Five Questions about Machine Translation, with Elaine Khoong, MD, MS

HEY GOOGLE, WHAT’S RUSSIAN FOR TAKE TWO PILLS AND CALL ME IN THE MORNING?

For over 25 million Americans who prefer a language other than English, language-related communication barriers in healthcare encounters are increasingly common. Can machine translation fill in when interpreters and translators are unavailable?

UCSF Action Research Center Associate faculty Dr. Elaine Khoong, MD, MS, shares her insight into the potential roles of AI to bridge language gaps between healthcare providers and patients.

 

Why is machine translation (MT) an important potential tool in healthcare? 

Worldwide, over 281 million people live in a country not of their birth. In the US, the number is at least 25 million. For all these people, language barriers in healthcare encounters may impede high quality care and impact health. 

In the United States, studies have consistently shown that patients with a non-English language preference rarely receive written patient instructions in their preferred language, especially when time-sensitive translations are needed. So, machine translation may help mitigate gaps in access to professional translations. 

 

How often are professional (human) translators not available in healthcare scenarios? 

Studies have shown that professional translations of care plans are available in less than 30% of outpatient settings. One study showed only 8% of patients with a non-English language preference received written discharge instructions in their preferred language. 

Quite frequently these types of patient-specific instructions are provided only in English. So, although lots of studies compare MT to human translations, in the real world, the risk of inaccurate MT should be compared to the risk of only providing English instructions. 

 

What are the clinical contexts in which machine translation could improve healthcare?   

Use of professional translation is uncommon, particularly in scenarios where patient-specific time sensitive written communication is needed (e.g., patient instructions after a visit, patient portal messages). While we continue to endorse working with professional translators when possible, machine translation may help bridge gaps when translators are unavailable. 

Specific to the findings in our study, we found that GPT had high rates of accuracy for English to Spanish patient-facing written communication. Importantly, we found that in this context, inaccurate translations were very unlikely to be harmful. This is consistent with other studies of large language models (like GPT). In our paper, we recommend that MT could be considered for low-risk communication from English to Spanish. 

 

What are some factors people in healthcare should consider before using machine translations? 

First, transparency is key--both for the patient and the clinician. Everyone has different risk tolerance. Both clinicians and patients need to know the risk of error and make an informed decision about this. There is a concept of machine translation literacy. If we see MT used more in healthcare, it's increasingly vital that clinicians have MT (and AI) literacy. 

Next, there are regulatory issues and policies to consider. There are some circumstances that don't allow MT use based on the section 1557 revised ruling, within the United States. I have seen some flexibility with how people interpret those rules, but right now MT use is not allowed for critical communication. 

Also, there are case-by-case instances that involve specific clinical context and content-- What are alternative options for translation? How important is communication? What additional strategies are being concurrently used to improve patient understanding? All these considerations matter.  

It’s also important to remember that MT does not need to be used in isolation. It can be used as a first pass before undergoing human review. 

 

How do these findings have the potential to change the quality of care for patients? And how do they signal a potential shift in the responsibilities of healthcare professionals? 

These findings support the growing literature that machine translation, and particularly LLMs, have value for English to Spanish written translations for content that is patient-facing and not too technical. We do not have the studies to demonstrate any impact on the quality of care, but I am hopeful that we are moving towards a time where a more formal evaluation of the impact on MT on outcomes (whether patient-reported, clinical, or utilization) is possible.   

I do not think machine translation shifts responsibilities of healthcare professionals, but I think they increasingly put pressure on health systems and regulators to provide more guidance on if and when these tools can and perhaps even should be used. I think we are moving more towards a situation where not adopting machine translation is going to cause more harm by inhibiting access than the harm of adopting machine translation in low-risk clinical scenarios. 

 

 

Find the original research, Evaluation of the accuracy and safety of machine translation of patient-specific discharge instructions: a comparative analysis, at BMJ Quality & Safety.