Clinical Research Scientist Seattle Children's Fairbanks, Alaska, United States
Background: Patients and families who use a language other than English (LOE) for medical care have limited access to healthcare team discussions on family-centered rounds (FCR) compared to those who are English-speaking. Artificial intelligence (AI) generated speech-to-text transcription and translation technologies have the potential to improve language access for families who use a LOE on FCR. Objective: 1) To evaluate the accuracy of AI-translated captions, with and without human supervision of the English speech-to-text, into Spanish, Arabic, and Vietnamese for FCR; and 2) to determine the acceptability and feasibility of AI-translated captions of FCR for families who use a LOE. Design/Methods: Parents of inpatients who use a LOE consented to audio-recording of their child’s FCR. FCR recordings were replayed through spf.io, an AI platform, with and without a human supervising the English speech-to-text transcription, while translation into Spanish, Arabic, and Vietnamese occurred. AI-only, human-supervised AI, and fully edited transcripts and translations were scored sentence-by-sentence on a 5-point Likert scale for audio quality, fluency, adequacy, meaning, and error severity scores.
Medical interpreters and parents of inpatients who spoke a LOE viewed a demonstration video about computer-translated captions on FCR and were interviewed. Topics included the acceptability and feasibility of this intervention, delivered with a consecutively interpreted summary at the end of FCR. Inductive content analysis of interview transcripts was performed guided by the Consolidated Framework for Implementation Research. Results: FCR recordings for patient families who used Spanish (n=5) or Arabic (n=1) for care were collected and processed through spf.io. Better audio quality was positively associated with improved fluency, adequacy, meaning, and severity scores. Human-supervised AI scores were higher than AI-only scores in the domains of fluency (4.68 vs. 3.88), adequacy (4.74 vs. 3.88), meaning (4.74 vs. 3.81), and severity (4.94 vs. 4.19).
Five Spanish medical interpreters, 5 Spanish-speaking parents, and 1 Arabic-speaking parent participated in interviews. Five out of 6 parents were interested in AI-translated captions on FCR. All 5 interpreters identified scenarios in which AI-translated captions of FCR could be useful, but all also had reservations.
Conclusion(s): AI-translated captions, particularly with human-supervision, may be a useful tool to improve language access, enhance understanding, and promote participation for patients and families who use a LOE on FCR.
