Text-based Patient – Doctor Discourse Online And Patients’ Experiences of Empathy

A common definition of empathy includes the skills of understanding others’ thoughts and emotions (cognitive empathy, mentalizing), sharing emotional states with others (affective empathy), and responding to others’ distress with care and compassion (De Waal and Preston, 2017; Levenson and Ruef, 1992).

Empathic responding is known to matter in face-to-face medical consultations. A systematic review by Derksen and colleagues found that medical doctors’ empathy was associated with patients’ satisfaction, enablement, and adherence to treatment, decreased anxiety and distress, and was related to better clinical outcomes (Derksen et al., 2013). More recent research has been in line with these findings, showing that empathy is linked with patients’ compliance to treatment plans (Attar and Chandramani, 2012), patient satisfaction (Menendez et al., 2015; Pollak et al., 2011; Wang et al., 2018), enablement (Mercer et al., 2012), self-reported well-being (Mercer et al., 2016), and self-reported treatment outcomes (Steinhausen et al., 2014). Yet, contrasting findings also exist, showing that surgeons’ empathy after a single visit was not associated with patient-reported depression symptoms – and although a weak relation was found between empathy and decreased pain, it was not considered clinically significant (Kootstra et al., 2018). The mechanisms through which empathy improves wellbeing likely relate to decreased patient stress as well as improved adherence and enablement. For example, a study by Xu and colleagues showed that medical doctors’ empathy was associated with decreased inflammation among patients with Chron’s disease and that this effect was mediated by better sleep, decreased stress, and improved self-efficacy (Xu et al., 2020).

As in all interactive settings, both verbal and non-verbal communication are important for conveying empathy during medical consultations. Non-verbal cues may include body posture, facial expression, eye contact, and tone of voice that are appropriate for the patient’s situation (Vogel et al., 2018). As described by Dean and Street (2014) empathic verbal communication can include verbally recognizing the patient’s emotions: "I know this is stressful to you”, validating their emotional responses and giving space to explore them further: "Tell me about what’s going on”, as well as taking therapeutic action and verbally reassuring the patient about an action plan: "We will figure this out together”.

In addition to the importance of empathy during face-to-face encounters, we have recently shown that medical doctors’ empathy also plays a role at an online clinic during text-based consultations (Martikainen et al., 2022). Patients who reported their doctors as more empathetic also reported lower concern about their symptoms and perceived their symptoms as less severe after the online encounter. Furthermore, patients who rated their doctors as more empathetic experienced that their symptoms had alleviated more than those who rated their doctors as less empathetic two weeks later. The exact communicational content improving patient experiences was not, however, a focus of the earlier work. Text-based interaction differs from face-to-face encounters in various ways, especially in terms of displaying empathy. To develop telemedicine encounters and to motivate good quality communication, it is important to understand, what type of communicational content supports the experience of empathy during text-based medical consultation.

Communication in chats is often considered to be ‘quasi-synchronous’, or a ‘hybrid’ of written and spoken language (Garcia and Baker Jacobs, 1999; Keng Wee Ong, 2011). Interaction in chats differs from spoken interaction mainly in terms of the affordances – the possibilities and constraints of a specific technology – that can be used for communication (Norman, 1988; Gaver, 1991; Hutchby, 2001). Affordances determine how the users can produce conversational actions and how the overall conversation can unfold in chats. Since chat communication is text-based, the ways in which conversational features are produced and received differ greatly from spoken communication. Although differences between spoken and chat communication are widely studied, researchers have avoided treating chats as "leaner" in comparison with spoken communication, and instead, examined them as a unique form of mediated discourse (Arminen et al., 2016). This way, the potential of text-based communication can be revealed, not only its deviations from spoken communication (Herring, 1999).

In spoken communication, one expectation is that turns that belong together will occur temporally adjacent to each another (Schegloff, 2007). However, turns in a chat conversation are not always expected to relate to the turn that has been posted immediately above (Meredith, 2019). Instead, multiple topics and conversations can be interwoven, and anyone can intervene between an initiated topic and its response (Herring, 1999). In addition, self- or other-initiated repair – which is a common feature in spoken communication to deal with troubles that arise in speaking, hearing, or understanding talk – function differently in chats. In chats, repairs usually include correcting a misspelling or an error of some other kind in the text (Meredith, 2019). In addition, although a message is posted in a chat in its entirety (i.e., without repairing it while producing it), participants often edit their messages while they still write them, for the message to better respond to something posted by another participant or a message that someone has posted in between (Garcia and Baker Jacobs, 1999). Thus, participants can utilize the affordances of the medium already during the message production and "repair" the message even before it is being sent.

