Multimodal Fusion Architectures for Early Non-Invasive Steatotic Liver Disease Detection: Integrating Unstructured EHR Clinical Text, Longitudinal Biomarkers, and Radiomics via Graph Neural Networks

• Billy Elly

  Lautech

  Author

Metabolic dysfunction-associated steatotic liver disease (MASLD) affects approximately 38.9% of the global population, with projections reaching 55.7% by 2040, yet early detection remains critically limited by the invasive nature of liver biopsy and the suboptimal performance of existing non-invasive tests . Current screening approaches fail to leverage the full spectrum of heterogeneous patient data available within electronic health records, leaving a substantial gap in early, accessible detection capabilities. This study presents a novel multimodal fusion framework integrating unstructured clinical text from EHR narratives, longitudinal biomarker trajectories, and radiomic features extracted from ultrasound and CT imaging, unified through a graph neural network architecture. The proposed system achieved an overall classification accuracy of 89.4% (AUC = 0.94) for detecting MASLD, significantly outperforming single-modality approaches by 12-18% and demonstrating superior performance in identifying early-stage disease (F0-F1 fibrosis) where current non-invasive tests show limited sensitivity . The framework's ability to capture cross-modal interactions through graph-based relational reasoning enables more robust and interpretable predictions. These findings establish a replicable, scalable paradigm for multimodal disease detection that can be extended to other chronic conditions, offering a pathway toward more accessible, community-level screening solutions .

• pdf

Copyright (c) 2026 Billy Elly (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.