Human-AI Co-Research on Design and Evaluation of Embodied Conversational Agent in Rehabilitation Contexts

Abstract

Despite strong evidence that adequate, repetitive home‑based rehabilitation improves functional recovery and independence, current rehabilitation delivery models still show significant gaps and limitations. Accordingly, we propose an AI-driven Embodied Conversational Agent (ECA) as a personal assistant to enhance continuity of care at home. It delivers natural‑language support on prescribed home exercises to reinforce experience‑dependent neuroplasticity, clarifies therapeutic principles and provides emotional and motivational support to sustain engagement. ECA is based on Large Language Models (LLMs) and may take virtual or physical form. However, physically instantiating ECAs into robots remains challenging, since many robots lack native support for key real-time interaction capabilities such as speech processing, gesture execution and attention tracking, while current LLMs can also be unreliable due to factual errors and inconsistency. These limitations restrict their direct deployment in clinical and rehabilitation settings, while early-stage ECA design is further constrained by restricted access to real users due to practical and ethical considerations. To expand the design space and accelerate iterative design refinement under constraints on direct human testing, we propose a novel Design-Based Research methodology for human-AI co-design and evaluation of ECA (co-AI DBR). Generative AI acts as a partner in the iterative knowledge building by supporting cycles of design, testing and refinement, enabled through the interplay of synthetic patient generation and real-code execution for simulation, emulation and evaluation of the ECA platform and its LLM-based conversational pipeline. To validate the methodology in a post-stroke rehabilitation context, a virtual ECA was first tested with virtual patients to assess the technical implementation, the accuracy of LLM-generated responses and their ability to provide therapeutic explanations, emotional and motivational support. Then a pilot deployment with the Furhat robot as ECA, using concerns raised by real participants (patient relatives and professionals), was conducted to further evaluate the ECA voice interface and augmented communication. LLM responses to questions from the participants showed higher lexical diversity (MTLD ≈ 134 vs. 93.9) and lower repetition (Yule's K ≈ 66.8 vs. 115.4) than responses to synthetically generated questions, while remaining fully factually consistent with zero