Worldwide, health experts are expecting an increasing shortage of medical personnel within the next few years [1-3]. Especially in rural areas, access to medical care is expected to decline [4]. Thus, it will become increasingly important for patients to inform themselves about their medical condition and to take the right steps based on this information. Symptom checkers (SCs) support this process of self-management [1]: these systems are defined as patient-facing decision support systems—typically using deep learning (eg, recurrent neural networks), Bayesian networks, or rule-based algorithms [5-8]—that enable laypersons to get preliminary diagnoses and recommendations for the level of care to seek based on their symptoms [8]. Like other sources of online health information [9], SCs provide health information in a convenient and scalable way.

One possible risk that emerges from the increasingly widespread usage of SCs is that they amplify existing health inequities. It is already known that racial/ethnic minorities, rural residents, and persons with a low income experience worse health care than others [10,11]. Ahmed et al [12] found that sociodemographic determinants, such as age, gender, education, and income, have an influence on the usage of electronic devices to access health information. Conversely, some authors suggest that mobile decision support systems could alleviate access problems as they provide accessible and easy-to-understand health information [13,14]. To prevent health inequities and to maximize the potential benefits of SCs, it is crucial to understand the factors that contribute to people using and not using them.

SCs commonly offer 2 features to their users: (1) Users can improve their self-diagnosis by obtaining a rank-ordered list of the most likely diagnoses, and (2) SCs can be used to assist with triage decisions. This means that they advise patients on whether it is necessary to seek care at all and, if so, how urgently (eg, instantly or within some days) they should visit which health care facility (eg, emergency department or general practitioner) [15]. Especially, the accuracy and safety of this triage function is an ongoing topic of concern for both patients and health care professionals. Their performance seems to be mediocre on average, with high variability between them [16,17].

Prior related studies have mainly focused on the effects of sociodemographic factors on the intention to use, trust in, or adherence to decision support systems in general. Age, gender, the level of education, and several individual factors have been already found to influence the interaction with and usage of SCs specifically. These findings will be briefly summarized next.

Users of SCs tend to be younger (with a mean age of about 40 years), and the willingness to use SCs seems to decline with increasing age [18-22]. However, younger users seem to find SCs more useful, but older users have been found to be more likely to recommend them [19,23].

Users also seem more commonly to be female (estimates range from 62% to 85%), although gender has also been reported to not impact the willingness to use such tools [19,20,22].

Lastly, users of SCs tend to have higher levels of formal education, which is associated with an increased likelihood of searching for diagnoses online [8,20,24-27].

In addition to sociodemographic factors, previous studies have found that people with higher eHealth literacy are more inclined to use mobile health apps in general and that a lack of computer literacy seems to be one of the greatest barriers to using SCs [22,28,29].

Another relevant interindividual trait is trust, as incorrect diagnoses and increased anxiety are the main concerns when using SCs [22,30]. Thus, (the propensity to) trust is found to impact the interaction with SCs as well [31].

Most quantitative findings on SC users stem to date from studies investigating samples using a single SC only. Although users generally perceive SCs as useful [32,33], a lack of awareness of these tools has been discussed as a potential barrier to broader adoption [34].

The aim of this work is to refine our understanding of how sociodemographic and interindividual characteristics influence the awareness, usage, and perceived usefulness of SC apps. In contrast to most of the previous research that is based on UK, US, or Canadian users of specific SCs (which might not be representative for SC users in general), we investigated a representative sample of German-speaking internet users. Building upon the existing literature, our paper focuses on factors previously shown to be relevant in other countries. Unlike most previous studies on SC users, we sampled not only users but also nonusers of SCs. This broader sampling approach allowed us to address more questions, for example, investigating the potential reasons for unequal usage of SCs across sociodemographic factors. Not being limited to the user group of a specific SC, our approach also yielded more generalizable findings concerning factors influencing usage in the population.

We conducted a cross-sectional online survey among German residents between July 15 and 26, 2022. Our aim was to sample 1000 participants. No prior sample size calculation was conducted, as the sample size was ultimately determined by the available budget. Considering that some participants were expected to respond incorrectly to control questions, we planned to oversample by 10% (resulting in about 1100 participants). Stratified random sampling was used to sample participants using the ISO 26362–certified sampling provider Bilendi/respondi [35]. We stratified the sample by gender, federal state, income, and age to reflect the German population [36]. Bilendi/respondi was selected because it is a commercial provider that is certified, offers panel surveys with stratified random samples, and has been used by other authors for surveying nationally representative samples in biomedical research [37-39].

