Independent thesis Advanced level (degree of Master (Two Years)), 20 credits / 30 HE credits
This thesis, in collaboration with Arkus-AI, explores the development of an AI-driven genetic counselor designed to enhance genetic counseling services at Eiira. By utilizing advanced large language models, this project aims to improve the quality and accessibility of genetic counseling for individuals at risk of inherited cancers.
Genetic counseling involves providing individuals or families with information and guidance about genetic conditions. Typically conducted by healthcare professionals trained in genetics and counseling, its main objectives are to assess the risk of inheritable genes, provide information about these genes, assist in making informed decisions about genetic testing and management, and offer emotional support.
The core objective of this research is to create an AI-based conversational genetic counselor that assists patients by answering questions related to their medical or cancer history. This research leverages pre-trained Large Language Models (LLaMA2 and Phi-3) and utilizes Transformer architecture with a document-based generation approach to optimize the performance of these models during counseling sessions.
Integrating AI/LLM technology into genetic counseling aims to enhance the precision, personalization, and accuracy of the advice provided to Eiira’s clients, empowering them to make well-informed health decisions. This project seeks to combine the analytical capabilities of AI with the specific requirements of genetic counseling, aiming to set a new standard in personalized medical consultation.
To achieve this goal, the methodology employed is Retrieval-Augmented Generation (RAG). Initially, data embeddings are created and stored in vector databases. The genetic counselor then reviews the user's medical report and generates a summary using the models and LLMChains(the process of linking multiple large language models (LLMs) together in a sequential manner to accomplish complex tasks), which link multiple large language models sequentially to handle complex tasks. This summary explains identified gene variants and suggested medications from the report. The summary is then presented to the user and stored in the context history of the conversation. When a patient submits a query, the relevant information is retrieved from the stored embeddings and sources and presented to the user.
This research tested and compared two models, LLaMA2 and Phi-3, finding that LLaMA2 outperformed Phi-3. The models' performance was evaluated using metrics such as perplexity, Rouge-1, cosine, and Jaccard similarity. While LLaMA2 required more time to generate responses due to its larger size and needed more powerful GPUs, substantial memory, and increased energy consumption, it demonstrated superior effectiveness overall.
2024.