The digital transformation and language barriers

Cardiovascular medicine is undergoing a digital transformation, with increasing reliance on technologies such as telehealth, home-based cardiac rehabilitation, wearable devices for remote monitoring, and smartphone applications to deliver patient care1. However, the evidence base supporting these digital innovations is limited: most trials and studies are conducted exclusively in English2.

This language barrier has significant implications for healthcare delivery and outcomes in the diverse population of the United States. As of 2020, over 67 million United States residents speak a language other than English, with 62% identifying Spanish as their primary language3. Within this Spanish-speaking population, more than 15.9 million individuals have limited English language proficiency3, making them dependent on Spanish translation for daily life and medical care4. As a result, emergency visit times are frequently prolonged and more frequent, post-discharge care is suboptimal, and there are higher readmission rates for individuals with limited English language proficiency5. Unfortunately, nearly one-third of healthcare systems fail to provide translation services or linguistically tailored health information6. Even when such services exist, professional translators are often unable to meet high volume and rapid turnarounds common in clinical practice7. Moreover, Hispanic/Latinx patients with limited English language proficiency often report dissatisfaction with both care quality and interpretation services received8.

Cardiovascular health differences and digital engagement in Hispanic/Latinx communities

These language barriers are particularly concerning given the cardiovascular health differences affecting Hispanic/Latinx communities. This population experiences higher rates of cardiovascular risk factors9, with recent data showing worsening in population-level prevalence and incidence of hypertension and diabetes10. Paradoxically, this is also a population that could greatly benefit from digital health innovations: over 91% of Hispanic/Latinx adults in the United States own smartphones, and compared to other racial/ethnic groups, they are more likely to use these devices to access health information11.

Despite this technological engagement, Hispanic/Latinx individuals with limited English language proficiency remain largely excluded from digital health research2. A recent analysis of digital health trials registered on ClinicalTrials.gov found that 51% explicitly required English language proficiency for participation, which is likely an underestimate2. This systematic exclusion may worsen the digital divide and prevent populations who might benefit the most from cardiovascular technological innovations from participating in the research that validates these tools and informs clinical decision-making12.

Large language models as translation tools

Large language models (LLMs) offer a promising solution to improve Hispanic/Latinx representation in digital health trials. While machine translation tools like Google Translate have existed for years, their medical applications have been limited by accuracy and usability/readability13. In contrast, newer LLMs such as ChatGPT, GEMINI, and DeepSeek—trained on vastly larger datasets and able to efficiently batch process large volumes of materials14—are positioned as potentially superior tools for English-to-Spanish translation in future digital health trials.

Emerging evidence supports the efficacy of LLMs in healthcare translation. Studies have demonstrated that LLM-translated materials (from English to Spanish) perform comparably to professional medical translators in medical accuracy and Spanish fluency for discharge instructions15 and post-operative care guidelines16. Additionally, LLMs have successfully translated kidney transplantation information from English to Spanish, achieving high scores for both medical accuracy and cultural sensitivity appropriate for Hispanic/Latinx populations17. These data suggest that LLMs could serve as valuable supplements to professional medical translators for health-related information, while also offering the unique ability of LLMs to provide conversational, reading-level appropriate, and culturally resonant translations18.

Limitations of large language model-mediated translation

Some limitations of LLMs should be considered. First, the risk of inaccuracy and “hallucination” (or the generation of false or misleading information) remains a valid concern with LLM use14. This risk can be substantially mitigated by restricting LLM use to direct English to Spanish translation tasks rather than content generation, leveraging their linguistic capabilities while avoiding their generative uncertainties. With this restriction, studies focusing specifically on health-related English to Spanish translations have highlighted the high accuracy rates (ranging in accuracy of translation from 4.70 to 4.94 out of 515,17 and scoring 100% for accuracy in 6/7 procedure-related discharge summaries16). Second, while translations may be technically accurate, they might lack clinical relevance or cultural appropriateness. However, recent iterations of LLMs pre-trained on a larger corpus of non-English language data appear to have produced more sophisticated LLMs that are linguistically adept and more in tune with cultural and contextual aspects of Spanish language17. Third, gender bias in LLMs is an ongoing challenge19, with analyses finding that females were frequently described with body-related adjectives and males with behavior-related descriptors20. Some investigators have addressed these concerns through careful prompt engineering17, though this remains an area requiring continued attention. Fourth, current LLMs lack the nuance to accommodate regional Spanish dialects (e.g., Argentinian versus Mexican Spanish), potentially limiting their effectiveness across diverse Hispanic/Latinx communities. Finally, the absence of regulatory oversight poses a practical barrier. Unlike approved medical devices, LLMs operate without formal healthcare regulation13. This regulatory gap creates a disconnect: while patients report greater satisfaction with LLM translations, clinicians remain hesitant to rely on them in health-related settings13.

