Look Who's Talking: Application of a Theory-Based Taxonomy to Patient–Clinician E-mail Messages

Background

Patient-centered communication includes consideration of patients' preferences and needs when providing care, enhancing the patient–clinician partnership, and including patients in the decision-making process when they desire inclusion. ¹ The Institute of Medicine recommends that communication be delivered through the modalities preferred by patients, and notes that electronic messaging (i.e., secure messaging) offers an opportunity to provide that communication between office visits. ^{2 –4} Therefore, electronic communication between patients and clinicians has the potential to provide patient-centered care between clinical visits when delivered in a patient-centered way. As with verbal communication, what is said (i.e., content) in a secure message is just as important as how the message is delivered.

There is supporting evidence of associations between electronic message use and some patient outcomes (e.g., glycemic levels in patients with diabetes), but less evidence among other outcomes (e.g., blood pressure among patients with hypertension). ⁵ Research does not demonstrate a consistent association between secure messaging use and health care utilization. ^{6 –10} These studies, however, quantified messages rather than considering message content. If patients are using secure messaging to (1) express uncertainty around their illnesses, (2) request information to reduce that uncertainty, or (3) facilitate self-management of their condition, then their clinicians' responses may impact how those patients manage their condition or change their decision to seek additional health care services. Classification of message content gives context to those messages and allows for analysis on what patients request of their clinicians and how clinicians respond, thereby providing insight into whether patients' requests are addressed, and addressed in a way designed to improve patients' care or understanding of their condition.

Introduction

This study evaluated the application of a theory-based taxonomy developed specifically to classify secure message content. The Uncertainty in Illness Theory (UIT) highlights the importance of credible clinical authorities in providing information to patients experiencing uncertainty around their symptoms, conditions, or health care experience. ¹¹ Reductions in uncertainty can lead to improved patient satisfaction, better treatment adherence, and ultimately, better outcomes. ¹ If patient-centered communication functions such as information exchange and enabling self-management and decision-making ¹² can be measured in secure message content, then associations with patient outcomes should be detectable following the direct and indirect pathways described by Street et al. ¹² Based on the UIT, ¹¹ this study's taxonomy (1) permits classification of patient-generated messages demonstrating uncertainty and (2) differentiates between clinician information sharing and action-oriented responses. In this study, we report on the pilot application of our taxonomy using a small sample of patient-initiated message threads generated during a single calendar year; we also explore whether differences existed in patient and clinician characteristics associated with different taxonomic codes.

Methods

This is a retrospective study using a random sample of 73 patients from among patients with diabetes, hypertension, or both conditions who sent at least one secure message using the Virginia Commonwealth University Health patient portal during the study period of January 1 and December 31, 2017. To be included in the sample, patients had to send at least one message that was not a prescription refill or scheduling request. Patients with diabetes and hypertension were identified based on the presence of diagnosis codes for the selected condition, either from at least two outpatient visits or from one inpatient visit during 2016.

We included only patient-initiated threads—inclusive of the initial patient-generated message and all responses—in the analysis. Messages were manually extracted from the electornic health records and imported into NVivo (QSR International, Australia) for coding. We grouped messages by patient and thread identifiers, allowing for message coding in the context of the full message thread. Coding units were constrained to the length of a single message and were frequently shorter. The taxonomy codes, or taxa, were defined to be mutually exclusive; we could apply multiple taxa to a single message if there were multiple requests or responses within the message.

Because this was a pilot study with the intent of determining if the taxonomy included all necessary codes to support both clinician- and patient-generated content, a single coder (D.M.H.-G.) applied taxa to all messages. Appendix Tables A1 and A2 list the patient- and clinician-generated taxa, respectively, and the associated definitions used during the coding process. We insured face validity of the taxa by leveraging taxa reported by other researchers, although the authors organized the taxa and developed and refined the taxa definitions according to the taxonomy's theoretical basis.

We detected differences in taxa frequency using chi-square analysis with the patient as the unit of analysis. Analyses included differences in coding frequency by patient demographics and health condition for codes assigned to both patient- and clinician-generated messages.

Results

The study sample included 658 patient-initiated message threads with 1,751 patient- and clinician-generated messages. We assigned a total of 2,055 taxa (i.e., taxonomic codes) to those messages. Table 1 lists the study sample's patient demographics.

