A Multi-Modal Deep Fusion Network: Enhancing Graft Survival Prediction in Non-Alcoholic Fatty Liver Disease Patients Prior to Liver Transplantation

NAFLD has become a major contributor to chronic liver conditions worldwide and is increasingly recognized as a significant factor influencing outcomes in liver transplantation. One of the key challenges clinicians faces is the accurate prediction of graft survival in NAFLD patients, owing to the complex and heterogeneous nature of the disease. Traditional models relying on single data modalities often fail to capture the full clinical picture, leading to suboptimal prediction performance. This paper aims to develop an accurate graft survival prediction model for NAFLD patients using multi-modal data that includes clinical, histological, and imaging features. We propose a novel Multi-Modal Deep Fusion Network (MMDF-Net), which combines a CNN for MRI image analysis with LSTM networks for sequential clinical data. The model was validated using the NHANES III dataset and benchmarked against existing methods such as LSTM, SVM, Random Forest, Logistic Regression, CNN, and XGBoost. MMDF-Net achieved a prediction accuracy of 94%, outperforming all baseline models by 6–12% in accuracy and showing an 8% improvement in F1-score. It also demonstrated superior AUC-ROC performance. Attention maps and calibration plots confirmed the model’s interpretability and reliability. Overall, the proposed method effectively improves graft survival prediction in NAFLD patients and offers a promising tool for enhancing clinical decision-making in liver transplantation.