Online trust theory is based upon many studies that have classified online trust as consumer trust in e-commerce [24,25]. When consumers feel that information is insufficient or asymmetric, trust can be used as a means to alleviate the asymmetry of perceived information, prompting them to make purchasing decisions [26,27]. McAllister divided trust into cognitive trust and affective trust [28]. Cognitive trust is mainly regarded as the consumer’s rational expectation of the ability and credibility of the trusted party and is mostly judged by the objective characteristic evidence of the trusted party's personal behavior and reputation. Affective trust is regarded as a trust attitude, which reflects the consumer’s feelings and self-consciousness toward the trusted party [29-31].

Choice in mHealth can also be regarded as a special online shopping behavior. In mHealth service, if the physician is regarded as the commodity, then the patient is equivalent to the consumer. Therefore, the study of the patient’s choice in mHealth is actually the study of the consumption behavior of this particular consumer group. Asymmetry of information results in trust playing an important role in the physician-patient relationship. At present, physician-patient trust is generally defined from the patient's point of view; it is defined as the patient's trust in the physician's ability to diagnose and treat them and trust that the physician will put their interests first [32]. Therefore, patient trust plays a very important role in constructing the patient choice model. In this study, trust is also divided into cognitive trust and affective trust. In addition, according to the study by Calefato et al [33], the patient's trust in the physician's abilities (ie, physician's professional skills, knowledge, and competence) is defined as the patient's cognitive trust; the patient's trust in the physician's benevolence (ie, physician's politeness, attitude, and willingness to help) is defined as the patient's affective trust.

The SOR model, proposed by Mehrabian and Russell in 1974, illustrates that the external environment influences the individual's attitude or behavior by influencing the individual's mental state. Stimulus refers to the external environmental factors received by the individual; organism refers to the internal mental state of the individual; and response refers to the attitude and behavior of the individual [34]. The model has been widely used in consumer behavior analysis. According to the model, a consumer (ie, organism) will have a conscious or unconscious psychological reaction after receiving an external stimulus, which involves both cognitive and affective aspects, and will then have an internal or external behavioral response to this stimulus [35-37]. In this study, physician information in the mHealth app, Chunyu Doctor, was considered an external stimulus that caused patients (ie, organisms) to generate trust, which in turn allowed them to make their choice of physicians (ie, response).

Physicians are the direct providers of medical services. In the field of mHealth, physician information is one of the most important considerations for patients when choosing medical services. Patients can get detailed information about a physician when browsing the physician's home page. This study extracted physician information from two major information sources: physicians and patients. Accordingly, three kinds of information were identified: physicians’ profile photo information, physicians’ nonprofile photo information, and patient-generated information.

In the past physician-patient relationship studies, information from physicians’ profile photos was often neglected, but the face is an important source of information. Sometimes a photo can promote trust between people [38]. This face-based assessment is very subjective [39], with emotion often generating trust in others, while physicians’ nonprofile photo information is different. For example, physicians in high-level hospitals often have more advanced medical skills and richer medical experience. As well, a physician's title is a reflection of their professional skills, medical experience, and work performance. These personal traits often lead patients to trust physicians at the cognitive level.

In online trust theory, user-generated information can enhance the trust of online sellers, reduce the risk perceived by buyers, and promote transaction behavior [40]. User-generated information in mHealth corresponds to patient-generated information. The main forms of this information are comments, feedback, and scoring, which can reflect online reputation [41]. Patient-generated information is usually the first-hand experience of patients with online consultations. Eysenbach proposed that, in the field of health care, experience-based credibility can be seen as an additional dimension of source credibility [42]. In that sense, similarity between experiences will enhance patients' perceptions of credibility. In addition, some studies have suggested that patient-generated information can reveal physicians' online behavior; reduce information asymmetry between physicians and patients [43]; prevent physicians from exaggerating and misinterpreting their professional ability, service level, and treatment effect; and help patients understand physicians' medical expertise and service attitude [44]. In conclusion, patient-generated information will affect patients' trust in physicians at the cognitive and affective levels, thus affecting their behavior in choosing physicians. Therefore, the following hypotheses are put forward:

Consumer expertise is a key factor influencing a consumer’s choice; it is the knowledge and experience, of which a consumer is subjectively aware, regarding a particular product or service [45]. High-expertise consumers are willing to take the initiative to look for information related to products and to make decisions through careful thinking because of their rich experience. Low-expertise consumers are more passive in searching for information, preferring to rely on marginal information, and are less willing to spend time thinking about new information.