One major difference between a chat and spoken communication is access to the embodied conduct of co-participants during interaction (Herring, 1999). Nonverbal features such as facial expressions, gaze, body position and movements, intonation, verbal stress, and rhythm of speech remain hidden in chats. This has implications for the display of empathy and affiliative responses in chat interaction. Affiliation refers to such actions with which a recipient demonstrates that they have access to and understanding of the speaker’s affective stance as well as display that they support or endorse this stance (Sorjonen et al., 2021; Stivers, 2008). Resources for delivering affiliative actions can be verbal (e.g., response cries such as "oh wow!” and assessments such as "that’s wonderful/terrible!”) or nonverbal (e.g., a nod).

Over the years, chat users have created substitutes for expressing embodied cues, socioemotional information, and empathy, by using textual resources and typographic marks (Walther et al., 2005; Keng Wee Ong, 2011). Emoticons (i.e., facial expressions produced typographically with letters, numbers, punctuation marks, and symbols), emojis (i.e., facial expressions, people, objects, places, and so on in a more pictorial form), and images in graphics interchange format (GIF) are also commonly used to give information, react to messages, and express emotions (Lyons, 2018). Users vary in terms of how much and often they use emoticons and may have different understanding of the meaning of specific emoticons (Lu et al., 2016). Thus, users may have a very different strategy for communicating with close ones in comparison to a doctor in a digital clinic.

Despite the possibility for using emoticons and emojis, it has been acknowledged that displaying empathy is still somewhat challenging in chat communication, because to understand the thoughts and feelings of the other, people often rely on facial expressions and body movements (Pfeil and Zaphiris, 2007). In addition, since research investigating the displays of empathy and intimacy in chat encounters has mostly focused on interactions between friends and family (e.g., (Hassib et al., 2017; Y. Hu et al., 2017), this challenge prevails especially in institutional chat encounters, such as online counselling and text-based helplines (Predmore et al., 2017; van Dolen and Weinberg, 2019).

In a recent study of Moylan et al. (2022) the volunteer participants in an online helpline reported that although expressing empathy is an essential part of their work, in the absence of vocal cues, such as a tone of voice and pauses, it was difficult for them to reach the help-seeker’s emotional state. This created uncertainty of accurately understanding the help-seeker’s needs and emotion and whether the emotional level of their own response was suitable for the situation. Moylan et al. (2022) concluded that the difficulty of expressing empathy over a text can turn out to be a barrier to mutual understanding between the help provider and help seeker. Apart from this study, research investigating the actual content of emotional and empathetic chat discussions is scarce. Thus, the ways in which participants can and seek to display empathy, especially in institutional chat encounters such as an online clinic, requires further attention.

Research from another context of text-based interaction – online communities and support groups – has shown that giving and receiving emotional support and empathy in a verbal form is one major characteristic of these communities (Pfeil and Zaphiris, 2007; Rodgers and Chen, 2017; Wright, 2000). As an example, Pfeil and Zaphiris (2007) analyzed messages in an online community that was directed for the elderly and identified categories in relation to the display of empathetic content. They divided these categories into the messages written by the empathy-seekers ("targets") and messages written by the ones displaying empathy ("empathizers"). In the targets’ messages, empathy-seeking was shown as self-disclosure in text units that described the general feeling that the target described, narratives of the target’s current situation, information about the medical situation of the target, and the target asking for others’ support or advice. In the empathizers’ messages, empathy was shown as providing light support in terms of showing interest to the target by asking for more information or clarification, displaying encouragement without going into detail, and speaking out best wishes for the target or the whole community. Empathy was also displayed as providing deep support in terms of deep emotional support towards the target, reassurance for the information, action, or feelings that the target reported, and giving help and advice concerning the target’s situation. Finally, the empathizers also displayed empathy with self-disclosure by describing being in a similar situation or having a similar problem with the target. These observations describe in detail how it is possible to display empathy in a written form and would be worth investigating in a chat context as well.

In this study our aim is to further analyze the text-based patient-doctor discourses from our previous study (Martikainen et al., 2022), to investigate what conversational characteristics are present during text-based consultations and how they relate to the patient experience. To categorize the patients’ and doctors’ utterances we use an adapted version of the Roter Interaction Analysis System (RIAS) (Roter and Larson, 2002) that we have modified to suit the purposes of analyzing online text-based consultations.

Our research questions (RQs) are:

The study was conducted in a private healthcare provider’s online service. The service is used to treat symptoms and diseases that do not require a physical examination. The patients can use the online clinic through a browser or a mobile application, logging in using their online banking credentials. After logged in, the patient can open a new conversation with the doctor. The consultations are charged per discussion independent of their duration. The doctors, providing care through the service, can access the patients’ previous medical records, write prescriptions, and invite the patient to a face-to-face check-up if needed. The interaction is text-based but the patients can also send photographs to the doctor.

