Hepatocellular Carcinoma Prediction using Large Language Model and Meditron

This project investigated the integration of machine learning, large language models (LLMs), and the Meditron medical language model to enhance early prediction and interpretation of hepatocellular carcinoma (HCC) in resource-constrained healthcare settings such as Kenya, where late diagnosis and limited access to advanced diagnostic tools remain major challenges. Using a design science and experimental research approach, an AI-driven clinical decision-support prototype was developed to process unstructured clinical conversations, extract relevant medical features, predict HCC risk, and generate clinically interpretable explanations. A Random Forest classifier was trained on a balanced secondary clinical dataset comprising equal proportions of HCC and non-HCC cases, with preprocessing techniques including median imputation and feature scaling applied to ensure data quality and model robustness. The trained model achieved an accuracy of 0.8780, an F1-score of 0.8780, and an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.9452, demonstrating strong predictive performance. We applied an LLM to convert free-text clinical notes into structured features suitable for the trained predictive model allowing automated HCC risk prediction from conversational input. Then the prediction outputs comprising probabilistic risk scores were submitted to the Meditron model that gave medically grounded interpretations of the results by contextualizing anticipated risk in respect to known clinical indications of HCC. We integrated LLM-based data extraction, machine learning-based prediction, and Meditron-driven interpretation, which enhanced transparency, explainability, and clinical relevance of the system. Results demonstrate that hybrid AI architectures that combine predictive modeling and domain-specific reasoning with medical language can provide accurate, interpretable, and scalable tools for early HCC risk assessment with great potential to support clinical decision-making and improve cancer care delivery in low-resource healthcare settings.