The core element of this thesis project lies in the integration of a Knowledge Graph within a Retrieval-Augmented Generation (RAG) architecture for a conversational system supporting the clinical management of rheumatoid arthritis. The aim of the work is the construction and evaluation of a chatbot based on Large Language Models (LLMs), supplemented by information retrieval techniques from validated sources, with the goal of providing reliable and clinically relevant responses. The system is designed for use both by general practitioners, in the triage phase and in referring patients to specialist rheumatology visits, and by patients themselves, for monitoring the course of the disease during periods of clinical remission. During the course of the work, different information retrieval strategies were implemented and compared: the use of embeddings and semantic similarity alone, the use of the Knowledge Graph, and a hybrid solution combining both approaches. The study showed that embeddings are sufficient for simple questions directly related to guidelines but are not adequate for more complex clinical scenarios, for which the integration of the Knowledge Graph appears to improve performance, making the chatbot more robust and reliable. This thesis is part of the DHEAL-COM HUB project, funded by the Italian Ministry of Health, which aims to develop digital health solutions and community-based medicine. In this context, the approach presented contributes to advancing the state of the art in applying Artificial Intelligence techniques for supporting triage and monitoring of chronic patients, laying the foundation for increasingly integrated and effective digital tools in clinical practice.
L’elemento cardine di questo progetto di tesi risiede nell’integrazione di un Knowledge Graph all’interno di un’architettura di Retrieval-Augmented Generation (RAG) per un sistema conversazionale a supporto della gestione clinica dell’artrite reumatoide. La finalità del lavoro è la costruzione e la valutazione di un chatbot basato su Large Language Models (LLM), affiancato da tecniche di recupero delle informazioni da fonti validate, con l’obiettivo di fornire risposte affidabili e clinicamente rilevanti. Il sistema è concepito per essere utilizzato sia dal medico di medicina generale, nella fase di triage e di invio a visita specialistica reumatologica, sia dal paziente stesso, per il monitoraggio del decorso della malattia in condizioni di remissione clinica. Nel corso del lavoro sono state implementate e confrontate diverse strategie di recupero delle informazioni: l’utilizzo dei soli embeddings e della similarità semantica, l’impiego del Knowledge Graph e una loro soluzione ibrida. Lo studio ha mostrato che gli embeddings sono sufficienti per domande semplici e direttamente correlate alle linee guida ma non per scenari clinici più complessi, per i quali l’integrazione del Knowledge Graph sembra migliorarne le performance, rendendo il chatbot più robusto e affidabile. Questa tesi si colloca nell’ambito del progetto DHEAL-COM HUB, finanziato dal Ministero della Salute, che mira allo sviluppo di soluzioni di sanità digitale e medicina di prossimità. In questo contesto, l’approccio presentato contribuisce all’avanzamento dello stato dell’arte nell’applicazione di tecniche di Intelligenza Artificiale per il supporto al triage e al monitoraggio di pazienti cronici, ponendo le basi per strumenti digitali sempre più integrati ed efficaci nella pratica clinica.
Utilizzo di Knowledge Graph come strategia di Retrieval-Augmented Generation all’interno di un sistema conversazionale per il supporto alla gestione delle malattie reumatiche
MARINO, SARA
2024/2025
Abstract
The core element of this thesis project lies in the integration of a Knowledge Graph within a Retrieval-Augmented Generation (RAG) architecture for a conversational system supporting the clinical management of rheumatoid arthritis. The aim of the work is the construction and evaluation of a chatbot based on Large Language Models (LLMs), supplemented by information retrieval techniques from validated sources, with the goal of providing reliable and clinically relevant responses. The system is designed for use both by general practitioners, in the triage phase and in referring patients to specialist rheumatology visits, and by patients themselves, for monitoring the course of the disease during periods of clinical remission. During the course of the work, different information retrieval strategies were implemented and compared: the use of embeddings and semantic similarity alone, the use of the Knowledge Graph, and a hybrid solution combining both approaches. The study showed that embeddings are sufficient for simple questions directly related to guidelines but are not adequate for more complex clinical scenarios, for which the integration of the Knowledge Graph appears to improve performance, making the chatbot more robust and reliable. This thesis is part of the DHEAL-COM HUB project, funded by the Italian Ministry of Health, which aims to develop digital health solutions and community-based medicine. In this context, the approach presented contributes to advancing the state of the art in applying Artificial Intelligence techniques for supporting triage and monitoring of chronic patients, laying the foundation for increasingly integrated and effective digital tools in clinical practice.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesi_MarinoSara.pdf
accesso aperto
Dimensione
3.26 MB
Formato
Adobe PDF
|
3.26 MB | Adobe PDF | Visualizza/Apri |
È consentito all'utente scaricare e condividere i documenti disponibili a testo pieno in UNITESI UNIPV nel rispetto della licenza Creative Commons del tipo CC BY NC ND.
Per maggiori informazioni e per verifiche sull'eventuale disponibilità del file scrivere a: unitesi@unipv.it.
https://hdl.handle.net/20.500.14239/33553