Korištenje velikih jezičnih modela kao izvora medicinskih informacija za pacijente i liječnike

master's thesis

• Title (english): Using large language models as a source of medical information for patients and doctors
• Author: Adrian Sallabi
• Mentor: Andrija Štajduhar (mentor)
• Committee member: Zdenko Sonicki (predsjednik povjerenstva)
• Committee member: Sanja Musić Milanović (član povjerenstva)
• Committee member: Andrija Štajduhar (član povjerenstva)
• Granter: University of Zagreb; School of Medicine; (Department of Medical Statistics, Epidemiology and Medical Informatics); Zagreb
• Defense date and country: 2024-09-11, Croatia
• Scientific / art field, discipline and subdiscipline: INTERDISCIPLINARY AREAS OF KNOWLEDGE; Biotechnology in Biomedicine (natural science, biomedicine and healthcare, bioethics area

Abstract

Veliki jezični modeli (eng. Large language models, LLM) predstavljaju najsuvremenije sustave za obradu prirodnog jezika, sposobne za rješavanje brojnih složenih zadataka u različitim industrijama, uključujući medicinu. Njihove potencijalne primjene u području zdravstva obuhvaćaju kliničku podršku, dijagnostiku te personaliziranu skrb. Međutim, prije njihove šire primjene u zdravstvenom sustavu, potrebno je temeljito razumjeti njihove prednosti i nedostatke. U ovom radu opisane su prednosti i ograničenja velikih jezičnih modela u medicinskom kontekstu, s naglaskom na njihovu pouzdanost pri generiranju medicinskih informacija. Prikazana je povijest razvoja ovih modela, te su detaljno obrađene mogućnosti njihove integracije u kliničku praksu. Razmotreni su etički izazovi poput zaštite privatnosti pacijenata, sigurnosti podataka i problematike halucinacija. Naposljetku, primjena velikih jezičnih modela prikazana je kroz praktični primjer izrade AI agenta za odgovaranje na pitanja pacijenata, temeljenog na GPT-4 modelu, koji je uspješno demonstrirao sposobnost rješavanja specifičnih medicinskih zadataka. Evaluirani su rezultati i objašnjene potencijalne implikacije šire primjene velikih jezičnih modela u zdravstvenoj praksi.

Abstract (english)

Large language models (LLMs) represent state-of-the-art systems for natural language processing, capable of solving numerous complex tasks across various industries, including medicine. Their potential applications in healthcare include clinical support, diagnostics, and personalized care. However, before their broader implementation in the healthcare system, it is crucial to thoroughly understand their advantages and limitations. This paper outlines the benefits and drawbacks of large language models in a medical context, with a focus on their reliability in generating medical information. The development history of these models is presented, and the possibilities for their integration into clinical practice are explored in detail. Ethical challenges, such as patient privacy protection, data security, and the issue of hallucinations, are also discussed. Finally, the application of large language models is demonstrated through a practical example of developing an AI agent for answering patient queries, based on the GPT-4 model, which successfully showcased its ability to address specific medical tasks. The results were evaluated, and the potential implications of the broader adoption of large language models in healthcare practice are explained.

• Keywords: medicinske informacije; veliki jezični modeli; GPT-4; klinička praksa
• Keywords (english): medical information; large language models; GPT-4; clinical practice
• Language: croatian
• URN:NBN: urn:nbn:hr:105:376847
• Study programme: Title: Medicine; Study programme type: university; Study level: integrated undergraduate and graduate; Academic / professional title: doktor/doktorica medicine (doktor/doktorica medicine)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Open access
• Created on: 2024-09-09 18:30:21