First, we conducted Study 1, a paper-pencil survey with COPD patients treated at a leading Swiss Hospital in the German-speaking part of Switzerland. Besides covering sociodemographic and health-related questions, the survey explored baseline differences in patient preferences for a deliberative, informative, interpretive, or paternalistic interaction style with a hypothetical health care CA.

The outcomes informed Study 2, an online experiment. For this study, we recruited COPD patients from four hospitals in the German-speaking part of Switzerland, from the Swiss Lung Association, and from an honorary led self-help association for COPD patients in the German-speaking part of Switzerland. We designed a between-subject online experiment where patients were randomly assigned to interact with a CA that followed either a deliberative or paternalistic interaction style. We chose these two styles since (1) we have already developed and experimentally tested the implementation of deliberative and paternalistic CA interactions in a recent study [29]; (2) there is the most information in the literature for these two styles regarding the moderating influences of sociodemographic and health-related variables; and (3) we expected to find significant effects when choosing the most and least preferred interaction styles as determined in Study 1. Both studies did not fall within the scope of Human Research Law, according to the local Swiss ethics authority, and thus did not require any formal authorization.

The primary objective of Study 1 was to explore whether general differences exist between interaction style preferences of COPD patients for their interaction with a CA. Thus, this part of the study was exploratory by nature and did not contain a detailed power analysis.

We conducted an a priori power analysis for Study 2 using R software (version 3.5.2) and the R package WebPower [30]. To identify a medium effect (f²=0.15) [31] in a hierarchical multiple regression with an alpha level of .05, a statistical power of 0.80, a reduced model with one predictor, and a full model with 13 predictors, a total of 127 participants was required.

For Study 1, we defined the following inclusion criteria: (1) COPD diagnosis, (2) age of 18 years or older, and (3) ability to speak German.

We defined the same inclusion criteria for Study 2. Here, the first inclusion criterion was checked before distributing the link to the online experiment. The link was only sent out to patients who were registered as COPD patients at any of the participating hospitals, the lung association, or the self-help association. In addition, patients also had to confirm that they have COPD during the online experiment. The second and third inclusion criteria were checked at the beginning of the online experiment. If patients did not confirm either being of age or being able to speak German, the experiment was automatically completed. There were no exclusion criteria.

Study 1 was administered as a paper-and-pencil survey and divided into four parts. Before starting with the actual questions, we provided general information about the survey (ie, we clarified the objective and structure, and provided an illustrative explanation of a CA-based intervention). After querying sociodemographic and health-related questions, the survey explored patients’ preferences regarding their interaction style with a hypothetical CA. Patients could choose from snippets portraying exemplary interactions with a CA, and each snippet depicted a different interaction style (see Figure 1 for an overview).

The procedure of Study 2 was as follows. Participants agreed to the study conditions and confirmed the study inclusion criteria (ie, being of age and being able to speak German). After querying standard demographic data (age, gender, mother tongue, and education), patients answered questions about their general health status and COPD. Patients were then randomly assigned to interact with a CA presenting a deliberative or paternalistic interaction style. The interactions were text-based and followed by a prescripted dialogue based on two predeveloped scripts. These scripts were developed and assessed in a recent study, where they were verified to be perceived as eliciting a deliberative or paternalistic interaction style between a CA and users [29]. During the conversation with the CA, patients could choose between one to three predefined answer options, which were identical in both conditions (ie, deliberative and paternalistic interaction styles). Deviations between the answer options only occurred when needed to keep the conversational flow realistic. Both interactions with the CA (ie, paternalistic and deliberative) were approximately of the same length and duration (38 conversational turns in the deliberative version and 32 in the paternalistic version with a reading duration of roughly over 12 minutes). After the interaction with the CA, participants were asked to evaluate the interaction with the CA on several dimensions. Details about the measures can be found in the measures section below.

