Explainable Cross-Disease Reasoning for Cardiovascular Risk Assessment from LDCT

Yifei Zhang, Jiashuo Zhang, Xiaofeng Yang, Liang ZhaoarXiv 2025 · DOI · arXiv

Introduction

Low-dose computed tomography (LDCT) is the established standard for lung cancer screening. However, these scans inherently capture the heart and surrounding vasculature, presenting a significant opportunity to screen for cardiovascular disease (CVD)—the leading cause of death globally. The "cardiopulmonary nexus" describes the deep physiological interconnection between the lungs and the heart; for instance, chronic lung inflammation or emphysema can directly stress cardiac function. This research introduces an explainable AI framework that leverages this nexus, mimicking clinical reasoning to provide accurate and transparent CVD risk assessments from routine lung scans.

Methodology: A Three-Stage Reasoning Framework

The proposed framework moves beyond traditional "black box" models by decomposing the diagnostic process into three collaborative stages:

  1. Dual-Pathway Visual Perception: The system utilizes a 3D CNN (MedicalNet) to extract fine-grained features from a localized cardiac region of interest (ROI). Simultaneously, it incorporates a "Lung Cancer Risk Prior" to capture the systemic biological stress of the patient. As illustrated below, the model identifies various pulmonary abnormalities across multiple anatomical planes, which serve as the foundation for cross-disease reasoning.
Detection of various lung pathologies including emphysema, bronchiectasis, nodules, and fibrosis across axial, coronal, and sagittal views.
  1. Agentic Pulmonary-to-Cardiac Reasoning: Acting as the system's cognitive core, an LLM (such as GPT-5-nano) equipped with medical knowledge retrieval identifies lung abnormalities and reasons about their impact on the heart. It retrieves established medical literature to link conditions like fibrosis to cardiac risks such as pulmonary hypertension, generating a logical chain of thought.
  2. Knowledge-Aware Multimodal Fusion: Finally, the system fuses raw 3D image features, lung risk scores, and the LLM’s reasoning embeddings to produce a final prediction for CVD screening and mortality.

Data and Experimental Setup

The framework was trained and evaluated using the National Lung Screening Trial (NLST) dataset, a high-quality cohort consisting of 32,136 LDCT scans from 10,395 subjects. The use of this longitudinal dataset is crucial, as it allows the AI to be evaluated on its ability to predict cardiac events and mortality that occurred years after the initial screening scan.

Key Insights: Reasoning as a Performance Driver

A striking finding of this study is that adding an LLM reasoning step does more than just provide explainability—it actually improves predictive accuracy. By forcing the model to align its visual features with medical logic, the system learns more robust, physiologically grounded patterns. The authors treat lung cancer risk not just as a separate diagnosis, but as a "systemic prior" for cumulative biological stress. This context helps the model understand that a patient with high lung risk is systemically more vulnerable to cardiac issues. Additionally, by isolating the cardiac region for the 3D CNN, the model avoids spurious correlations from irrelevant anatomical structures.

Results: State-of-the-Art Accuracy and Interpretability

The framework outperformed several high-profile baselines, including the Sybil model and the M3FM foundation model:

  • CVD Screening: Achieved an AUC of 0.919, a significant improvement over the previous state-of-the-art of 0.892.
  • CVD Mortality: Achieved an AUC of 0.838, outperforming image-only models.

Qualitatively, the model provides human-verifiable rationales. While visual heatmaps (Grad-CAM) show the model focusing on coronary arteries and ventricular walls, the LLM produces explanations such as: "Pleural effusion and fibrotic changes reduce pulmonary compliance... leading to hemodynamic overload that elevates cardiovascular risk."

Conclusion and Future Work

This framework demonstrates that integrating medical reasoning into AI models can bridge the gap between high-performance deep learning and clinical interpretability. While currently limited by its retrospective nature and the latency associated with LLM APIs, future work will focus on prospective clinical trials and the development of smaller, specialized local models to enhance real-time utility in hospital workflows.

Summary updated

BibTeX
@misc{zhang2025explainable,
  author = {Yifei Zhang and Jiashuo Zhang and Xiaofeng Yang and Liang Zhao},
  title = {{Explainable Cross-Disease Reasoning for Cardiovascular Risk Assessment from LDCT}},
  year = {2025},
  eprint = {2511.06625},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.06625},
}
Figure