Translation approaches to support systemic anti-cancer therapy consent for individuals with limited English proficiency

Language discordance between patients and clinicians is common. In the UK, over one million people report limited English proficiency (LEP) [1]. LEP is a key social determinant of cancer incidence, prevention and control [2‐4], which may reflect both the association of LEP status with other markers of social disadvantage, and the direct effects of language discordance on health communication [2, 5]. Poor healthcare communication with LEP patients leads to increased healthcare costs [1] because of delayed diagnoses and increased healthcare utilisation, as demonstrated among patients with multiple myeloma [6]. The absence of suitable translation and interpreting services also decreases treatment adherence among LEP populations with cancer [7, 8], reduces their participation in oncological clinical trials [9, 10] and limits their participation in research [10].

Language discordance creates additional challenges when obtaining informed consent for systemic anticancer therapy (SACT)—a complex, risk-laden and technical therapeutic approach within which understanding treatment intent is critical [11]. Whereas guidelines exist for translation of written participant information and consent forms in a clinical trial setting [12], there is no such guidance for routine oncology practice. Conventional documentation in routine oncology consultations, including written information material and structured SACT consent forms, is often provided in English only [13], with reliance upon verbal interpretation for communication with LEP patients.

Unsupervised machine translation, defined as the use of computerised algorithms to generate translations without human validation, is widely used by healthcare professionals, health organisations and patient charities [14], such as through integration of the neural machine translation tool Google Translate into accessibility software (such as ReciteMe) [15], online triage tools and patient information apps. For high-resource languages with a large linguistic corpus to train machine translation tools, machine translation has shown promise. For example, a recent study showed high accuracy in translating paediatric instructions into Spanish using Chat-GPT (Open AI), but emphasised that “[t]ranslation performance is language specific […] results cannot be generalized to other languages” [16].

These technologies increase access to information resources for LEP patients in their preferred language [16, 17], but their use must assess clinical risks in oncological settings [18]. Best practices are for machine translation outputs to be reviewed by professional translators [19], but this is often omitted in real-world situations [14]. Furthermore, existing research assesses the accuracy of machine translations, rather than their impact on translation users [17, 20]. Specifically, the impact of different translation approaches on comprehension in a SACT consent context has not been formally evaluated.

To address these unanswered questions, our study had two co-primary aims: (1) to compare the effectiveness and quality of professionally translated and unsupervised machine-translated versions of SACT patient information in a low-resource language (i.e. a language with a limited linguistic corpus to train machine translation tools), and (2) to assess the impact of providing a professionally translated English–Bengali bilingual consent form on both objective comprehension and self-rated understanding during an interpreted SACT consent consultation. We chose myeloma as a case study, given its relative unfamiliarity and non-curative treatment intent. We studied Bengali- and Sylheti-speaking communities in the UK because of their high rates of LEP and poorer health outcomes compared to the wider UK population [21].

Between 23rd April and 8th May 2025, 123 individuals were recruited and randomised during the Booklet Component, of whom 92 proceeded to the Consent Component. All participants received their allocated intervention. Flow of participants through the study is summarised in Fig. 1.

Bild vergrößern

This study aimed to assess the impact of different Bengali translation approaches for SACT patient information booklets and consent forms. In the Booklet Component, we observed a higher rate of critical errors introduced by machine translation compared to professional translation, but that only 16% of participants across both translation arms correctly understood treatment intent after reading. In the Consent Component, we observed significantly higher rates of both measured comprehension of treatment intent and subjective confidence in understanding, in participants receiving a bilingual consent form.

The role of unsupervised machine translation in healthcare remains contentious. Without human review, machine translations can create clinical risks, which vary by language [17]. A recent study comparing neural machine translation tools to professional translations in the translation of discharge summaries showed equivalent quality for Spanish (a high-resource language), but inferior quality for Chinese, Vietnamese and Somali translations (lower resource languages) [20]. Our results support caution in the use of unsupervised machine translation tools, especially in low-resource languages such as Bengali. If healthcare organisations do provide unsupervised machine translations through accessibility tools, they should be aware of risks and should provide disclaimers in the target language [32]. Producers of patient information can also maximise “translation-friendliness”, for example by avoiding idioms and complex sentences, acknowledging that other users may access the material via machine translation [32]. Many computational linguists and clinical researchers suggest that the next step for machine translation (including large language models) should be human-centred machine translation [33].

Despite the clear difference in translation quality in the two arms of the Booklet Component, we observed a low level of comprehension in participants after reading translated written patient information, whichever translation method (machine or professional) they were randomised to. This is despite choosing a booklet designed for accessibility and readability, suggesting that low health literacy or other communication factors may be additional barriers beyond the availability of translated resources. Other formats of patient information, such as audio and video, may be more effective in some linguistic groups and deserve further study.

Participants’ understanding of treatment intent improved significantly in both arms of the Consent Component compared to the Booklet Component, likely reflecting cumulative learning and the additional benefit of an interpreted verbal consent consultation. The improvement was substantially greater in the group receiving the English–Bengali bilingual SACT consent forms, who demonstrated significantly higher comprehension and confidence than participants provided with English-only forms. Two mechanisms may explain the improved comprehension and confidence seen with the bilingual consent form. First, in addition to hearing the interpreted consultation, participants could directly read the information in Bengali. Second, interpreters could read aloud Bengali translations of specialist terminology, rather than interpreting these terms spontaneously. This latter benefit may be even more marked if less experienced interpreters are interpreting treatment consent consultations. However, the combination of written translation and verbal interpretation was not universally effective: within the bilingual consent form group, 40% of participants did not correctly understand the intent of treatment. Further investigations are necessary to ascertain whether using accessible spoken Bengali, normally understood by Sylheti speakers, contributed to the limited understanding of the intent of treatment for the bilingual consent group.

Our study has several limitations. First, we recruited healthy volunteers rather than patients with cancer. Patients face additional psychological and cognitive demands which we cannot account for in our study, though such burdens would likely make reliable translations even more valuable. The salience of treatment intent would also be greater in a patient with cancer than a healthy volunteer. Second, our simulation had partial fidelity to real-life clinical practice, with small groups of participants undertaking consent consultations rather than individuals. Third, we measured comprehension and confidence immediately after each intervention, rather than looking at recall later. Fourth, we did not include a formal measurement of health literacy in Bengali [34]; this limits our ability to assess the contribution of health literacy to our study endpoints. Fifth, our study was exploratory and not definitively powered to our objectives, resulting in wide confidence intervals for effect sizes. Finally, machine translation of Bengali is likely to perform worse compared to languages where machine translation tools have had a larger linguistic corpus for training (e.g. French, Spanish, German), and poor performance of machine translation may not apply across other language groups.

We recommend adopting bilingual interlinear consent forms as a cost-effective and relatively simple measure to improve communication and health equity for LEP patients. Bilingual consent forms may be effective in other language-discordant health contexts beyond SACT, including clinical trials, especially when personalisation can be achieved without writing extensive text (e.g. through ticking relevant boxes). Further work is needed to assess the role of machine translation in healthcare communication, but current use should be cautious. Finally, other methods beyond written booklets should be explored to provide health information for LEP patients.