EHR-Integrated interactive app helps patients understand their health records

In recent years, federal regulations like the 21st Century Cures Act have made it mandatory for healthcare systems to offer patients electronic access to their health records. This has led to broad adoption of Fast Healthcare Interoperability Resources (FHIR) APIs that let apps pull data directly from EHR systems.

But having access doesn’t mean the data is useful. Over half of adults in the U.S. read at or below a sixth-grade level, and about one in five Americans speaks a second language at home. Many struggle with digital literacy or find health records too technical to understand. These limitations mean that the benefits of digitized records often reach only the most educated or tech-savvy.

Researchers at Stanford think they might have a solution. It’s a mobile app called LLMonFHIR, and it combines artificial intelligence, natural language understanding, and voice interaction to make patient records easier to understand.

LLMonFHIR acts as a sort of translator that sits between patients and their medical records. When asked a question about a patient’s health data, it’ll give an answer in plain English, Spanish, or any other preferred language.

Health-record answers, now jargon free

The app connects to a patient’s health records through Apple’s HealthKit platform, which pulls in data from various healthcare providers. Patients can then ask questions in natural language, spoken or typed, and receive customized responses. These might include explanations of medication regimens, summaries of lab results, or suggestions for managing chronic conditions.

LLMonFHIR works by using a large language model (LLM) and a method known as retrieval-augmented generation (RAG). When asked a question, it doesn’t try to process a patient’s entire medical history at once. Instead, it picks out the relevant parts, which makes it faster and more accurate.

The app can read its answers aloud, switch between languages, and adjust the level of detail to suit the user’s needs.

But does it work?

To evaluate its effectiveness, researchers conducted a pilot study with five physicians who reviewed app-generated responses to patient questions. These questions were based on synthetic health records created to mimic real-world complexity. Physicians were asked to judge each answer’s accuracy, clarity, and relevance using a five-point scale.

Overall, LLMonFHIR scored well. Questions about medications like how to take them, possible side effects, and allergy risks consistently earned top marks. The answers were generally precise, easy to follow, and included helpful safety information, like when to defer to a doctor.

When it came to summarizing health conditions or giving lifestyle advice, the results were still strong but more mixed. Reviewers noted that the model sometimes included irrelevant personal details. like education level or social history, when asked to focus on medical issues. In some cases, it offered generic advice rather than personalized guidance.

The biggest variability showed up in questions about lab results. While some responses were spot-on, others left out key details or failed to interpret the results meaningfully. This reflects one of the ongoing challenges in designing AI tools for healthcare: the systems need to not only retrieve the right data, but explain it in ways that are specific, contextual, and actionable.

Next up, opening the app to everyone

Like all early-stage technology, LLMonFHIR has its caveats. The app currently works only on iOS, which may limit its reach. Some people still lack access to smartphones or the internet entirely. And while the model is good at answering many types of questions, it occasionally overlooks important context or delivers inconsistent outputs, issues that developers are working to address.

The team plans to expand the app to Android, incorporate more rigorous testing with real patients, and improve how the model processes time-based data like medication timelines and recent versus past diagnoses. There are also plans to explore running more parts of the model locally on users’ devices, which could improve privacy and reduce dependency on cloud infrastructure.

Overall, LLMonFHIR represents a promising step toward making digital health records more accessible and understandable for patients who have historically been left behind. For patients who’ve always struggled to make sense of their medical records, whether because of language barriers, reading difficulties, or confusing medical jargon, this technology has the potential to increase engagement, support self-management, and improve outcomes.