Deep learning can predict cardiovascular events from liver imaging.

Area of Science:

• Cardiovascular medicine
• Medical imaging
• Artificial intelligence

Background:

• Cardiovascular disease is a leading cause of death, with metabolic alterations playing a key role.
• The liver, as the primary metabolic organ, may indicate early cardiovascular risk.
• Current methods for quantifying cardiovascular risk are challenging.

Purpose of the Study:

• To explore the use of transformer neural networks on liver MRI data for enhanced cardiovascular risk prediction.
• To develop a non-invasive imaging-based biomarker for cardiovascular risk stratification.

Main Methods:

• A vision transformer backbone was trained in a self-supervised manner on liver MRI data from the UK Biobank.
• The trained encoder was used to predict cardiovascular outcomes without manual feature selection.
• Performance was evaluated using fivefold cross-validation, comparing predicted risk scores against actual outcomes.

Main Results:

• The model, trained on 44,672 liver MRIs, achieved a mean AUC of 0.70 (95% CI: 0.69-0.72, p <0.001) for predicting major adverse cardiac events.
• Statistical analyses (ANOVA, log-rank tests) confirmed significant differences in risk scores and survival across prediction groups (p <0.001).

Conclusions:

• Vision transformer-based models show promise as quantifiable biomarkers for cardiovascular risk assessment.
• These models capture subtle metabolic and vascular information from liver MRIs, indicating strong predictive performance.
• Further prospective studies and external validation are needed to confirm clinical utility.