To conclude the experiment, participants had to answer a short COPD health literacy quiz. The questions were based on the standardized Bristol COPD Knowledge Questionnaire (BCKQ), a multiple-choice questionnaire developed to measure the disease-specific knowledge of COPD patients [32]. Finally, patients reported on their perception of the length of the study and could leave some free-text feedback. All questions of Study 2 can be found in Multimedia Appendix 1.

We used Qualtrics, a software-based online survey and data collection platform, for the online experiment and for randomly assigning patients to one of the two experimental settings (ie, interaction with a CA using a deliberative interaction style and interaction with a CA using a paternalistic interaction style). We further used Collect.chat, a commercially available chatbot software, to develop the CA dialogues, and iframe to embed the CA into the Qualtrics HTML.

We recruited the participants for Study 1 from the pulmonology department of a leading Swiss hospital. The patient database included around 1300 patients, and we contacted all by mail. Eligible patients received the printed survey and a letter containing information about the survey and participation conditions. We also provided a prefranked return envelope to reduce the necessary effort to reply to the survey and to minimize financial expenses for the participants. The postal send out took place on April 9 and 10, 2019, and we started to receive responses 1 week later. We received replies for 2 months in total, and the majority of replies reached us in the first 3 weeks.

The patients of Study 2 were recruited during a 3-month period from February to April 2020 at six study sites in Switzerland. The study sites were the pulmonology departments of four hospitals in the German-speaking part of Switzerland, the Swiss Lung Association, and an honorary-led self-help association for COPD patients. Participants were recruited via email or postal mail or in-person by participating health care professionals on-site. When they were recruited in-person, participating health care professionals handed out a flyer to potential participants. This flyer contained some information on the study and a link to the online experiment (see Multimedia Appendix 2 for the flyer). Participating health care professionals had previously received study instructions from the study authors. When patients were recruited via email, they received an email from the participating hospital, the Swiss Lung Association, or the self-help association. In this email, the participating contact person explained the study and asked for voluntary participation. The email further contained the same flyer as used for the on-site recruitment. When participants were recruited via postal mail, they received a letter from the participating hospital with study information and a link to the online experiment, which they had to type into a web browser. The postal letter is presented in Multimedia Appendix 3. We followed a multichannel and multisite recruiting process. More precisely, we contacted 903 patients at three study sites via email and postal mail. Additionally, 110 flyers were printed and displayed at two further study sites. Here, the flyers were displayed in the study sites’ waiting rooms and were free to pick up for the patients. There were no additional prompts by medical personnel to participate in the study. The email and postal mail had the same design and included exactly the same content as the flyer. At the sixth study site, 27 local COPD patient groups were contacted via email. Importantly, the resulting response rate has to be interpreted as a lower bound estimate since patients might have received an invitation more than once. For example, they could have been regular patients in hospital X and, at the same time, also a member of a COPD patient organization that distributed study invitations.

To start the online experiment, participants had to either click on the study link in the email or type the URL in their browser in case they received a postal letter. A simplified URL (ie, only portraying a shorter number of signs) was created with Bitly [33] to access the online experiment and to reduce barriers to participate. The experiment was available online from April to August 2020.

For Study 1, we gathered basic sociodemographic (age, gender, and education) and health-related data (GOLD COPD level) [34]. GOLD stands for Global Initiative for Chronic Obstructive Lung Disease and is an internationally used scale for classifying the severity of COPD [34]. Patients reported sociodemographic and health-related data before being presented with the four main physician-patient interaction styles.

Study 2 was pretested with two pulmonologists of a leading Swiss hospital (both specialized in COPD), one advanced practitioner nurse in lung disease, and 10 PhD students of the Swiss Federal Institute of Technology in Zurich. We aimed to identify possible problems in terms of clarity, accuracy, and relevance for assessing health-related measures that are specific for COPD. Based on the feedback, a few changes were made to improve the wording of the questions and the order of the items.