The data were collected as part of a study investigating the role of doctors’ empathy in patients’ experiences online and testing augmentations to an online anamnesis questionnaire to support patient experience as described earlier (Martikainen et al., 2022). The anamnesis questionnaire was created as part of the online consultation service, it is filled in by the patients when checking in to the online clinic, before interacting with the doctor. The questionnaire includes drop-down menus as well as spaces for the patient to describe their symptoms and requests in their own words. The doctor opens the discussion after reading the patients’ answers to the questionnaire.

Patients’ perceptions of doctors’ empathy were assessed using a Finnish translation of the Consultation and Relational Empathy (CARE) questionnaire (Mercer et al., 2004). The patients answered ten questions on the doctors’ ability to convey empathy on a five-point scale (0 – poor, 1 – fair, 2 – good, 3 – very good, 4 – excellent), they also had a possibility to state that the question did not apply to the situation. Internal consistency of the scale was found to be excellent (Cronbach's α = 0.97). A mean empathy score was calculated for participants who considered at least three out of ten of the questions as applicable (n = 159).

The patients also filled out basic demographic information on their educational level, income, gender, and age. Patient’s stress levels were assessed with the Perceived Stress Scale (Power, 2003).

The data were collected from June to November 2019. Patients’ interest to participate in the study was enquired at the end of the digital anamnesis questionnaire described above. After the online appointment, each patient willing to participate received and signed a written informed consent form to indicate whether their chat discussions could be used for the research.

Altogether 209 adult patients participated in the study. Of these participants, valid chat conversation data were available in 201 cases. In the 8 non-valid cases the full dialogue was not recorded due to technical problems. Of the participating 201 patients 135 (67.2%) were women, 64 (31.8%) had up to high school level education, 72 (35.8%) had a Bachelor’s degree or equivalent and 65 (32.3%) had a Master’s degree or higher education. For the analyses on patients’ experiences of doctor’s empathy, 172 (86%) participants with valid conversation data filled out a follow-up questionnaire within two weeks of the online encounter with the doctor (M = 4.9, SD = 3.5 days). Of these participants, 2 had met with the same doctor face-to-face after the meeting at the online clinic. These participants’ data was excluded from the analyses regarding the perceptions of empathy since meeting the doctor face-to-face might bias the patients’ evaluations of online empathy.

Most of the doctors working at the clinic during the data collection were experienced general practitioners or occupational physicians, while some were also specialized in pediatrics, otorhinolaryngology, and gynecology. All had previous experience of the online clinic work. The doctors had not received communication training prior to the study by the employer.

The patients’ consent was obtained only after the dialogue with the doctor had already taken place, thus the patients already knew what type of information they were choosing to share for the study. The doctors were informed about the data collection before it started. No identification information was collected from the doctors at any point. The data was stripped from all information that might have even indirectly led to identification of the participants and the data was stored on a secured server only accessible for the researchers involved in the study. The Research Ethics Committee in the Humanities and Social and Behavioural Sciences of the University of Helsinki approved the study protocol.

The consultations were investigated via the RIAS method modified for the purpose of this study. In addition to coding the discussions between the doctors and the patients, the patients’ answers and requests written in their own words in the anamnesis questionnaire, filled in before the consultation started, were also coded.

The RIAS system has been used to study various kinds of medical interaction settings, including consultations with general practitioners (Bensing, 1991), and medical specialists (Ong et al., 1998). Although most of the studies have examined spoken communication between patients and doctors in face-to-face settings, RIAS has also been modified for technology-mediated interaction in video-based telemedicine consultations (Miller and Nelson, 2005). To our knowledge, RIAS has not been applied for classifying chat-based encounters in earlier research.

The original RIAS includes 62 categories for classifying the communicative utterances: 34 items for doctor communication and 28 items for patient communication (Ong et al., 1998). These 62 categories are merged into clusters and the clusters are further divided into Instrumental/Task focused talk and Affective/Socio-emotional talk (Miller and Nelson, 2005; van den Brink-Muinen et al., 2002). In this paper we will use the terms Instrumental and Affective talk in line with van den Brink-Muinen and colleagues (van den Brink-Muinen et al., 2002).

The clusters pertaining to Instrumental talk are: "Biomedical talk", "Psychosocial talk", and "Procedural statements" and categories included in Affective talk are: "Social talk", "Agreement", "Rapport building", and "Facilitation" (see Table 1 for a description of each category and cluster).

In this study, we used an iterative coding strategy to modify the RIAS to fit the text-based data. The RIAS categories were not translated into Finnish as they were used only by the researchers. The process of modifying the RIAS is described in detail in Appendix 1. After the modifications were finalized, two coders (authors 3 and 4) analyzed all the recorded data and a mean value between the two coders was used to indicate the number of utterances pertaining to each conversational category. One new cluster, "Technology related exchange”, and one new category, "Sick-leave related talk” (pertaining to the Biomedical talk cluster) were identified during the process (Table 1).

