Improving In-person Interpreter Utilization in Complex Care: Findings from a Stepped-Wedge Cluster Randomized Trial of Integrated Artificial Intelligence

Comprehensive Summary

Barwise et al. performed a stepped-wedge randomized trial to determine whether a machine learning (ML) algorithm integrated with informatics would increase the use of in-person language interpreters among patients with non-English language preferences (NELP) and reduce time to first in-person interpreter visit in complex medical cases. From May 2023 to June 2024, 35 units at Mayo Clinic Rochester were randomized and transitioned between phases at 60-day intervals. Eligible participants were acute care inpatients ≥18 years with NELP and a complexity score ≥3. Exclusion criteria included patients who were non-verbal, used sign language, or whose records were confidential. The ML algorithm, implemented through software called "Tower Control," calculated a complexity score based on electronic health record (EHR) data, including length of stay, procedures, level of care, clinical notes, events, and need for palliative care. The score ranged from 0 (least complex) to 9 (most complex). The control group included 749 unique cases receiving standard of care where providers requested in-person interpretation services without reminders. The intervention group included 672 unique cases where the research team generated a daily list of NELP patients and notified language services staff, who then prompted bedside nurses via a secure chat "nudge." Overall, 42% of eligible participants received in-person interpretation (44% intervention vs. 41% control). The intervention showed a trend toward increased odds for receiving services (OR = 1.4; p = 0.061). Among those who received an interpreter (n = 542), time to first interpreter visit did not significantly differ between groups (p = 0.87).

Outcomes and Implications

This study is the first to employ AI to promote interpretation services. Despite not reaching statistical significance, the trend toward improved interpreter utilization (OR = 1.4) suggests potential clinical benefit and supports the need for a larger multicenter trial. The authors note that the need for interpreters may have exceeded available resources as more units were integrated into the intervention group, potentially explaining the lack of significant findings. Active outreach serves as a nudge to remind busy clinicians about the importance of interpretation services. The algorithm also supports the interpretation services personnel by helping them anticipate demand and prioritize complex patients during staffing shortages. Due to the positive impact on workflow, the language services team continues to use the intervention. EHR-based nudges represent a practical approach to mitigating underuse of interpreter services in busy clinical settings. Important limitations include the lack of patient outcome data, exclusion of remote interpretation services, and omission of social determinants of health in the complexity score calculation.