For the main analyses and for assessing the various aspects of the two interaction styles, we gathered basic sociodemographic data (age, gender, and education), health-related data (GOLD COPD level, years since COPD diagnosis, and disease literacy), and interaction-related data (willingness to change, relationship quality, intention to continue interacting, and working alliance).

Preferences for the paternalistic, informative, deliberative, or interpretive style are depicted by gender, age, educational levels, and disease severity. Out of 181 participants who started the survey, only 139 participants who completed the CA preference task were included in the final sample of Study 1. Moreover, 22 additional participants were excluded because they reported not having COPD. The final sample consisted of 117 participants with a mean age of 65.7 years and a mean GOLD classification value of 2.87. Of the 117 participants, 66 were male. Descriptive statistics can be found in Table 1. R scripts for all tables can be found in Multimedia Appendix 6. Missing data were dealt with by list-wise deletion because of the small number of participants having missing values for a variable of interest.

We conducted hierarchical multiple regression modeling to predict participants’ self-reported interaction quality with the CA, that is, willingness to change, relationship quality, intention to continue interaction, and working alliance, based on (1) the type of CA (paternalistic/deliberative), (2) patients’ demographics (age, gender, and education), and (3) COPD-related measures (GOLD, COPD disease literacy, and experience with COPD). Each outcome was predicted in a three-step procedure. The first block added to the model was the CA type (labelled “model 1”). The second block contained the CA type and participants’ demographics (labelled “model 2”), and the third block consisted of the CA type, participants’ demographics, and COPD-related measures (labelled “model 3”). Relationship quality, intention to continue interaction, and working alliance were measured on a metric scale. We calculated hierarchical linear regressions for those outcomes and logistic regression for the binary outcome willingness to change. As in Study 1, missing data were dealt with by list-wise deletion.

In summary, we found evidence that age and experience with COPD inform participants' preferences for a deliberative or paternalistic interaction style of the CA. Older participants reported better interaction-related outcomes when interacting with a paternalistic CA, whereas younger participants reported that when interacting with a deliberative CA. Participants with fewer years of personal experience with COPD reported better interaction-related outcomes when interacting with a paternalistic CA, whereas those with more years of personal experience reported that when interacting with a deliberative CA. We did not find evidence for gender, disease level, and disease literacy. Irrespective of the CA type, disease literacy positively predicted both dimensions of working alliance, and participants with fewer years of experience with COPD reported higher perceived support in goal agreement by the CA and were more motivated to continue the interaction with the CA. A more severe disease level was associated with higher motivation to continue the interaction with the CA. Participants who worked with a paternalistic CA were more likely to change their behavior based on the intervention. Thus, our results indicate that knowing the age and years of experience of a patient with COPD can help to decide which interaction style to choose for the patient in order to increase interaction-related outcomes for the patient.

In this work, we investigated the preferences of patients with COPD for specific interaction styles of health care CAs. The interaction style between health care professionals and patients has long been recognized as a key success factor for chronic disease management and final treatment success [44,45]. Given the rising number of chronically diseased patients and the associated financial and personal burdens, CAs represent scalable and ubiquitous digital tools to support chronic patients and relieve human health care professionals. A systematic approach for inducing two specific interaction styles into CAs in a health care setting has previously been developed and validated [29].

In our first study, we determined baseline differences for preferred interaction styles between 117 COPD patients and CAs. We showed that differences in preferences for specific interaction styles for the interaction between chronically diseased patients and CAs exist. In our second study, we explored the patterns of preferences for two specific interaction styles in 123 COPD patients. We found evidence that younger patients reported better interaction-related outcomes when interacting with a deliberative CA, while older COPD patients reported better interaction-related outcomes when interacting with a paternalistic CA. Additionally, COPD patients with longer personal experience with the disease reported better interaction-related outcomes when interacting with a deliberative CA. Moreover, COPD patients with lower COPD disease literacy reported better interaction-related outcomes when interacting with a paternalistic CA. Gender, disease severity, or disease literacy did not affect any preferences for specific interaction styles. Nevertheless, we found evidence that disease literacy, in general, positively predicted both dimensions of working alliance independent of the interaction style.

