Digital Twin (DT) systems offer promising opportunities in healthcare by enabling continuous monitoring and personalized analysis of physiological data. Yet, for patients, the outputs of these systems often remain difficult to interpret due to technical complexity, lack of contextualization, and limited transparency. This lack of interpretability can undermine patient engagement, trust, and ultimately the effectiveness of DT systems in supporting self-management and communication with care providers.
This thesis explores how Large Language Model (LLM)-based conversational interfaces can function as interpretation layers within DT architectures, translating structured physiological data into understandable, patient-facing insights. Focusing on heart rate as a representative signal, we investigate how Retrieval-Augmented Generation (RAG) techniques can enhance the clarity, contextual relevance, and trustworthiness of responses in patient-facing interactions. To do so, we developed a local prototype that combines rule-based heart rate categorization with a retrieval-enabled LLM (LLaMA 3.2), operating on simulated datasets to preserve privacy. The system generates contextual, comprehensible answers without offering clinical advice, using hourly summaries of heart rate logs as input.
A mixed-methods evaluation involving nine participants was conducted to assess both technical performance and user experience. Technical accuracy was measured using RAGAs metrics, including faithfulness, answer relevance, and context recall, while user experience was evaluated through semi-structured interviews and standardized usability questionnaires. Results indicate that the system produced relevant, understandable, and empathetic responses, helping participants identify trends in their simulated heart data. However, broader queries led to decreased response consistency, and the use of simulated rather than real patient data limited the generalizability of findings.
The main contribution of this thesis lies in embedding RAG within a DT framework to enable interpretable patient communication of heart rate data. This approach addresses key challenges in hallucination reduction and contextualization, providing a foundation for future systems that aim to translate complex physiological signals, starting with heart rate, into accessible and trustworthy explanations for end users.