Co-Design and Creation of a Documentation Support System for Care Workers with Large Language Models

Abstract Developing meaningful and wanted technologies for the care context presents a significant challenge. This study illustrates the development of a documentation support prototype created with continuous involvement of care stakeholders. The process is examined in terms of its strengths and weaknesses, and key design requirements are identified. We employ a co-design approach involving 43 unique care stakeholders (24 care workers, 19 residents) across four study phases conducted in Vienna residential care homes. Each phase incorporates divergent and convergent methods to identify challenges and define solutions. Four mock speech interactions were recorded for feasibility testing, followed by 40 recordings of real care interactions (avg. ca. 11 min). Qualitative data from interviews with care workers is analyzed using thematic analysis, and system performance was evaluated using ROUGE-1, BERTScore, and care worker annotations. Based on the recordings and interview data, we identify topics in care worker–resident dialogues that are relevant or irrelevant to documentation. We found that the system consistently generates relevant information for documentation (55% on average). Of the labeled topics, 29.6% were identified as not relevant for documentation. We identify transcription errors and lack of speaker diarization for part of the reason for the remaining 15.4% of summaries consisting of wrong segments. We present the design outcomes of a multi-phase co-design study. Our findings demonstrate that the system can extract care-relevant information from care interactions, and we discuss insights related to the co-design process, as well as potential implications for work practices and privacy.