The study included participants who were at least 18 years old, and excluded underage participants and those who refused to consent. Moreover, we excluded data from analysis if a participant answered one of the embedded control questions incorrectly. Upon completion, participants received a payment of EUR 1.00 (US $1.08) for their participation.

This study was approved by the Ethics Committee of the Charité – Universitätsmedizin Berlin (EA4/018/22). Prior to enrollment, participants provided informed consent and volunteered to take part in the survey. The study was conducted and reported according to the Checklist for Reporting Results of Internet E-Surveys (CHERRIES) guideline [40].

We developed a survey in German and administered it as an online questionnaire using the Unipark EFS Survey [41]. The authors and a sufficiently large convenience sample (N=9) [42] from the authors’ personal and professional network conducted pretests of the survey to ensure comprehensibility of the questions and usability of the online survey and identify any technical issues. We rearranged the survey sections and simplified the language of the questions following the pretest. All collected data were stored in EFS Survey accessible only to the authors. Participants filled out the survey remotely upon an invitation from the sampling provider, and they were prevented from participating more than once by assessing their pseudonymized ID assigned by Bilendi/respondi. The assigned ID was not shared with the authors.

Overall, the survey had 4 sections: (1) sociodemographic and interindividual characteristics, (2) questions about previously received diagnoses and medical care, (3) the usage of technology and health apps in general, and (4) the usage of SC apps in particular.

The questions about demographics and characteristics included age, gender, the level of formal education, the federal state (Bundesland) participants reside in, the municipality size, the disposable income (assessed using the Organisation for Economic Co-operation and Development [OECD]–modified scale [43]), their migration background, and their self-efficacy (measured using the Allgemeine Selbstwirksamkeit Kurzskala, ASKU, [44]).

In the second section (diagnoses and medical care), we asked participants to fill out the Minimum European Health Module (MEHM) [45] to rate their self-perceived health, including activity limitations and chronic morbidity. We also presented them with a selection of different diseases (for which officially approved health apps are available in Germany), and they could choose all diagnoses that applied to them, including depression, panic or anxiety disorder, and chronic pain. Health care usage was assessed by asking for the insurance type (statutory, private, other, or none), whether they have a permanent general practitioner (yes/no), how often they visited a general practitioner within the last 12 months (open numerical text field), whether they are undergoing psychotherapy, and whether they have been hospitalized as an inpatient in the past 12 months (yes/no). We included psychotherapy in our definition of health care usage as the German statutory health insurances cover mental health services and many digital health apps, including SCs, are sought for psychiatric or psychosomatic issues.

In the third section, we asked participants how often they use the internet (several times a day, once a day, several times a week, several times a month, or less than once a month) and we assessed their affinity for technology interaction (using the Affinity for Technology Interaction [ATI] scale [46]). We also assessed their health app usage by asking whether they generally use health apps (yes/no).

In the last section, we gave participants a description of SCs and asked them about SCs using 3 steps: First, we asked whether they know about SCs (yes/no). If they affirmed, they were asked whether they had used them before (yes/no). If they did, we asked them to rate their usefulness on a 5-point Likert scale with the levels 1=“not useful at all,” 2=“rather not useful,” 3=“sometimes useful, sometimes not,” 4=“rather useful,” and 5=“very useful.”

We embedded 2 control questions in the questionnaire asking participants to select a particular answer option to a mock question (eg, “Please select ‘does not apply’”).

The data were analyzed exploratively—all values (including P values along with other measures of statistical inference) should therefore be interpreted in a hypothesis-generating manner. We included robustness checks (see Multimedia Appendix 1) adjusting P values for multiple testing using the Benjamini-Hochberg procedure to verify that the results remained valid after correction. Our significance level was set to .05. For income, we controlled for unreasonable data by excluding outliers (defined as the top 2.5% and the bottom 2.5% income).

First, we compared those aware of SCs and those unaware of them. Second, we compared SC users with nonusers, (1) in all participants and (2) in a subset of those being aware of SCs, to assess factors that may contribute to the willingness to use. Lastly, to assess factors influencing the perceived usefulness of SCs, we included only data from participants who had used SCs before. We divided usefulness into “not useful” (indicated by selecting “not useful at all” or “rather not useful”), a middle category (“sometimes useful, sometimes not”), and “useful” (indicated by selecting “rather useful” or “very useful”).

