Analyzing patient secure messages using a fast health care interoperability resources (FIHR)-based data model: Development and topic modeling study

Background: Patient portals tethered to electronic health records systems have become attractive web platforms since the enacting of the Medicare Access and Children's Health Insurance Program Reauthorization Act and the introduction of the Meaningful Use program in the United States. Patients can conveniently access their health records and seek consultation from providers through secure web portals. With increasing adoption and patient engagement, the volume of patient secure messages has risen substantially, which opens up new research and development opportunities for patient-centered care. Objective: This study aims to develop a data model for patient secure messages based on the Fast Healthcare Interoperability Resources (FHIR) standard to identify and extract significant information. Methods: We initiated the first draft of the data model by analyzing FHIR and manually reviewing 100 sentences randomly sampled from more than 2 million patient-generated secure messages obtained from the online patient portal at the Mayo Clinic Rochester between February 18, 2010, and December 31, 2017. We then annotated additional sets of 100 randomly selected sentences using the Multi-purpose Annotation Environment tool and updated the data model and annotation guideline iteratively until the interannotator agreement was satisfactory. We then created a larger corpus by annotating 1200 randomly selected sentences and calculated the frequency of the identified medical concepts in these sentences. Finally, we performed topic modeling analysis to learn the hidden topics of patient secure messages related to 3 highly mentioned microconcepts, namely, fatigue, prednisone, and patient visit, and to evaluate the proposed data model independently. Results: The proposed data model has a 3-level hierarchical structure of health system concepts, including 3 macroconcepts, 28 mesoconcepts, and 85 microconcepts. Foundation and base macroconcepts comprise 33.99% (841/2474), clinical macroconcepts comprise 64.38% (1593/2474), and financial macroconcepts comprise 1.61% (40/2474) of the annotated corpus. The top 3 mesoconcepts among the 28 mesoconcepts are condition (505/2474, 20.41%), medication (424/2474, 17.13%), and practitioner (243/2474, 9.82%). Topic modeling identified hidden topics of patient secure messages related to fatigue, prednisone, and patient visit. A total of 89.2% (107/120) of the top-ranked topic keywords are actually the health concepts of the data model. Conclusions: Our data model and annotated corpus enable us to identify and understand important medical concepts in patient secure messages and prepare us for further natural language processing analysis of such free texts. The data model could be potentially used to automatically identify other types of patient narratives, such as those in various social media and patient forums. In the future, we plan to develop a machine learning and natural language processing solution to enable automatic triaging solutions to reduce the workload of clinicians and perform more granular content analysis to understand patients' needs and improve patient-centered care.