In consumer behavior research, consumer expertise plays a significant role in explaining consumer decisions and responses to products [46,47]. In this study, patient expertise refers to the patient’s knowledge or experience of mHealth services. Since patient-generated information is unique information in mHealth compared with traditional medical services, we believe that patient expertise will play a moderating role between physician information and trust. That is, the higher the expertise of patients, the more they will rely on patient-generated information in choosing a physician; however, low-expertise patients will rely more on the physician’s profile photo and nonprofile photo information. Hence, we hypothesize the following:

Corritore and Wiedenbeck [22] proposed that in online medical care, physician-patient trust is one of the most important factors to ensure the success of online health services. Gefen et al [48] believed that consumers’ intentions to use certain goods or services mainly depended on consumers' cognitive trust and affective trust in external factors. Therefore, in this study, patients’ trust is considered to influence patients’ choice of physician in mHealth. Hence, we hypothesize the following:

Based on our hypotheses, this study illustrates a research model that explains the influence of physician information on the patients’ choice of physician in mHealth, as shown in Figure 1.

The purpose of this study was to investigate the impact that physician information had on patients’ choice of physician in an mHealth platform and to identify the information that played a key role in this choice. Research has shown that people's ability to process information within their short-term memory is limited [49]; therefore, even though the information provided in the mHealth app, Chunyu Doctor, is comprehensive, a patient’s choice of physician is based only on a limited amount of useful information. In previous online medical studies, scholars have often used their own experiences and related theories to select the information they think is more useful, built regression models by crawling data to uncover the relationship between the physician information and the number of physicians’ consultations, and judged the impact of physician information on patients’ choice of physician [50-52]. This method is subjective and it is easy to overlook some potentially critical information. Humans obtain much external information through their vision and the images to which their attention is focused [53]. Through eye-tracking experiments, researchers can capture the information upon which people focus their attention more objectively and accurately. Therefore, in this study, eye-tracking experiments were used to screen physician information that had a key impact on patients’ choice of physician.

The Chunyu Doctor app contains 12 different types of physician information; for the universality of the experimental results, we did not consider the physicians' departments or specific communication between physicians and patients. The types of information we were concerned with were divided into three categories: physicians’ profile photo information, physicians’ nonprofile photo information, and patient-generated information. This information is shown in Table 1.

In this study, the 12 types of physician information (see Table 1) were classified into high-level and low-level information. Hospital, title, and educational background levels were classified by the government (eg, hospitals are classified into Class A tertiary hospitals and general hospitals); because of this, other information levels, except physicians’ profile photo information, were set with reference to the Chunyu Doctor app and all of them were pretested using a questionnaire. The questionnaire showed that the various information levels were significantly different; the specific classification is shown in Multimedia Appendix 1.

