Deep learning can predict cardiovascular events from liver imaging

Gregory Patrick Veldhuizen, Tim Lenz, Didem Çifçi, Marko van Treeck, Jan Clusmann, Yazhou Chen, Carolin V. Schneider, Tom Luedde, Peter W. de Leeuw, Ali El‐Armouche, Daniel Truhn, Jakob Nikolas KatherJHEP 2025 · DOI

TL;DR

Researchers have developed a deep learning model using liver MRI scans to predict major adverse cardiovascular events (MACE). This AI approach shows promise as a non-invasive tool for cardiovascular risk assessment, rivaling established scores and potentially enabling earlier, personalized interventions.

What Problem Does This Paper Address?

Cardiovascular diseases (CVDs) are a leading cause of death and morbidity worldwide. Accurately identifying individuals at high risk of MACE is crucial for timely intervention and improved patient outcomes. Current risk assessment models, while valuable, often rely on traditional risk factors and may not fully capture the subtle, early indicators of CVD. The liver, as a central metabolic organ, is increasingly recognized for its role in metabolic alterations that can precede cardiovascular disease. However, quantifying these liver-related changes for predictive purposes has been challenging. This paper aims to bridge this gap by exploring the use of deep learning on liver MRI data to predict cardiovascular risk.

Methods

The core of this study lies in the application of a vision transformer (ViT), a powerful neural network architecture, to analyze liver MRI scans. The researchers employed a self-supervised learning (SSL) approach, specifically using Momentum Contrastive Learning (MoCo) v3, to pretrain the ViT model. This SSL pretraining allows the model to learn meaningful feature representations from unlabeled liver MRI data, overcoming the common bottleneck of requiring large, manually labeled datasets.

After pretraining, the ViT's encoder weights were frozen, and its output feature vectors were fed into a linear classifier head. This classifier was then trained in a supervised manner to predict MACE. To understand which parts of the MRI were most influential for the model's predictions, the researchers generated heatmaps by combining image gradients with original pixel values, highlighting areas of high importance. The model's performance was evaluated using fivefold cross-validation.

Datasets

The primary dataset used in this study is the UK Biobank. Specifically, the researchers utilized liver MRI scans from a subset of participants. The dataset comprised 44,672 liver MRIs for the training and cross-validation process. For external validation, a separate cohort of 750 participants was used. The study also references the use of SCORE2 (Systematic Coronary Risk Evaluation 2) and its variants (SCORE2-Diabetes, SCORE2-OP) as established risk assessment tools for comparison.

Key Insights and Ideas

  • The Liver as a Predictive Organ: The study strongly suggests that the liver, through its imaging characteristics, contains significant prognostic information for cardiovascular risk, acting as a metabolic indicator of early disease.
  • Vision Transformers for Medical Imaging: The successful application of vision transformers, pre-trained with SSL, demonstrates their potential for extracting complex patterns from medical images without extensive manual annotation.
  • No Manual Feature Engineering: A key advantage highlighted is the model's ability to learn predictive features directly from the images, bypassing the need for manual feature selection or labeling, which can introduce bias.
  • Focus on Key Anatomical Structures: The generated heatmaps reveal that the model consistently focuses on specific anatomical regions within the liver and surrounding vasculature (hepatic veins, inferior vena cava, abdominal aorta), suggesting their critical role in cardiovascular health.
  • Complementary to Existing Tools: The model shows promise not as a replacement, but as a complementary tool to existing risk scores like SCORE2, potentially offering enhanced risk stratification, especially for individuals with intermediate risk.

Key Results

  • The deep learning model achieved a mean Area Under the Receiver Operating Characteristic Curve (AUC) of 0.70 (95% CI: 0.69-0.72) in predicting major adverse cardiac events (MACE) through fivefold cross-validation.
  • The model demonstrated significant discriminatory power to stratify patients based on risk, with pairwise comparisons of prediction score groups showing highly significant stratification.
  • Survival analysis revealed a statistically significant difference in survival outcomes between participants with prediction scores above and below 0.5 (log-rank test statistic = 43, p <0.001).
  • The model's performance was comparable to, and in some aspects potentially superior to, established risk scores like SCORE2. For instance, the precision-recall curve of the DL model outperformed SCORE2, especially at lower recall levels.
  • The model showed a statistically significant difference in prediction scores across different BMI categories, with higher scores in obese and overweight groups, aligning with known CVD risk factors.

Limitations and Future Work

The authors acknowledge several limitations:

  • Retrospective Nature: The study is retrospective, which can introduce biases and limit generalizability.
  • Single-Slice MRI: The reliance on single-slice liver MRI images might limit the comprehensive capture of all relevant information.
  • Limited Liver Health Metrics: The inability to incorporate detailed liver fat measurement data from the UK Biobank due to access policy changes is a significant limitation. While participants with known liver disease were excluded, subclinical or undiagnosed conditions could still influence results.
  • Lack of Fully Independent External Validation Cohort: This raises questions about the model's generalizability to diverse populations.

Future work should focus on:

  • Prospective Studies: Conducting prospective studies to evaluate the real-time application of the model in clinical settings.
  • External Validation: Validating the model with external datasets from diverse populations to ensure robustness and generalizability.
  • Integration into Clinical Practice: Further research is needed to confirm findings and improve model performance for integration into clinical workflows, potentially as an adjunct to existing risk assessment tools.
  • Incorporating More Data: Exploring the inclusion of additional clinical and imaging data, as well as detailed liver health metrics, to enhance predictive accuracy.

Summary updated

Abstract

Our study demonstrates that deep learning applied to liver MRI can predict cardiovascular risk, highlighting the role of the liver as a metabolic indicator of early cardiovascular disease. These findings are significant for clinicians and researchers seeking non-invasive, imaging-based biomarkers for cardiovascular risk stratification, particularly in patients who might not yet exhibit overt symptoms. If validated in prospective studies, this approach could enhance current risk assessment models, allowing for earlier and more personalized interventions in high-risk individuals. However, further validation is necessary before clinical implementation, ensuring broad applicability and integration into existing prevention frameworks.
BibTeX
@inproceedings{veldhuizen2025deep,
  author = {Gregory Patrick Veldhuizen and Tim Lenz and Didem Çifçi and Marko van Treeck and Jan Clusmann and Yazhou Chen and Carolin V. Schneider and Tom Luedde and Peter W. de Leeuw and Ali El‐Armouche and Daniel Truhn and Jakob Nikolas Kather},
  title = {{Deep learning can predict cardiovascular events from liver imaging}},
  booktitle = {JHEP},
  year = {2025},
  doi = {10.1016/j.jhepr.2025.101427},
}
Figure