A Large Language Model-Based Mobile Application for Querying FHIR Data: Overcoming Functional, English, and Health Literacy Barriers to Patient Electronic Health Record Access

Abstract

Background: Recognizing the critical role of patient access to electronic health records (EHRs) in enhancing healthcare quality and patient empowerment, federal legislation such as the 21st Century Cures Act and the Office of the National Coordinator for Health Information Technology (ONC) Cures Act Final Rule mandates nationwide EHR interoperability through Fast Healthcare Interoperability Resources (FHIR) application programming interfaces (APIs). Despite these advances, key barriers to patient EHR access–low functional, English, and health literacy–persist, impeding equitable access to these benefits.

Methods: We present LLMonFHIR, an open-source mobile application that uses large language models (LLMs) to allow users to “interact” with their health records at any degree of complexity, in various languages, and with bidirectional text-to-speech functionality. The application uses OpenAI’s GPT-4, leverages Retrieval Augmented Generation (RAG) function calling, and was built using Stanford’s Spezi framework for the rapid development of modern, interoperable digital health applications. A pilot evaluation of physician perspectives on the accuracy, relevance, and understandability of the model’s responses was conducted using the SyntheticMass FHIR patient dataset.

Findings: LLMonFHIR received generally high accuracy, understandability, and relevance scores from physician reviewers, averaging 4·53, 4·73, and 4·53 out of 5, respectively, across tested queries. The application adapted responses to different patient profiles, effectively translating FHIR data into patient-friendly and non-English languages. Challenges summarizing health conditions and retrieving lab results were noted, with variability in responses and occasional omissions underscoring the need for precise pre-processing of data.

Interpretation: LLMonFHIR’s ability to generate responses in multiple languages and at varying levels of complexity, along with its bidirectional text-to-speech functionality, give it the potential to empower individuals with limited functional, English, and health literacy to access the benefits of patient-accessible EHRs. Prior to implementation, patient validation, improved resource identification, and on-device execution are necessary to ensure patient safety, enhance privacy, and reduce costs.

Funding: None.

Declaration of Interest: None of the authors have a competing interest. The LLMonFHIR application and the Stanford Spezi ecosystem are open-source and licensed using the MIT license. Stanford Spezi is being used to foster a digital health ecosystem and teach the next generation of digital health leaders.

Ethical Approval: On the basis of the information provided, the IRB has determined that this project does not meet the definition of human subject research as defined in federal regulations 45 CFR 46.102 or 21 CFR 50.3. No further IRB review is required. Please see your HSR application form in eProtocol for the completed determination and any additional instructions. Protocol Number: 75096 (NEW). Review Type: HSR. Protocol Director: Aalami. Department: Surgery - Vascular Surgery. Protocol Title: LLMonFHIR Synthetic Data Validation Study.

Keywords: FHIR API, EHR, mHealth, LLM