Regarding physicians’ profile photo information, based on a study by Ert et al [54], two dimensions were used to distinguish the information level: attractiveness and trustworthiness. A total of 60 participants—28 males (47%) and 32 females (53%)—rated 40 physician profile photos based on their perceived attractiveness and trustworthiness. We used photos of physicians from the official website of Peking Union Medical College Hospital; 20 males and 20 females, aged 30-40 years old, were selected. In the photos, all physicians wore glasses, white coats, and light sweaters or shirts; none had beards, accessories, or visible makeup. All photos were front-facing portraits of physicians smiling. Each participant had to look at all 40 physician profile photos. For each physician's profile photo, participants were asked to answer the following questions: “Do you think the physician is attractive?” and “Do you think the physician is trustworthy?” These were asked in order to estimate the perceived levels of attractiveness and trustworthiness of the profile photos. The questions were scored on a 7-item Likert scale, ranging from 1 (totally unattractive/untrustworthy) to 7 (very attractive/trustworthy); the order in which the photos appeared to participants was random. Eight profile photos were selected based on the photos’ average attractiveness and trustworthiness levels; four photos had the highest and four photos had the lowest average attractiveness and trustworthiness levels, respectively. There was a significant difference between high-level and low-level profile photos (F₁=94.58, P<.001); however, there was no significant difference between the four high-level profile photos (F₃=0.622, P=.60) and the four low-level profile photos (F=0.292, P=.83). In addition, there was no significant difference between male and female physicians (F₁=2.547, P=.11).

In this study, 12 types of physician information were grouped (see Table 1) and divided into three major categories. Each category was classified according to the rated levels (eg, high-level patient-generated information contained high consultation numbers, high praise rates, high satisfaction, and high gratitude expressed; low-level patient-generated information contained low consultation numbers, low praise rates, low satisfaction, and low gratitude expressed). This was used to generate eight mHealth physician home pages. In summary, the eye-tracking experiment was a 2 (physicians’ profile photo information: high or low level) × 2 (physicians’ nonprofile photo information: high or low level) × 2 (patient-generated information: high or low level) within-group experiment, in which eye-tracking data and questionnaire data were collected.

A total of 42 undergraduate and postgraduate students from Beihang University participated in the experiment; participants were publicly recruited via the Chinese social media platform WeChat. In order to avoid gender bias, 21 men (50%) and 21 women (50%), aged 20-25 years old, were recruited. All participants had normal visual acuity or corrected visual acuity, with astigmatism below 200 degrees, and had online shopping experience. Before beginning the experiment, each participant signed an informed consent letter and registered their personal information. After the experiment, participants were compensated with CNY ¥30.

The eye-tracking equipment used in this study was the Tobii T120 eye tracker (Tobii Technology). The eye tracker was integrated into a 17-inch, thin-film-transistor display with a resolution of 1280 × 1024 and a sampling rate of 120 Hz. The instrument has a built-in eye-tracking server; the participants did not need to wear any equipment but did need to watch the experimental material displayed on the monitor. While the participants looked at the screen, the eye tracker automatically recorded eye-tracking data of the participants. Eye-tracking data can be analyzed and extracted using the Tobii Studio software associated with the eye tracker.

In order to simulate the mHealth treatment situation as realistically as possible, the experimental interface was based on the physician’s home page on the Chunyu Doctor app. In order to avoid interference from other information, only the aforementioned 12 types of physician information were included. Each physician's home page was the same size and, as much as possible, the display format of the various types of information was consistent with the original app. The experiment began with an introduction:

During the introduction to the experiment, two unfamiliar types of information were explained, namely, gratitude expressed (ie, offering extra thank-you tips) and physician's topic (ie, number of physician's articles published). Participants were able to view the experimental stimuli (ie, the Chunyu Doctor app) by clicking the left mouse button on the screen. In order to ensure that the participants were able to fully observe the information about the eight physicians, each physician's home page was fixed for 1 minute. The eye-tracking experiment was complete after participants browsed the home pages of the eight physicians. Participants were then asked to complete a questionnaire; participants needed to rate the usefulness of the 12 types of information by rating the following statement for each on a scale from 1 (strongly disagree) to 7 (strongly agree): “I think this information is very useful for me in helping me choose this physician.” At this point, the experiment has been completed.