We conducted comparative analyses of these subsets by comparing all characteristics using summary statistics (mean and SD for metric variables, absolute numbers, percentages, and 95% CIs for binary, multinomial, and ordinal variables). For inferential analyses, we used Welch t tests (for metric variables with groups of different sample sizes); chi-square tests (for binary and multinomial variables), or Fisher exact tests when any cell contained less than 5 observations; and Mann-Whitney U tests (for ordinal variables). To quantify effect sizes, we used Cohen d for t tests, the phi coefficient (φ) for 2×2 chi-square tests/Fisher exact tests, Cramer V for more than 2×2 chi-square tests/Fisher exact tests, and the Glass rank biserial correlation coefficient rg for Mann-Whitney U tests. Further, we visualized selected characteristics in raincloud plots [47].

In Multimedia Appendix 1, we provide the following additional analysis: To explore perceived usefulness in more detail, we correlated usefulness with other binary (point-biserial correlation) and continuous variables (Pearson correlation) and visualized it in a heatmap using 1 column of a correlation matrix.

We used R version 4.1.2 [48] and the tidyverse packages [49] to manipulate and analyze the collected data. We also used the packages rstatix [50] to compute summary statistics and correlation matrices; DAAG [51] to assess the variance inflation factors; ggdist [52] and gghalves [53], in addition to ggplot2 [54], for data visualization; and DescTools [55] to compute CIs. For effect size computation, we used rstatix (Cohen d), rcompanion (Glass rank biserial correlation coefficient rg) [56], DescTools (Cramer V), and the psych package (φ) [57].

To make the Results section more concise, mostly statistically significant results are reported in tables summarizing group comparisons of participant characteristics. Detailed tables outlining all findings and inferential statistics are provided in Multimedia Appendix 1.

A total of 1555 people accessed the survey, of which 400 (25.7%) did not complete it. Moreover, 4 (0.3%) participants were screened out: 2 (50.0%) for indicating to be younger than 18 years and 2 (50.0%) for not providing informed consent. We excluded the data of 67 (4.3%) participants due to incorrect answers to at least 1 of 2 control questions. As a result, we included the data of 1084 (69.7%) participants in our study. About 1 in 6 participants (177/1084, 16.3%) indicated having previously heard about SCs. Of these, 40.1% (71/177)—equating to 6.5% (71/1084) of the total sample—reported having used an SC at least once before. Participants’ characteristics (with all collected variables) are shown in Table 1.

Participants aware of SCs were commonly younger (mean 38.8, SD 14.6 years, vs mean 48.3, SD 15.7 years; P<.001), were more commonly female (107/177, 60.5%, vs 453/907, 49.9%; P=.015), had higher formal education levels (eg, 72/177, 40.7%, vs 238/907, 26.2%, with a university r college degree; P<.001), and on average reported a higher net household income (mean EUR 2173.96, EUR SD 992.83 [mean US $2347.88, SD US $1072.26], vs mean EUR 1813.79, SD EUR 865.37 [mean US $1958.89, SD US $934.60]; P<.001) than the remaining study participants. About three-quarters of the participants being aware of SCs (128/177, 74.0%) reported prior experience of using health apps in general in contrast to the remaining participants, of which less than half reported this (389/907, 43.5%; P<.001). They also showed higher scores on the ATI scale (mean 4.17, SD 0.93, vs mean 3.66, SD 0.99; P<.001). We found only little differences between these groups regarding self-reported general health and migration background. A summary of all characteristics and interindividual differences between those who knew SCs and those who did not is provided in Table 2 and Table S1 in Multimedia Appendix 1.

First, we present results comparing users and all other participants, allowing us to compare the characteristics of users and the general public. To assess the inclination to use SCs, we contrasted the characteristics of users and nonusers in the subset of participants who were aware of these tools.

Of the 71 users, 29 (40.8%) considered SCs (rather) useful, while 18 (25.4%) found them (rather) not useful. The remaining one-third of participants (24/71, 33.8%) reported that SCs were sometimes useful and sometimes not useful to them.

Between users considering SCs useful and those who did not (disregarding those who found them sometimes useful and sometimes not), all sociodemographic variables except for age revealed differences; see Table 5. In our sample, males were 4 times (16:4) more likely than females (13:13; odds ratio [OR] 4.0) to rate their experience with SCs as useful. The difference in general usefulness was statistically significant (P=.002). A higher net household income was also strongly associated with usefulness (mean EUR 2591.63, SD EUR 1103.96 [mean US $2798.96, SD US $1192.28], among those considering SCs useful vs mean EUR 1626.60, SD EUR 649.05 [mean US $1756.73, SD US $700.97], among those who did not; P<.001). Additionally, a higher level of formal education was found among participants rating the usefulness of SCs favorably (55.2% vs 22.2% with a university or college degree; P=.016).

