The study was approved by the ethics committee of Dongying People’s Hospital (DYYX-2023-168). Information about this study was explained to the participants before the start of the study, and consent was obtained and signed online by each participant for informed consent. Participants completed a web-based questionnaire. Data were anonymized. Participants who completed the experiment were compensated with gifts of approximately ¥5 (US $1 = ¥7.09).

In this study, we used a DCE to quantify the preferences of a youth population for attributes of an AI-based medication counseling service. DCEs are a widely used technique in health economics [40-43]. DCEs are attribute-based benefit measures that present individuals with hypothetical scenarios. It is assumed that respondents use an AI-based medication counseling service for guidance on their medication behaviors and questions. The service used in a given scenario was characterized by specific attributes; the respondent was repeatedly asked to choose among a multiple-option set based on 6 attributes, and the respondent’s assessment of the service depended on the level of these 6 attributes. In examining patient preferences and the willingness to pay (WTP), DCEs were well placed to capture preferences from the patient’s perspective in order to derive recommendations on how to optimize health services from the patient’s perspective, which can be an important factor in ensuring that optimal health services are provided in a given situation [44,45].

The determination of attributes and levels in DCEs is a crucial step and was performed as follows: first, conducting literature reviews or expert interviews to develop the initial attributes; second, streamlining data to avoid respondent fatigue (with a recommended upper limit of 6-7 attributes; excessive attributes may increase the burden on the respondent, thus affecting the quality of the respondent’s answer); and third, eliminating inappropriate attributes through further literature review and refining their wording [46-48].

In the initial phase of our study, we tackled key issues in pharmacy services by reviewing literature on AI integration in health care and pharmacy, followed by group discussions and in-depth interviews with clinical pharmacists. Subsequently, to deepen our understanding, it was hypothesized that patients’ preference or satisfaction with AI-based medication counseling services would be influenced by specific attributes identified in this experiment, and based on this hypothesis, 15 respondents without a medical-related background were selected to explore their primary needs in AI-based medication counseling services. Finally, after consulting with relevant experts, respondents ranked their preferences for the identified attributes.

Based on this comprehensive process, 6 key attributes were identified for our experiment: (1) granularity, (2) linguistic comprehensibility, (3) symptom-specific results, (4) access platforms, (5) content model, and (6) costs. The levels for each attribute were assumed and set in a specific order; for instance, the symptomatic nature of the results ranged from 60% to 100%. Table 1 lists each attribute and its corresponding levels. The above 6 attributes were combined with AI, which can mimic human decision-making to communicate with patients and provide general or refined medication counseling services. The patient’s condition can be evaluated by processing large amounts of complex data, enabling the creation of personalized medication plans, including audio or video outputs based on patient needs. Detailed explanations of the attributes and levels are shown in Multimedia Appendix 1.

In the design of DCEs, 2 basic principles were observed: balance and orthogonality [49]. Balance indicates that each attribute level occurs with equal frequency in the attributes, for example, level 1 in attribute 1 equals to level 2 in attribute 2. Orthogonality indicates that each pair of levels occurs with equal frequency in all attribute pairs. We used STATA (version 16; StataCorp LLC) for the orthogonality design, which produced 24 choice sets that were randomly divided into 3 versions, each containing 8 choice sets. To prevent participants from having no preferred option when choosing between products A and B, the option “neither” was set. Each choice set had a total of 3 options: offer A, offer B, and neither. Test-retest choice sets were added (ie, respondents were asked to consider the same choice set twice) to check for internal consistency. As a result, each version contained a total of 9 choice sets.

The questionnaire consisted of 2 parts. In the first part, known as the demographic question, we aimed to assign basic information about the respondents: age, gender, highest level of education, income, general place, and status of residence. In the second part, there were 9 tasks (including a consistency check option); an example of the tasks is presented in Figure 1.

