Trustworthy multi-modelling for Liver Cirrhosis diagnosis using retrieval-augmented generation

Abstract

Liver cirrhosis is one of the leading causes of morbidity and mortality worldwide, disproportionately affecting populations in low- and middle-income countries where access to advanced diagnostic tools is limited. Traditional diagnostic methods such as liver biopsy are invasive, expensive, and often unavailable, contributing to delayed detection and poor treatment outcomes. Existing AI-based diagnosis tools struggle with interpretability and often fail to integrate diverse clinical inputs. This work addresses these challenges by providing a trustworthy multi-model diagnostic tool for liver cirrhosis that integrates machine learning on clinical data, computer vision on ultrasound images, and a Retrieval-Augmented Generation chatbot trained on 40 PubMed clinical trials and its exact match accuracy was 95%, its ECE was 0.25,invariance accuracy was 85% and its inference time was 7 seconds. Using late fusion, the system combines outputs to enhance the accuracy and generalization capability, during training hence achieving a training accuracy of 76.73%, validation accuracy of 90.26%, with a training loss of 0.4374 and validation loss of 0.3318, indicating strong performance and effective generalization to unseen inputs. Among machine learning models, the XGBoost Classifier achieved 95.9% accuracy, while VGG16 performed best in ultrasound image classification with 74% test accuracy. LIME, SHAP, and Grad-CAM XAI techniques were integrated to show features contribution towards model output. By delivering accurate, interpretable, and accessible liver cirrhosis diagnostics, this tool holds promise for enhancing early disease detection and supporting clinical decision making, particularly in low-resource settings. Its integration of late fusion multimodelling technique, XAI and Retrieval Augmented Generation approach contributes to trustworthy AI adoption in health care advancing both patient outcomes and diagnostic equity.