Mean number of individual words written during the patient-doctor interaction was 57.0 (SD = 33.9) by the doctors and 38.5 (SD = 39.4) by the patients. During the consultations, the most common topics included respiratory infections (19%), musculoskeletal problems (19%), urinary tract infections (17%), eye infections (11%), and eczema or other skin problems (8%).

Table 2 shows that Biomedical talk was the most common type of talk among patients (67.2%) and doctors (52.2%), followed by social talk (12% for patients and 16.6% for doctors). Doctors also used more procedural statements (9.8%) and technology related talk (9.1%), when compared to the patients (2.5 and 1.8%, respectively).

Based on the 201 analyzed chat encounters, the following three original RIAS categories were never used: Legitimizes; Asks questions (open ended) — psychosocial/feelings; Asks questions (open ended) — lifestyle.

As shown in Table 3, inter-rater reliability was high for Social (ρ  = 0.950), Biomedical (ρ  = 0.939), and Technology related talk (ρ  = 0.833), and moderate for Procedural statements (ρ  = 0.693) and Agreement (ρ  = 0.687). The reliability was poor for Rapport building (ρ  = 0.193), Facilitation (ρ  = 0.008) and Psychosocial talk (ρ  = 0.341).

The categories with low inter-rater reliability were also very rarely found in the data, which may explain why they could not be reliably assessed. When an average by rater 1 and 2 was calculated, the mean per conversation (doctors and patients combined) for facilitation utterances was 0.33 (SD = 0.56), mean for rapport building was 0.72 (SD = 0.84) and mean for psychosocial talk was 0.58 (SD = 1.45).

Table 4 shows that the amount of doctors’ social talk (Spearman’s ρ  = 0.23, p = 0.003) and procedural statements (Spearman’s ρ  = 0.24, p = 0.002) were positively correlated with the patients’ evaluations of doctors’ empathy. The correlations remained significant also after controlling for patients’ age, gender, stress levels, and the number of words written by the doctors. However, the effect sizes of the significant associations remained weak (ρ  < 0.30).

This study has various strengths. The study was conducted in a real-life environment with a relatively large number of participants. We used previously validated methods for assessing medical doctors’ empathy (Mercer et al., 2004) and categorizing the patient-doctor discourse (van den Brink-Muinen et al., 2002; Ong et al., 1998). Two raters categorized all the data and only the categories yielding acceptable reliability were used in the further analyses.

The RIAS method has been used extensively in studies focusing on face-to-face clinical encounters (e.g., Ong et al., 1998; Bensing, 1991; van den Brink-Muinen et al., 2002), and it is one of the few validated methods for medical interaction analysis(Roter and Larson, 2002). Furthermore, it has been modified for video-based telemedicine encounters as well (Miller and Nelson, 2005). According to Miller and Nelson (2005) two of the main weaknesses of RIAS are that it does not account well for situations with multiple participants and for non-verbal communication. In this sense applying RIAS in text-based consultations may be more accurate as non-verbal communication does not occur and only two participants are involved in the dialogue.

The study also has some weaknesses. First, using the RIAS method with text-based data might not recognize all the nuances of the patient-doctor discourse. Although we were able to reliably recognize several communicative categories, some of the analyzed categories showed low interrater reliability. It would be important to conduct future studies focusing on the discourse content using a qualitative approach that could better consider the fine-tuned differences in the communication that may lead to improved experiences of empathy. Second, although this coding scheme is applicable to text-based medical context like the one studied here, the findings from this study may not be generalizable to other types of medical consultations, since this study was conducted at a digital clinic in which only medical problems not requiring a face-to-face visit were treated. Analyzing consultations with more difficult medical problems or emotional topics might yield different results as compared to our study. Also, the applicability of the coding scheme to medical consultations outside Finland should be tested in future studies. Some elements of the coding scheme might not be useful in all contexts (e.g., the sick leave utterances), however, having the possibility to use these elements may improve the accuracy of the method, while they can simply be left without ratings in contexts where they are not needed.

Although the partial correlations accounted for the number of words written by the doctor (indicating the length of the conversation), and patient’s age and gender, other confounder such as the doctor’s specialization or whether the patient was interacting with the doctor for the first time may have affected the results, a future study should investigate these potential confounders more closely. It is also of note that a validation study of the modified method was not carried out and we did not assess intra-rater reliability. These issues should be focused on future research.

Furthermore, it should be noted that when analyzing agreement between coders we calculated the frequency of the codes per dialogue. This may have overestimated the agreement rates since it does not consider the agreement in timing when the codes were applied.