A total of 272 questionnaires were recovered, of which 254 (93.4%) were valid. The sample distribution characteristics are shown in Multimedia Appendix 2. Among the participants, 44.5% (113/254) were male and 55.5% (141/254) were female. A total of 87.4% (222/254) of the participants were under 30 years old and all of them had online shopping experience. In line with the fact that young people are more receptive to new things, mobile device users tend to be younger and older people mostly use mHealth apps via their children. A total of 42.1% (107/254) had never used mHealth services while 57.9% (147/254) had used them; 17.7% (45/254) had used them one or more times a week. These findings could help explain the patient expertise in mHealth of the study participants. Based on the above analysis, the study sampling was considered reasonable.

The reliability and validity of the measurement instruments were tested [58,59], as shown in Tables 4 and 5. SPSS, version 22.0 (IBM Corp), was used to analyze the reliability of the data and to measure the Cronbach alpha values of each construct. The principal component analysis showed that the Kaiser-Meyer-Olkin value of the sample was 0.904, which indicated that it was suitable for factor analysis. SmartPLS 3.0 (SmartPLS GmbH) was used to carry out confirmatory factor experiments [60] to analyze the validity of the data, including convergence validity and discriminant validity. Table 4 shows that the Cronbach alpha values of all constructs were greater than .8 and that the composite reliability values were also greater than 0.8, which shows that the measured constructs had better reliability. The factor loads of all variables were greater than 0.7. The average variance extraction (AVE) values were greater than 0.5, which indicated that the constructs in this study had good convergence validity [61].

For the test of discriminant validity of the measurement scale, we can compare the square root of each factor’s AVE and its correlation coefficients with other factors [61]. The results of the discriminant validity measurement are shown in Table 5, in which the enlarged diagonal values are the square root of each factor’s AVE. Table 5 shows that the square root of each factor’s AVE is greater than the corresponding correlation coefficients, which indicates that the discriminant validity between the constructs in this study was quite good.

Partial Least Squares Structural Equation Modeling (PLS-SEM) has advantages over Covariance-Based Structural Equation Modeling (CB-SEM) or general linear regression modeling in processing sample data that do not obey normal distribution—physician information is a class variable—involving multiple constructs and small sample size prediction experiments [62]. Therefore, we used SmartPLS 3.0, an analysis software commonly used in partial least squares path modeling, to construct the whole structural equation; we used bootstrapping with 5000 resamples to assess the statistical significance of path coefficients [63].

The results show that 77.1% of the variation in patient choice could be explained by the model (R²=0.771). Table 6 shows that most of the hypotheses in this model were supported. Physicians’ nonprofile photo information had a significant positive impact on cognitive trust. Physicians’ profile photo information had a significant positive impact on affective trust, which supports H1a and H1b. Patient-generated information had a significant positive impact on cognitive trust and affective trust, which supports H1c and H1d. Cognitive trust and affective trust had significant positive effects on patients’ choice of physician, which supports H3a and H3b. However, the hypotheses regarding the moderating effect of patient expertise were only partially supported. Patient expertise had significant moderating effects on the influence of patient-generated information, both on cognitive trust and affective trust. That is, when patient expertise was higher, the level of information generated by patients was higher and they experienced more cognitive and affective trust; thus, H2c and H2d were supported. However, patient expertise had no significant moderating effect on the influence of physicians’ nonprofile photo information on cognitive trust nor on physicians’ profile photo information on affective trust. Therefore, H2a and H2b were not supported.

To sum up, this study showed that in mHealth, physician information affected patients’ choice of a physician through trust. In addition, it should be noted that patients with different levels of expertise will have different experiences. The experimental results showed that patient expertise had no significant influence on the relationship between physicians’ profile photo information and affective trust nor between physicians’ nonprofile photo information and cognitive trust. It may be because physicians’ profile photo information and physicians’ nonprofile photo information are widely known types of information that exist in online medical treatment; no matter how much knowledge and experience patients have regarding mHealth care, it will not change the impact that these two types of information have on patients' choice of physician. However, patient expertise had a significant impact on the relationship between patient-generated information and cognitive trust and on the relationship between patient-generated information and affective trust. The interaction diagrams of the moderating effects are shown in Figures 3 and 4.