Evaluating the Feasibility and Utility of Machine Translation for Radiation Therapy Patient Education Materials.

Patients with limited English proficiency have increased healthcare challenges. Providing patient education in the patient's language has the potential to improve health outcomes, reduce system costs, and improve patient satisfaction. Healthcare providers remain apprehensive regarding the contextual accuracy of neural machine translation to assist the efforts of human translators and interpreters. Google Translate (GT) now uses deep learning techniques to translate whole sentences, which is more sensitive to context and grammar. The purpose of this research was to investigate the feasibility and utility of using machine translation (Google Translate) to translate radiation therapy patient education materials. Five Radiation Therapy patient education pamphlets were chosen for translation based on their high usage and importance at our institution. The pamphlets chosen for the study were: 1) Taking care of your skin during radiation therapy; 2) What to expect when getting radiation therapy for prostate cancer; 3) What to expect when getting radiation therapy to the abdomen and pelvis; 4) What to expect when getting radiation therapy to the head and neck; 5) What to expect when getting palliative radiation therapy to the lungs. Each pamphlet was translated from English into the following five languages: Vietnamese, Punjabi, Simplified Chinese, Portuguese, and Spanish, using both GT and professional human translators. These languages were chosen based on data on the most common language needs in the local hospital network. An additional group of professional human translators (3 per language) conducted an independent, blinded review of the translated pamphlets, evaluating the domains of fluency, adequacy, meaning, and severity. When comparing human and machine-translated material scores, humans scored higher in every domain, with t-tests determining the differences to be significant (p=<0.001). Total scores for machine-translated materials ranged from 13.33-15.27 out of a possible 20 for GT and 14.5-17.28 for human-translated material. In the most sensitive domain of severity, which identifies the resultant risk to patient outcomes based on mistranslation, machine scores ranged from 3.27-4.40 out of a possible 5 compared to 3.60-4.93 for human translations. Analysis of variance (ANOVA) found significant differences in scores between languages with Spanish and Vietnamese regularly scoring better in both human and machine translated versions. This research suggests that using machine translation tools such as Google Translate to translate whole sentences in radiation therapy pamphlets from English to 5 common languages is feasible. While there remain significant differences in quality between human and machine translation, it appears that the latter does not pose a significant risk of adversely affecting patient health outcomes, potentially opening the path for its limited use in patient education. [ABSTRACT FROM AUTHOR]