This paper is especially important for the development of personalized CAs in the context of digital health care, with a focus on chronic diseases. To our knowledge, this is the first investigation that systematically evaluated the preferences of chronic patients for their interaction style with CAs. While CAs have primarily been developed portraying a single interaction style for every human counterpart interacting with them, medical research has long stated the crucial importance of deploying personal interaction styles in order to improve patient satisfaction [40], treatment adherence, and final treatment outcome [46,47]. Addressing the gap in the literature regarding differentiated and personalized interaction styles for patient-CA interactions and adding to the growing body of literature on CA personalization [10,11], this paper now provides the first evidence that chronic patients report better interaction-related outcomes when interacting with CAs that display personalized interaction styles.

The findings of these two studies further inform the pairing of chronic patients to CAs that are personalized at the level of their interaction style. While medical research postulates the relevance of five factors (gender [23,24], age [23,24], disease level [26,48], personal experience with a disease [49], and disease literacy [50,51]) that influence the patient-physician interaction, we showed that not all of these aspects are similarly important when it comes to coupling chronic patients with CAs. Our first results indicate that knowing the age and personal disease experience of the patient is sufficient to decide which interaction style results in increased interaction-related outcomes for the patient at hand. While these are the first results from a restricted sample in an experimental setting, the implications for CA deployment could be significant. Especially from a privacy perspective, the findings would reduce the amount of personal patient data needed to achieve an advantageous CA-patient allocation, as only these two data points can be gathered instead of obtaining a whole plethora of personal data. In addition, these two data points can be easily collected at the start of the patient-CA interaction, without any specific (medical) knowledge needed for assessing them. This could reduce the work of health care professionals, whose time is limited and costly, as the best-fitting CA could be allocated based on responses to a simple digital questionnaire at the beginning of the patient-CA interaction. Notwithstanding these potential possibilities, more research needs to be done to be able to robustly understand patients’ preferences in detail so that industry applications can be developed and reliably used.

This paper has several strengths. First, we followed a two-step approach by determining baseline differences of COPD patients’ preferences for their interaction with a CA in Study 1 and subsequently expanding the findings to Study 2, a between-subject online experiment. Second, we deployed a systematic and validated approach for inducing two specific interaction styles into the patient-CA interaction [29]. Third, we continuously ensured an objective approach by integrating both theoretical knowledge and applied medical expertise into the development of the experiment. We did this by closely collaborating with medical professionals. Here, we worked together with not only medical experts on COPD (the chronic disease subject of this paper), but also health care professionals from other fields to reduce the risk of bias. In addition, we integrated the views of both senior and novice health care professionals to reflect traditional paternalistic-based training and current shared decision making–based training. Fourth, we focused on investigating the preferences of a specific target population, that is, patients with COPD. This focus on a relatively homogeneous patient group allowed us to delve into depth and gain a profound understanding of their preferences.