Table 2 displays the taxa frequency of patient-generated messages. Sharing information with clinicians was most common (33%) and included self-reporting and clinical updates at 6% each, as well as responding to clinicians' messages (21%). Twenty-six percent of patient-generated content related to information seeking (logistics and symptoms or condition).

Eighteen percent of the threads (n = 116) included no clinician message response; another 11% (n = 72) included documentation of a phone response. Table 3 presents taxa distribution across clinician-generated messages. Twenty-eight percent of clinician-generated messages reflected information sharing (medical guidance and orientation to procedures); 17% of messages included a clinician's request for additional information from the patient. Almost 11% of clinician-generated messages reflected the clinicians denying, or deferring responses to, patients' requests. Another 18% reflected only partial fulfillment (13%) or acknowledgement (6%) of patients' requests.

Statistical differences in taxa were observed by patient age, health condition, sex, and race. Forty-seven percent of patients younger than 60 years received at least one message from a clinician that recommended they be seen in the office compared with 21% of patients aged 60 years or older (p = 0.0160). Clinicians were also more likely to recommend to patients with only diabetes that they be seen in the office (45%) compared with either patients with both conditions (14%, p = 0.0207) or patients with only hypertension (38%, p = 0.0542).

More patients with only diabetes sent medication refill requests (68%) compared with patients with both diabetes and hypertension (36%, p = 0.0346). A higher proportion of patients with only diabetes requested a new or change in medication (45%) compared with patients with only hypertension (14%, p = 0.0121).

Differences based on patient sex and race approached significance. A higher proportion of female patients (40%) received recommendations from clinicians that they be seen in the office than male patients (21%, p = 0.092). Similarly, white patients were more likely to send messages seeking logistical information (73%) and sharing appreciation or praise (35%) over black patients (50%, p = 0.0582, and 14%, p = 0.0567, respectively). Proportionally, white patients (80%) were more likely to receive at least one e-mail from a clinician with information sharing content than black patients (61%, p = 0.0811). Thirty percent of white patients received at least one message from a clinician with praise or appreciation content compared with 11% of black patients (p = 0.0591).

Discussion

With this study, we demonstrated how this taxonomy classifies patient- and clinician-generated content consistent with the UIT. It is the first study to describe patient characteristics associated with secure message content.

Consistent with the UIT, this study demonstrates that almost 4 in 10 patient-generated messages reflected some level of patient uncertainty. A critical feature in managing patients' uncertainty around their illnesses includes the availability of trusted authorities (e.g., clinicians), ¹¹ which is enhanced between office visits through the use of secure messaging. ³ Those trusted clinical sources may educate patients about their health condition and provide support as patients adapt to their health status, leading to better self-management and improved health outcomes. ^11,12 We demonstrated that the majority of patient-generated content involves patients seeking or sharing information with their clinicians and almost half of clinician responses included information sharing and request fulfillment content that should reduce patients' uncertainty.

One-quarter of clinician responses left patients' requests unfulfilled or unaddressed (combining partially fulfilled, deny, defer, and acknowledged codes), and 18% of threads were left unanswered. This lack of response or fulfillment may increase patients' uncertainty and lead to increased stress and poorer outcomes. ^{12 –14} Even if those threads for which a clinician response was not documented had a response through another form of communication (e.g., phone), that would not necessarily be considered patient-centered since clinicians were not using the patients' preferred communication modality.

One limitation of this study is that taxa were assigned to patient and clinician messages in isolation; in reality, each message is part of a message thread that is similar to an in-person encounter. As such, there may be bidirectional communication as options are discussed, dismissed, and new options agreed upon. This was in evidence by the fact that 17% of clinicians' content was information seeking and 21% of patients' content were responses to clinicians. Future studies should examine whether pairings of patient- and clinician-generated codes within threads are associated with outcomes; for example, is there a difference in outcomes if a clinician responds with information seeking before providing an information sharing response, rather than deferring or providing an information sharing response initially?

