Key points are not available for this paper at this time.
Medical chatbots are becoming essential components of telemedicine applications as tools to assist patients in the self-management of their conditions. This trend is particularly driven by advancements in natural language processing techniques with pre-trained language models (LMs). However, the integration of LMs into clinical environments faces challenges related to reliability and privacy concerns. This study seeks to address these issues by exploiting a privacy by design architectural solution that utilises the fully local deployment of open-source LMs. Specifically, to mitigate any risk of information leakage, we focus on evaluating the performance of open-source language models (SLMs) that can be deployed on personal devices, such as smartphones or laptops, without stringent hardware requirements. We assess the effectiveness of this solution adopting hypertension management as a case study. Models are evaluated across various tasks, including intent recognition and empathetic conversation, using Gemini Pro 1.5 as a benchmark. The results indicate that, for certain tasks such as intent recognition, Gemini outperforms other models. However, by employing the “large language model (LLM) as a judge” approach for semantic evaluation of response correctness, we found several models that demonstrate a close alignment with the ground truth. In conclusion, this study highlights the potential of locally deployed SLM as components of medical chatbots, while addressing critical concerns related to privacy and reliability. • Exploration of integrating language models into medical chatbots. • Comprehensive examination of challenges including reliability and privacy concerns. • Insights into the complexities of balancing data privacy with effective language model-based healthcare solutions in practical resource-limited settings. • Adoption of open-source small language models and experimental evaluation of their performances. • Focus on self-management of chronic hypertension.
Magnini et al. (Wed,) studied this question.