Pilot Project 1: AI Heals (Health Advances Through Language Solutions)?

City College Of New York, New York NY

Investigators

Linked publications & trials

Abstract

Among patients who speak English less than “very well”, cancer outcomes are often less favorable, in part due to lacking professional medical interpretation services. Technology promises efficient, scalable remote interpreting solutions to bridge these challenges. However, there is no evidence-based gold standard for technology- based interpreting. Two technology-based, people-rendered methods used for remote interpreting are 1) Remote Consecutive Medical Interpreting (RCMI; “audio consecutive”), the most commonly utilized, and 2) Remote Simultaneous Medical Interpreting (RSMI), “UN-style” simultaneous interpreting applied to the medical encounter, which may closely approximate a same language encounter, decrease interpreting errors, and improve outcomes. Further, with artificial intelligence (AI) solutions, there is potential for less expensive, more scalable interpreting services delivery in the form of AI Simultaneous Medical Interpreting (AISMI). We will compare RCMI (audio consecutive) (Arm 1), versus RSMI (UN style) (Arm 2) and versus AISMI (AI UN style) (Arm 3) with actors who speak various languages playing patients who speak English less than very well and playing English-speaking providers, to determine comparative error rates of clinical significance and efficiency. We will write 50 English scripts incorporating content reflecting actual medical oncology encounters. The patient portion of the scripts will be translated into various languages. We will have a pool of 20 bilingual interpreters. Additionally, we will train a commercially available AI system to provide AISMI for speakers of various languages. Each script will be human-interpreted twice, once via RSMI and once via RCMI, and by the AISMI system. There will be 150 interpreted simulations per language (appointments acted from the scripts with the addition of unscripted live interpreting) that will be audio-recorded and transcribed (50 for each type of interpreting). We will determine differences in interpreting error rates of clinical significance (primary outcome) and in interpreting efficiency of utterance (secondary outcome). Additionally, we will investigate the potential acceptability of AISMI with surveys among patients who speak English less than very well and clinicians and administrators. 26 million people in the U.S. have limited English proficiency, speaking English less than “very well”; lack of language interpretation services has been linked to poorer cancer outcomes among patients with limited English proficiency. There is little information available on technology-based solutions to providing medical interpretation services. We will compare the error rates and efficiency of technology-enabled interpretation methods, including remote simultaneous interpretation (UN style) delivered by artificial intelligence.