This study found that patient characteristics are associated not only with the types of messages patients send to their clinicians but also with their clinicians' responses as well. Prior research demonstrated differences in messaging use by patient demographics, ^{15 –23} but none explored differences in patient demographics associated with message content. Epstein and Street ²⁴ noted that a number of patient characteristics, including those identified in this study, mediate the relationship between patient-centered communication and health outcomes. In addition, written forms of communication may present challenges for older patients and individuals with low health literacy. ²⁵ Studies on technology-mediated communication such as e-mail have found that senders' and receivers' genders influence message content and its interpretation. ^{26 –28} It is therefore important to understand the influence of these characteristics on the messages that patients send. Future studies that employ larger sample sizes should explore which patient demographics are significantly associated with message content in adjusted analyses.

We present here the first study to describe differences in clinician messaging responses based on patient age, sex, race, and health condition. Our analyses are descriptive because of the small sample size: the large number of taxa across a small sample makes regression analyses unreliable. Future studies should increase the number of patients to permit adjusted analyses that might determine if the differences we observed by patients' race, age, and sex are statistically significant after controlling for patients' characteristics such as health status and condition, demographics, and other social determinants such as health literacy.

The differences we identified by race were borderline significant, with smaller proportions of black patients seeking logistical information and receiving information sharing content compared with white patients. The lower information sharing rate from clinic staff may be related to the lower rate of information requests from black patients; however, because our analyses examined each taxon in isolation rather than as part of the full conversation, we cannot be certain of this fact. Future studies should explore the full call-and-response nature of the conversation by analyzing the pairing of the message content relative to patient characteristics (e.g., is there a difference in clinicians' responses to patients by race among patients who sent logistical information?). Two-thirds of studies in a literature review of the effects of race on patient–physician communication reported that black patients had fewer acts of participation during their physician visits ²⁹ ; the lower percentage of black patients requesting logistical information may be a reflection of that attitude.

More research must be performed to determine why there appear to be differences by race in receiving praise or appreciation from clinicians. Street et al. noted that patient-centered communication was influenced by racial concordance between patients and clinicians, ^30,31 but other research found no differences by race. ^32,33 Our study did not incorporate characteristics of the clinicians. Future studies should incorporate clinicians' age, race, and sex, which have been shown to influence patient–clinician communication.

We conducted this small study to determine if the codes developed for the taxonomy were adequate to cover all concepts communicated between patients and clinic staff through secure messaging. As such, we did not apply a robust coding process, in which multiple coders applied the taxa to messages, coding was compared and discrepancies discussed, and resolutions resulted in refined taxa definitions. The fact that this study employed a single coder, who was also the developer of the taxonomy itself, is less than ideal. ³⁴ Our hope is that this taxonomy will be applied in a more rigorous manner to larger samples to determine if these results can be reproduced in different settings and for different conditions.

These differences we report in this study are important to understand because clinicians' behaviors influence patients' use of secure messaging; for example, patients were more likely to initiate their own threads if their clinicians initiated more message threads, responded quickly, or had higher overall response rates. ³⁵ Furthermore, secure messaging is a favorite functionality on patient portals, ^36,37 and many patients expressed intention to send messages to their clinicians if given the opportunity. ^18,38 Moreover, patients reported that message responses were generally of higher quality and they felt less rushed compared with phone communication. ³⁹ Given strong patient preferences to use secure messaging coupled with the impact that clinicians' behaviors have on patients' use of secure messaging, it is important to understand the source of the differences in clinicians' responses based on the patients' characteristics observed here. Furthermore, there is a need to understand if those differences are associated with patient outcomes. If clinicians treat patients differently through secure messaging and those differences are associated with health or health care utilization outcomes, disparities in care and health status may result. Future studies should explore these relationships further and whether there are characteristics associated with the clinician that impact the relationships.

Conclusions

This taxonomy provides opportunities to explore secure message content in new ways and to assess the associations different message content may have with patient outcomes. In addition, with a small study population, this research was able to signal that message content varies by patient characteristics. Perhaps most importantly, clinician responses varied by patient characteristics, indicating the potential of differential care provided to patients through secure messaging.

It is important to remember that patient-centered care should be provided in the modality preferred by the patient. ¹ This study demonstrated that almost one in five patient-initiated message threads received no message response from a clinician. Responses provided through other modalities such as phone calls may not be fully patient-centered because patients' communication preferences are not being considered.

The findings from this pilot study present many opportunities for future research. Application of this theory-based taxonomy permits research that examines differences in message content across patient and clinician populations, as well as content associations with patient outcomes. The latter is critical to understanding how to encourage additional secure message use and provide education to clinicians on how to best communicate using secure messaging.