A proposed hybrid workflow

Balancing the advantages and limitations of current LLM technology, we propose a hybrid workflow to increase Hispanic/Latinx participation in digital health trials while maintaining translation quality and cultural sensitivity (see Fig. 1). Our proposed workflow operates in two phases: first, LLMs would automatically translate all text-based trial materials—including consent forms, procedural instructions, and other patient information—from English to Spanish. Second, professional medical translators would review these automated translations for accuracy, cultural nuance, and clinical appropriateness. This step of human oversight is essential, as LLMs are not approved or regulated by a regulatory body and hence require further supervision at this stage and time. This hybrid approach strategically leverages the strength of LLMs (rapid, high-accuracy initial translations), while human experts ensure quality control. The workflow addresses key limitations of standalone LLM translation—potential inaccuracies, cultural blind spots, and regulatory concerns21—through professional oversight, while dramatically reducing the time and cost barriers that currently prevent many digital health trials from including Spanish-language arms. The efficiency gains would allow professional translators to focus on reviewing and refining existing translations (rather than translating materials from scratch), significantly reducing both project timelines and costs. It should be noted that this hybrid workflow is dependent on the availability of trained medical interpreters, who may not be available in all healthcare-related settings6. Moreover, given regional variations in Spanish (i.e., differences in vocabulary, idioms, etc.) and current limitations in LLMs, this workflow will be reliant on trained medical interpreters familiar with these language nuances to ensure that there is cultural alignment with the diverse population of Spanish-preferred-speaking individuals. However, with potential remote/fully-digital implementation, this approach could make Spanish language trial participation feasible for a broader range of research studies and potentially expand access to cutting-edge cardiovascular digital health innovations for Hispanic/Latinx communities who have been largely excluded from such trials and studies to date2.

Fig. 1: Integration of large language models to improve participation of non-English language fluent individuals in digital health trials.
figure 1

This figure highlights the proposed hybrid workflow, leveraging both large language models and human expertise, which we believe will increase participation of non-English language fluent individuals in digital health trials. We also highlight emerging capabilities of large language models for future implementation in digital health trials. LLM – large language model. Created in BioRender. Kim, D. (2025) https://BioRender.com/ybmpmh3.

Full size image

Future applications of large language models in digital health trials

Beyond their translation abilities, LLMs offer several emerging functionalities that could enhance digital health trials for Hispanic/Latinx (and other non-English language fluent) populations (see Fig. 1). First, LLMs can potentially be used to create audio content via text-to-speech (e.g., verbally conveying digital trial materials or information)22. This could make digital trial materials accessible to participants with low literacy or limited written language proficiency. Additionally, the generative properties of LLMs could enable the creation of visual aids and simplified representations of complex documents, such as informed consent forms, potentially improving patient comprehension across educational backgrounds. However, such use of LLMs would have to be closely monitored and regulated given the potential for misuse and hallucination. The reverse capability of LLMs (i.e., speech-to-text)23 could also be used to allow for real-time patient-reported outcome collection in patient’s native languages, preserving important linguistic nuance that may be lost in traditional translation processes. This capability could enable more authentic and comprehensive data collection from diverse populations. Finally, looking toward more distant future applications (and requiring much more generation of evidence at the foundational level), agentic AI systems could function as virtual health navigators24, providing motivational support in both a culturally and linguistically sensitive fashion, potentially improving both the enrollment and retention of patients in digital health trials.

Conclusions

In summary, the digital transformation of cardiovascular medicine presents both an unprecedented opportunity and critical challenge for addressing the digital divide. While technological innovations hold immense promise for improving cardiovascular care, their benefits risk being unequally distributed if Hispanic/Latinx populations with limited English language proficiency continue to be excluded from research that validates these tools. Our proposed LLM-assisted workflow offers a practical pathway forward, leveraging technological advances to overcome traditional barriers to inclusive research while maintaining the quality and cultural sensitivity essential for meaningful participation. Moreover, emerging applications of LLMs have potential to further improve recruitment and retention in future digital health trials. Additionally, while this Comment has focused on the Hispanic/Latinx population with limited English language proficiency living within the United States, this hybrid workflow can be applied to nearly any racial or ethnic minority group via thorough pre-training and/or fine-tuning of an LLM in the intricacies of languages and dialects. By implementing such approaches while being mindful of weaknesses of LLMs, we can ensure that the digital revolution serves to bridge, rather than widen, existing differences in health-related outcomes for all communities.