This work also has limitations. First, we tested baseline differences between all four major interaction styles in Study 1 (a paper-pencil study), but we only tested personal preferences for the deliberative and paternalistic interaction styles in Study 2 (online experiment). These are the two interaction styles where a systematic and validated approach for inducing the specific interaction style into the CA-patient interaction for a digital online setting exists. They further represent the start point and endpoint of a hypothetical ethical development process of a model patient-physician interaction [52]. Nevertheless, the results from Study 1 indicated that some patients might have personal preferences for other interaction styles than these two. The preferences for these interaction styles need to be investigated by future research. In addition, we could see differences in the preference allocation between the studies. While older participants in Study 2 preferred a paternalistic CA when evaluating relationship quality, 57% (4/7) of participants who chose the paternalistic style in Study 1 were in the youngest of the tested age ranges. This difference could be assumed to have originated in the respective study setup. Participants could choose from one single interaction snippet of the four main different interaction styles, and participants in Study 2 had a real interaction with a CA, but only the deliberative or paternalistic CA. As such, further research is needed to robustly understand patients’ preferences, especially when it comes to a real interaction with a CA. Second, the study population only included German-speaking patients based in one country. It could be that other languages or regions influence different interaction style preferences and personal requirements. Third, disease-related patient inputs (eg, GOLD status and years of diagnosis) were self-reported and hence not verifiable. Additionally, not all patients reported their GOLD status. Fourth, we only modeled the first part of an initial interaction between a patient and a CA. In reality, patients would need to interact over a prolonged period of time when a CA supports their chronic disease management. Fifth, we used a paper-based snippet of a hypothetical patient-CA interaction in Study 1 and a prescripted and rule-based CA in Study 2. While these two approaches were necessary because of the study condition in Study 1 (we sent a physical letter to the patients) and to control the experimental condition in Study 2, both approaches have their limitations when it comes to emulating a naturalistic patient-physician interaction. We are aware of the increasing number of artificial intelligence (AI)–based CAs [4] as well as voice-based CAs for health care purposes [53]. We believe this could be an interesting path for future research in this context of personalized patient-CA interaction styles. AI-based CAs could not only interact in a more naturalistic way by utilizing unconstrained written, spoken, or visual input [4], but also further adapt dynamically to personal developments, for example, the level of disease literacy of their human users.

In general, we advise future research to put a stronger focus on the investigation of patients’ personal preferences for specific interaction styles when interacting with CAs based on the long-known importance of this factor in the human patient-human physician context. In detail, we see specific possibilities for future research motivated by the limitations of this study and as an extension of it.

First, we advise future research to expand and test the used systematic approach for inducing two specific interaction styles to more interaction styles. As discussed above in the Strengths and Limitations subsection, we could see variations in preferences when patients could choose from four interaction styles (but only having one written interaction snippet to choose from) and a setting where they could interact for longer with either a deliberative or paternalistic CA. This could provide valuable insights into how many different interaction styles for patient-CA interactions are needed. In addition, we recommend future research to study the development of patients’ preferences over time. It would be highly relevant to determine whether such preferences stay stable or dynamically evolve over time. Second, we suggest the development and evaluation of CAs in other languages besides German and in more diverse geographical settings to investigate the effects of language and regional specificities on patient-CA interaction styles. Third, we recommend focusing on the preferences of patients suffering from different diseases (both acute and chronic). We suggest focusing on differences within as well as between diseases to understand any influencing factors of the medical condition at hand in detail. Fourth, future research could expand our experiment and develop a more extended interaction between patients and interaction style–personalized CAs. This could bear interesting findings to further understand dynamic developments of personal preferences for interaction styles between patients and CAs. Fifth, we believe the development and implementation of AI-based CAs that are able to interact more naturally and adapt dynamically to the patient at hand could yield interesting results in the field of patients’ personal preferences for their interaction with CAs.

The interaction style between patients and physicians is recognized as a critical parameter for patient satisfaction, treatment adherence, and subsequent treatment outcome and, as such, also plays a paramount role for chronic disease management. So far, CAs as ubiquitous and scalable digital tools have mainly utilized a single interaction style for every patient, thus ignoring the relevance of personalized interaction styles. In this paper, we showed that chronically diseased patients exhibit preferences for different interaction styles when conversing with a digital health CA. Our results provide evidence that patients’ age and personal experiences with the disease inform their preferences for a specific interaction style. Hereby, this work provides insights into the rising trend of personalized CAs in health care. We envisage a future where every chronic patient gets paired with a CA exhibiting the right interaction style at the right moment and dynamically adapting to the needs of the patient, thereby allowing for a satisfying and fulfilling patient-CA interaction that supports the best possible treatment outcomes and disease management.