The formula for determining the minimum sample size for a DCE [47] was based on the thumb rule of Johnson and Orme [50]: N≥500 c/ta, where c is the maximum number of levels containing the attribute, t is the number of option sets, and a is the number of options in each option set (including the exit option) [51]. In this experiment, c=5, t=24, and a=3. Calculating the sample size according to the above formula gives a minimum sample size of 312 for this experiment.

For this study, in terms of sampling methodology, we proceeded in 3 stages, and questionnaires were collected using the Questionnaire Star platform. In the first stage, we identified 3 cities in Shandong province based on their gross domestic product level and city rank: Jinan (2nd), Dongying (9th), and Tai’an (12th). This allows our findings to cover regions with different levels of economic development as a way to ensure the comprehensiveness and fairness of data collection. In the second stage, we considered young people as the survey population in this study. In a survey conducted in 2023 among Chinese college students (with 7055 valid questionnaires), more than 84.88% of the college students reported that they had used AI tools and that they were interested in emerging technology products; therefore, we controlled the survey location by selecting only 2 neighborhoods and a university to perform the survey. The third stage was to perform convenience sampling. Questionnaire collection was performed through 2 methods by the investigators: (1) professional investigators conducted site visits to designated locations to administer the questionnaires, ensuring strict adherence to the inclusion and exclusion criteria to maintain the quality of the data collected, and (2) investigators also posted recruitment information on a web-based platform. Participants who met the location and eligibility criteria were subsequently contacted online by the investigators, who provided detailed explanations regarding the questionnaire and its content. Participants in this study were required to meet the following conditions and sign informed consent: (1) aged between 18 and 44 years [52], (2) can complete the web-based questionnaire on their own or with the help of an investigator, and (3) able to understand the meaning of each entry in the questionnaire. The data collection period was from October to December 2023, and finally, 389 questionnaires were distributed.

To ensure the quality of the responses, we provided centralized training to the enumerators prior to the official launch of the survey, including a detailed explanation of DCE and how to communicate with the respondents. Only those who passed the training were allowed to join the survey team. When the survey was conducted, the investigator explained the questions to the respondents one by one, and some respondents encountered difficulties in completing the questionnaire on their own, for example, operational inconvenience (login restriction, poor network causing slow loading, too many questions making the respondents fatigued, outdated functionality in respondents’ cell phones to load questionnaires and fill them out online). In such cases, the investigator verbally explained the questions to the respondents emphasizing the important parts in order to assist in completing the questionnaire and to ensure the authenticity and validity of the questionnaire. Each respondent was informed that the questionnaire was completely confidential before taking the survey. Questionnaires that met the following exclusion criteria were discarded: (1) those completed in less than 400 seconds, (2) those that failed the consistency check, and (3) those in which all options in the DCE section were identical. After applying these exclusion criteria, a total of 340 valid questionnaires remained, resulting in a response rate of 87.4% (340/389).

STATA version 16 was used for the statistical analysis, and we used a random parameter logit model to assess people’s preferences in choosing an AI-based medication counseling service. During the model estimation process, we utilized 500 Halton draws for simulation to enhance the accuracy and efficiency of the integration. Furthermore, in the random parameter logit model, we assumed that the parameters have a latent normal distribution to capture the heterogeneity in preferences among individuals. The advantage of the random parameter logit model is that it allows the parameters to vary randomly across individuals, portraying individual heterogeneity through the distribution of the parameters and better capturing the variability of preferences between individuals. We assumed that all coefficients of the attribute level were random distribution and freely correlated. Cost was treated as a continuous variable in the model to facilitate the calculation of WTP, that is, the relative monetary value that the youth population assigns to various aspects of AI-based medication counseling service. The ratio of regression coefficients of different attributes indicates marginal substitution rates. The ratio of other attributes’ coefficients to the cost coefficient reveals the population’s WTP, applicable across various contexts to assess preferences for different attributes and levels [53-55]. A monetary evaluation of the attributes was conducted as follows. The WTP for a single level was calculated by dividing the estimated β of the attribute level by the negative of the estimated costs β. A positive estimate means that people were willing to pay for that level, while a negative estimate indicates unwillingness to pay. Although cost was treated as a continuous variable for WTP calculations, it was considered a categorical variable for the calculation of relative importance (RI). In order to determine each attribute’s RI in relation to all other characteristics or levels, the marginal utility range of each attribute was divided by the total of all attribute ranges. Moreover, we performed subgroup analyses by income and education levels and calculated WTP values by using a mixed logit model to evaluate the influence of demographic factors on cost sensitivity.

