The Application of Large Language Models for Treatment Recommendations in Oral Cancer

Abstract

The thesis explores the use of Large Language Models (LLMs) to assist in radiotherapy treatment recommendations for patients with oral cancer, which is a subtype of head and neck cancer (HNC) that involves a complicated process of clinical decision-making. The research thesis examines the manner in which sophisticated LLM-based models could be used to support clinicians through the analysis of structured clinical data and subsequent creation of guideline-congruent treatment recommendations. Four LLMs, namely, BioGPT, Llama, GEMMA, and Meditron were the chosen LLM architectures for radiotherapy treatment recommendations according to the patient characteristics and tumour staging information. In the study, a quantitative experimental design was adopted and anonymised clinical datasets were used. Ordered patient records have been converted to natural language prompts, which can be used to interact with language models. Quantized Low-Rank Adaptation (QLoRA) was used to implement parameter-efficient fine-tuning, which enables the models to acquire task-related knowledge without having to retrain the whole network. To assess the performance of model predictions, typical performance metrics were used, such as accuracy, precision, recall, F1-score and confusion matrix analysis. The findings show that LLM-based architectures are able to process structured clinical information and provide precise treatment advice. It was revealed from the experiments that Llama-3.1 showed the best predicting capacity among all the tested models with an accuracy rate of about 93-94%, whereas BioGPT had good biomedical capabilities and provided predictions with an accuracy rate of about 86-87%. The comparative analysis also revealed that the performance in model architectures differed, thereby demonstrating the significance of domain-specific training and prompt engineering. In general, the results are indicative of the fact that the LLMs can be used as clinical decision-support systems in oncology. Even though such systems cannot substitute professional judgment, they may improve the multidisciplinary decision-making processes, helping clinicians to analyse the patient data and provide evidence-based treatment.