Medicinsk rådgivning med stora språkmodeller: förbättring av självutvärdering genom multiagentsystem
Year: 2025
Type: other
Abstract: The recent advances in the field of Large Language Models have led to their adoption in a wide range of fields. Their usefulness has been proven in many fields, such as in programming and in writing to assist and improve code and text. However, many fields which doubtlessly could benefit from adopting the use of Large Language Models, have remained hesitant to do so. One of the fields that has remained hesitant is the field of medicine. The safety-critical nature of the medical field requires a certain degree of technological conservatism, but due to the sensitive nature of medicine, some people might prefer to interact with what they might perceive to be a non-judgmental system, rather than a real person. Additionally, such a system could be used to streamline the process of receiving medication and perhaps alleviate potential understaffing issues. This thesis attempts to assert whether an LLM based system could be safely implemented in a medical counseling scenario, by designing a rudimentary system for a single medication, an emergency contraceptive pill. The system must be able to adhere to a strict conversation template, but still retain some conversational flexibility. Furthermore, it must be able to keep track of what is known about the current customer and based on said information select the right pill. The thesis presents a prototype implementation of an LLM agent-based system which adheres to the aforementioned requirements. The thesis covers methods for constructing such a system, focusing on methods for ensuring adherence to conversational requirements, minimizing hallucination and ensuring high-quality, factual responses. Furthermore, the thesis discusses the efficiency and potential issues of these methods, ethical and privacy considerations and ways of evaluating a system of this variety. Based on these topics of discussion and observations made during development of the system, areas of future work are addressed. The findings based on the prototype implementation suggest that it is possible to implement a system of this variety; the methods employed increased the determinism of the system output a significant amount. However, there is yet work to be done on perfecting the methods used and further generalizing the system to be more easily adapted to different contexts.
Source: Doria (University of Helsinki)
Author: Sabel, Joar
Language: English
Subfield: Health Informatics
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7115334493