JMIR Publications

ABSTRACT

Background:

Conversational voice artificial intelligence (AI) assistants can automate routine postoperative follow-up calls in high-volume, low-complexity pathways such as cataract surgery. However, there are concerns that digital health tools can widen health inequalities if language access and inclusive design are not core considerations.

Objective:

To establish community-defined requirements for an equitable voice AI assistant, by exploring the health care experiences and language-related barriers of Turkish-speaking migrants in the United Kingdom (UK) as an exemplar.

Methods:

We conducted a Patient and Public Involvement (PPI) co-design focus group with 7 Turkish-speaking adults receiving ophthalmic care. The first phase explored participants’ experiences of ophthalmic care in the UK. In the second phase, participants evaluated a pre-recorded prototype Turkish-language telephone call from a voice AI assistant to a Turkish-speaking volunteer. The session was delivered bilingually, recorded with consent, transcribed, and analyzed using reflexive thematic analysis. The voice AI assistant was delivered over the phone with automatic speech recognition and neural text-to-speech, and comprised a dialogue manager layered on a large language model (LLM) to support open-ended conversation constrained to a postoperative review protocol.

Results:

Participants described a recurring communication-workaround pattern in which pathway delays and limited language support led to reliance on family members and loss of privacy. In this context, a language-concordant voice AI for standardized postoperative follow-up was conditionally acceptable provided specific safeguards were met. Priorities included advance notice and clear provenance of calls, robust caller verification, explicit privacy assurances, a standard Turkish accent with clear articulation and slower pace and tolerance for regional dialects, interpersonal warmth and empathy, genuine interactivity for questions and clarification, accessibility for low vision and low literacy, and escalation to clinicians for complex or sensitive issues. These requirements were synthesized into a 10-point checklist covering preparation, verification, confidentiality, clarity and pace, voice qualities, empathy, interactivity, dialect handling, accessibility, and efficiency.

Conclusions:

For patients facing significant language barriers, conversational voice AI may complement existing services when implemented with clear verification, privacy protections, and defined scope with clinician oversight. The co-designed 10-item checklist provides a pragmatic blueprint for safe, empathic, and culturally responsive multilingual deployments that leverage existing telephone infrastructure. These principles will guide the further development of a multilingual voice AI. It will be tested in a forthcoming clinical trial as an automated, telephone-based assistant that makes routine follow-up calls after cataract surgery in 10 languages.