In this paper, we study the capabilities of large language models (LLMs) to adapt a concert moderation to diverse expertise levels of listeners. Our proof-of-concept concert moderator is based on retrieval-augmented generation (RAG) and uses few-shot audience modelling to infer listener’s expertise. We study the capabilities of the system to adapt to three different listener’s expertise levels. Two open domain LLMs are compared: gpt-oss:20b and llama3. The recognised differences among the models suggest that they vary in how directly they reproduce versus paraphrase retrieved information while maintaining semantic alignment.

Conversational agents (CAs), such as medical interview assistants, are increasingly used in healthcare settings due to their potential for intuitive user interaction. Ensuring the inclusivity of these systems is critical to provide equitable and effective digital health support. However, the underlying technology, models and data can foster inequalities and exclude certain individuals. This paper explores key principles of inclusivity in patient-oriented language processing (POLP) for healthcare CAs to improve accessibility, cultural sensitivity, and fairness in patient interactions. We will outline, how considering the six facets of inclusive Artificial Intelligence (AI) will shape POLP within healthcare CA. Key considerations include leveraging diverse datasets, incorporating gender-neutral and inclusive language, supporting varying levels of health literacy, and ensuring culturally relevant communication. To address these issues, future research in POLP should focus on optimizing conversation structure, enhancing the adaptability of CAs’ language and content, integrating cultural awareness, improving explainability, managing cognitive load, and addressing bias and fairness concerns.

Ensuring equitable access to digital therapeutics (DTx) is essential to avoid healthcare inequalities in an era of increasing digitization. This requires DTx to be tested with users from diverse populations, which is often not realistic due to time and resource constraints. In this paper, we propose the use of large language models (LLMs) to simulate diverse patients. Specifically, we manually create a patient vignette that characterizes a specific population group. Variations of this vignette are used for role-prompting a commercial LLM, GPT-4, instructing the LLM to take on the role described in the patient vignette and act accordingly. We investigate if the LLM stays in its given role. To do this, we simulate a medical anamnesis interview with the role-prompted LLM and analyze its responses for compliance, coherence, correctness, containment, and clarification. Our results show that GPT-4 generates compliant, coherent and clinically valid responses, including information that is not explicitly stated in the provided patient vignette.

In healthcare, agency refers to the ability of patients to actively participate in and control their health through collaborating with providers, informed decision-making and understanding health information. Conversational agents (CAs) are increasingly used for realizing digital health interventions, but it is still unclear how they are enhancing patient agency. This paper explores which technological components are required to enable CAs impacting on patient agency, and identifies metrics for measuring and evaluating this impact. We do this by drawing on existing work related to developing and evaluating healthcare CAs and through analysis of a concrete example of a CA. As a result, we identify five main areas where CAs enhance patient agency, namely by: improved access to health information, personalized advice, increased engagement, emotional support and reduced barriers to care. For each of these areas, specific technological functions have to be integrated into CAs such as sentiment and emotion analysis methods that allow a CA to support emotionally.

Textual entailment has been recognized as a generic task that captures major semantic inference needs across many natural language processing applications. However, to date, textual entailment has not been considered in a cross-corpus setting, nor for user generated content. Given the emergence of Medicine 2.0, medical blogs are becoming an increasingly accepted source of information. However, given the characteristics of blogs( which tend to be noisy and informal; or contain a interspersing of subjective and factual sentences) a potentially large amount of irrelevant information may be present. Given the potential noise, the overarching problem with respect to information extraction from social media is achieving the correct level of sentence filtering - as opposed to document or blog post level. Specifically for the task of medical intelligence gathering. In this paper, we propose an approach to textual entailment with uses the text from one source of user generated content (T text) for sentence-level filtering within a new and less amenable one (H text), when the underlying domain, tasks or semantic information is the same, or overlaps.