It is maybe the most common translation tool on earth, and at least the most famous one, so is it suited for medical encounters?
Google Translate at its core is a text to text solution, and later on was added with a speech to speech capabilities. When we think of it as “medical interpreter” in medical encounters the speech-to-speech components gets the full focus.
Here are some factors to consider
- Not Tuned to Medical Jargon - Google Translate over the years was evaluated on its text-to-text Accuracy, and in some research was examined against medical written language - the latest study was related to text to text translation of medical instructions in 2021 (here) - for various languages it worked less than well. This is where the English input was inserted as text with 100% accuracy.
In our labs, we also managed to see many errors, in particular in medicine names - and it seems Google Translate is more tuned to everyday language.
- Adding the “Speech Recognition” part adds complexity to Google Translate accuracy.
Now, in our context there is even more expectation from Google Translate - to conduct “oral interpretation” - meaning speech-to-speech translation. This adds even more complexity due to the speech recognition required component - the need to transcribe the source language to text and only later to translate. Speech-to-Text makes even another potential point of failure in accuracy in addition to the errors that may exist in the text-to-text translation. Seems like Google Translate put the “conversation” part as a secondary feature that is not on the main page.
- No errors indication - Google Translate will always “translate” - and there’s no self-proofreading, or any indication of scoring / accuracy. In medical settings, there should be extra layers of carefulness, and even interactive notifications to users when there is less confidence in the results.
- Word by word and not the “message” on “Conversation Mode” - when invoking Google translate for its speech part, it is working simultaneously, and almost immediately produces the translation in the destination language as the user talks - “literal translation” if you will. This leaves almost no room for “convey” or “reconstruct” the broader message in a more flawless & seamless way in the other language.
- There is no patient context considered - Google Translate has no context of its use, the parties involved and the settings. For example in many languages (e.g. Arabic) gender plays a big role in how to say things for example: “how do you feel today” in Arabic would sound different between asking a female patient and a male one - but Google Translate will always translate it to the male phrase, which, naturally, can damage the cultural sensitivity and respect that is so important in the medical encounters. Consider other sociocultural factors Google Translate is a uncontrolled like language formality, patient age that also influence the contextual interpretation
- Google Translate not stating it is HIPAA compliant. In fact, it’s the other way around - Google emphasized that they use the data they collect Google Translate to “Improve our machine learning models”. With over 1 billion people who use Google Translate - this seems a main source of data for Google. Hence - health providers need to assume that every PHI data potentially is going to be repurposed to Google needs, eg for future model training or other internal needs.
However - in our conversations with providers while building No Barrier, many of them acknowledged that they found themselves in situations where they use Google Translate, even though they are aware that this is not allowed. The situation usually was the last resort - for example ER reception at 04:00 am where no interpreter was available and there was a need to conduct intake.
In No Barrier we are building the first AI Medical Interpreter that would address the various considerations, among them medical accuracy, HIPAA and context awareness for the medical settings.