Among the 340 participants included in this study, 161 (47.4%) were males and 179 (52.6%) were females; 262 (77%) participants were in the age range of 18-24 years. Approximately 80.6% (274/340) of the participants were not living alone and 79.4% (270/340) were unmarried. Approximately 85.9% (292/340) of the participants had tertiary education or above. Specific data are shown in Table 2.

According to the mixed logit model results shown in Figure 2, all attributes were statistically significant, and there were varied preferences across attribute levels. We found that users were most concerned about whether the service was free or not. Preferred service attributes included refined medication counseling services (with general medication counseling services as the reference; β=0.51; P<.001), easy-to-understand language (with difficulty in understanding language as the reference; β=0.81; P<.001), 100% symptom-specific results (with 60% symptom-specific results of the results as the reference; β=3.18; P<.001), 80% symptom-specific results (β=1.17; P<.001), 90% symptom-specific results (β=2.06; P<.001), through WeChat applets (website access as reference; β=0.66; P<.001), and content model in video form (text as reference; β=0.86; P<.001).

The 100% level of symptom-specific results was strongly preferred by the respondents, and the same answer was obtained during the face-to-face conversations with the respondents. These coefficients reflect the RI of each attribute level to user preferences; P values test the statistical significance of the coefficients; and standard errors and 95% CIs provide credibility to the coefficient estimates. An in-depth analysis of these preferences can provide guidance for the design and optimization of AI services.

Among the 6 attributes in this study, the most important factor influencing the young population to make a choice was the symptom-specific results (RI=36.99%), followed by cost (RI=30.24%), access platforms (RI=6.53%), content models (RI=12.25%), linguistic comprehensibility (RI=9.13%), and granularity (RI=4.87%).

As shown in Table 3, participants were willing to pay an additional ¥3.84 for fine-tuned medication counseling. For linguistic comprehensibility, participants were willing to pay an additional ¥6.14 for clearer language delivery. For symptom-specific results, participants were willing to pay ¥24.01 more for 100% results, ¥15.55 more for 90% results, and ¥8.82 more for 80% results. For access platforms, respondents were willing to pay an additional ¥4.99 to access the counseling service via WeChat applet. For the content model, participants were willing to pay ¥6.54 more for video-based content delivery. We evaluated the influence of demographic factors on cost sensitivity by using subgroup analyses, with WTP estimated based on the results. Specific results are shown in Multimedia Appendix 2.

In this study, we analyzed Chinese youths’ preferences for AI-based medication counseling services, and we identified 3 standout attributes: symptom-specific results, content model, and linguistic comprehensibility [56]. Respondents showed a strong preference for highly symptom-specific results, which were correlated with higher WTP. Based on our survey data, we suggest that AI-based medication counseling services should prioritize the enhancement of symptom-specific outcomes, which may help improve user satisfaction and WTP.

A study [57] has shown that AI technology can improve diagnostic accuracy in certain areas, but the effects may vary in different contexts. Regarding content models, digital formats such as voice and video are preferred over text probably because they are more engaging and easier to comprehend [58]. Text explanations can make patients feel passive because of varying levels of understanding. Video explanations enriched with health science information could improve patient attention and health literacy [58-60].

