JMIR Publications

ABSTRACT

Background:

Patient portals, tethered to Electronic Health Records (EHR) systems have become attractive online platforms since the Medicare Access and CHIP Reauthorization Act (MACRA) and the “Meaningful Use” program in the United States. Patients can conveniently access their health records, seek consultation from providers, and share their viewpoints on health care services via secure messaging in patient portals. With the increasing adoption and patient engagement, the volume of patient secure messages in free text format has risen substantially, which opens up new research and development opportunities for patient-centered care.

Objective:

We propose to develop a data model for patient secure messages based on the Fast Healthcare Interoperability Resources (FHIR) for identifying and extracting significant information. We also create an annotated corpus to analyze the contents of patient secure messages with topic modeling techniques.

Methods:

We initiated the first draft of the data model for patient secure messages and annotation guideline by analyzing FHIR resources and manual review of 100 sentences randomly sampled from over 2 million patient-generated secure messages from the online patient portal at Mayo Clinic Rochester between February 18, 2010 and December 31, 2017. We then annotated an additional set of 100 randomly selected sentences using the Multi-purpose Annotation Environment (MAE) tool and updated the data model and annotation guideline iteratively until inter-annotator agreement was satisfactory. After that, we created a larger corpus by annotating 1,200 randomly selected sentences. We eventually calculated the frequency of identified medical concepts in it and performed topic modeling analysis to learn the hidden topics of patient secure messages related to 3 highly mentioned micro-concepts - fatigue, prednisone and patient visit.

Results:

The proposed data model has a 3-level hierarchical structure of health system concepts including 3 macro-concepts (e.g., foundation and base, clinical, and financial macro-concepts), 28 meso-concepts (e.g., condition, medication, and appointment), and 85 micro-concepts (e.g., attributes of the meso-concepts). The annotated corpus contains 34% foundation and base macro-concepts, 64.4% clinical macro-concepts, and 1.6% financial macro-concepts. The top 3 meso-concepts among all the 28 meso-concepts are condition (20.41%), medication (17.13%) and practitioner (9.82%). Topic modeling identified meaningful hidden topics of patient secure messages related to fatigue, prednisone, and patient visit.

Conclusions:

Our data model and annotated corpus enable us to identify and understand important medical concepts mentioned in the patient secure messages and prepare us for further Natural Language Processing (NLP) analysis of such free texts. The data model could be potentially used for automatically identifying and analyzing 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 NLP 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.