment without requiring additional scans or manual calcium scoring.
• Modularity: The design separates risk priors, perception, reasoning, and cardiac evidence, allowing individual components to be updated or replaced (e.g., with better local LLMs) without retraining the entire system or changing the evaluation protocol.

The authors remain modest regarding clinical deployment, noting that the generated rationales are mechanism-level explanations rather than causal proof, and that prospective validation across diverse scanners and populations is required before clinical adoption.