For linguistic comprehensibility, plain language is preferred, with clear and understandable explanations being critical. AI’s lack of emotional response in counseling was noted as a concern. AI emotion recognition has the potential to improve communication in the future, but more research and data are needed at this stage to validate this hypothesis—maybe the interaction could be made more empathetic and reassuring.

Respondents demonstrated a strong preference for WeChat applets over websites as access platforms because of their simplicity and user-friendliness. Further research could explore the factors underlying this preference. WeChat applets offer smooth, user-friendly access, while web pages may contain disinformation and are less convenient to save and revisit. WeChat applets are increasingly popular due to their simplicity, lack of installation requirement, large user base, and low development costs [61]. Preference for WeChat applets and easy-to-understand language may reflect the preferences of the respondents for simple everyday use across a number of previously indicated attribute levels.

In terms of medication guidance, respondents expect AI to provide more detailed and refined advice compared to traditional methods. However, the one-question, one-answer format of AI services can limit the patients in asking follow-up questions or explore further, thus causing them to overlook important issues. Comprehensive counseling addressing all aspects of drug use is essential.

Finally, cost was also a key factor considered by the respondents, with a significant preference for low-cost or free medication counseling. This is consistent with the results of previous related DCE studies for pharmacy services [62,63]. Respondents were willing to pay ¥24.01 for 100% symptom-specific results. This implied that the respondents were not concerned about the high cost of counseling services for highly symptomatic patients. It is possible that the respondents anticipate receiving more precise counseling outcomes and, thus, a better quality of health care. According to the subgroup analysis results, high-income individuals demonstrated greater payment capacity and were more willing to pay a premium for higher-quality health care services. Additionally, high-income groups tended to place more value on personalized and precise services. Similarly, individuals with higher education levels often had greater demands for information accessibility and comprehension, reporting easy-to-understand language and video content models more effective in meeting their needs. Higher income and education levels were associated with a greater WTP for premium services and a higher value placed on personalized services.

Symptom-specific results and costs formed the largest portion of RI, and we believe that symptom-specific results are directly related to treatment outcomes and patient satisfaction. Of the attributes we provided, only symptom-specific results and costs were quantified by specific numbers, which allowed respondents to compare and assess the extent of services more intuitively when making choices. This suggests that young people are likely to combine economic pragmatism, that is, finding the optimal solution with limited resources when making decisions.

A well-designed policy framework is essential for the successful integration of AI-based medication counseling services into health care systems. Policymakers need to consider the key findings regarding symptom-specific results, content model, linguistic comprehensibility, and the need for cost-effective, user-friendly platforms.

This study has several limitations. First, this study was conducted on a specific population: young adults aged 18-44 years. Future product designs should expand the scope of this study, considering the perspectives of various demographic groups to maximize the benefits of AI-based medication counseling service. For example, studying preferences among older populations or testing AI-based services with different health conditions could provide a broader view of AI’s role in medication counseling. Addressing the specific needs of different groups will be the focus of our next efforts. Second, this study uses convenience sampling in the third step of sampling, and only 1 province was selected to conduct the survey; the data can be used in the future to improve the representativeness of the sample in a wider range of areas by increasing the sample size, employing more random sampling techniques, and extending the data collection period.

This study aims to determine young respondents’ preferences for using AI-based medication counseling services in China. Our findings indicate that the 6 attributes that respondents found most appealing were 100% symptom-specific results, refined medication counseling, WeChat applet platforms, easily understandable content, video content mode, and no payment required. In summary, when young people choose AI-based medication counseling services, policymakers should pay more attention to considering economic practicality and personalized needs when designing services for young people so as to meet their expectations for high-quality and reasonably costly health care services, and better understanding their needs and promote AI-based medication counseling as a way to solve problems, increase engagement and medication adherence, and establish them as the first to benefit from AI-pharmacy integration and improve the safety of medication and medication literacy of users. These findings will provide valuable insights for policymakers to drive the implementation of AI-based medication counseling services.