Higher self-efficacy scores were also associated with rating SC usage as useful: Visual analysis of the association between self-efficacy and usefulness hinted at a threshold effect; see Figure 2. Although users with a self-efficacy score above 3.5 commonly found SCs useful (25:10), users below this threshold did not (3:8; OR 6.6). Similarly, most participants (9:1) scoring very high (>5/6) on the ATI scale considered SCs useful, while the majority (3:4) of users with a low score (<3/6) did not. Mean and median scores for affinity for technology, however, were similar across these 2 groups (see Figure S1 in Multimedia Appendix 1). Although in lesser magnitude, these observations held when including participants rating their previous experience as “sometimes helpful, sometimes unhelpful.”

Although users rating their experience as useful self-reported higher general health and lower restrictions for health reasons than users with unhelpful experiences with SCs, these findings were not statistically significant and had a small effect size. Rates of all 3 indicators of health care usage (inpatient hospital stay within the past year, number of physician visits within the past year, currently undergoing psychotherapy) were higher among participants considering SCs useful. In contrast, users currently suffering from panic or anxiety disorder more commonly considered SCs not useful than useful (see Table S4 in Multimedia Appendix 1).

Although the usage of other health apps in general was strongly associated with the awareness and usage of SCs, it was not associated with considering SCs useful.

Table 6 summarizes the gender distribution of participants who were aware of SCs, used them, and considered them useful: women seemed to know and use SCs more frequently. However, the proportion of users among those who knew about SCs was similar for men and women, but men found SCs more commonly useful.

We used a sample stratified by gender, state of residence, net household income, and age. However, since we used an online questionnaire, there was a risk of selection bias—choosing people who had a higher affinity for technology than the general population. Franke et al [46] found differences in the affinity for technology between samples recruited online and offline in a validation study. A university and social media sample had a mean affinity for technology of 4.14, while a random sample in German cities (using pen and paper) had a mean affinity for technology of 3.58. Our study sample’s average ATI score was 3.74 and thus slightly higher than expected for a random sample among both “offliners” and “onliners.” As most of our participants (>91.9%) indicated using the internet multiple times a day, we certainly missed the population subgroup with low technological affinity.

Although the sample size was suitable for the aim of this work, the subsets of those aware of and those using SCs were rather small. Thus, we might have missed important associations of a smaller effect size and overestimated the degree of other associations by chance. Due to the nature of exploratory analyses, which investigate a multitude of associations at once, our findings are subject to the multiple testing problem. However, we conducted robustness checks correcting P values. Although we replicated findings from previous studies (eg, users being more often female, younger, and well educated), other results, such as men finding SCs more useful than women, need to be replicated in future studies. Especially, the hypotheses we derived from the presented findings must be replicated in further studies—for example, that a lack of awareness is the driver behind people’s intention to use SCs, rather than sociodemographic differences.

Our data stem from a cross-sectional online survey of the German population. Thus, all the data are self-reported data. Especially concerning mental health, we might have missed participants with a mental illness who are not open to reporting this in a survey.

As a final point, we evaluated only the subjective usefulness of SCs, not their objective usefulness (eg, facilitating safer, more informed decision-making and guiding users toward appropriate health care facilities). SCs were, for example, perceived as more useful by men, but whether they led to better decisions or other favorable outcomes remains unproven.

Our findings hint at a misalignment between factors associated with using and benefiting from SCs. That is, some groups could potentially benefit from these tools but are uninclined to use them or are unaware of them, while other groups could be inclined to use SCs but seem to not benefit from them. Based on these observed misalignments, our data suggest that SCs currently fail to alleviate inequalities in access to health care despite their high availability and convenient service: Users with higher educational levels, income, and self-efficacy are more likely to report benefiting from SCs, while users with lower self-reported health and mental health issues are less likely to do so. Simply expanding the awareness of SCs among the general population to reduce unequal awareness and unequal usage might leave this unequal distribution of benefits intact.

Ultimately, how the outlined misalignments between using and benefiting from SCs will evolve when a greater part of the general population becomes familiar with SCs and decides to use them remains an open question.

Our findings provide some indication that SCs’ full potential—in terms of users considering them useful—might be tapped by the majority and late adopters (ie, those adopting an innovation after the commonly younger and more educated [68,69] innovators and early adopters). However, health variables unrelated to the inclination to adopt technology early—such as the specific medical concern, past medical history, and the context an SC is approached with—are associated with perceived usefulness and might be a bigger factor influencing whether an SC offers its user a valuable service. To get consumers and patients who are more likely to benefit from SCs to use them, they must have a low-barrier point of contact in the patient journey. Thus, integrating them into the standard health care system might prove a more fruitful path forward than simply promoting (or discouraging) the stand-alone use of such apps.