Patient preferences and willingness-to-pay for AI-enabled blended type 2 diabetes care by digital experience and socioeconomic status: a discrete-choice experiment in China

Abstract

Objectives

To elicit stated preferences and willingness-to-pay (WTP) for artificial intelligence (AI)-enabled blended care in type 2 diabetes mellitus (T2DM), and to examine preference heterogeneity by digital experience and socioeconomic status (SES).

Design

Cross-sectional discrete choice experiment (DCE).

Setting

12 community health centres in Jiaozuo and Puyang, Henan Province, China. Data were collected between June and August 2025.

Participants

423 adults diagnosed with T2DM for at least 6 months, recruited using consecutive convenience sampling from routine follow-up appointments. Of 769 participants who completed the survey, 346 were excluded following prespecified data quality criteria (retention rate: 55.0%).

Outcome measures

Outcome measures included preference weights and WTP (in Chinese Yuan, ¥) for five DCE attributes: monthly subscription fee, recommendation source, feedback modality, in-person follow-up frequency and expert oversight, estimated using mixed logit models. Simulated uptake probabilities for tailored service packages across four user profiles were computed.

Results

Among 423 participants, 80.6% had never used AI tools. Price was the dominant driver of choice (62.7% relative attribute importance). Profound preference heterogeneity emerged across subgroups: rural residents (n=78) were highly price-sensitive but preferred physician endorsement (WTP ¥17.58 (US$2.55), 95% CI ¥5.98 to ¥29.17); female participants (n=224) valued guideline recommendations (WTP ¥18.45 (US$2.67), 95% CI ¥7.81 to ¥29.40); and diabetes app users (n=34) were the least price-sensitive but showed a negative preference for AI instant feedback, instead preferring human dietitian feedback. Expert oversight carried a consistent negative WTP across all profiles. Targeting tailored service bundles to intended subgroups increased uptake by 8–16 percentage points compared with non-targeted bundles.

Conclusions

A ‘digital experience paradox’ exists whereby digitally experienced users view human interaction as a premium service, while underserved groups rely on specific trust markers such as physician endorsement. To avoid widening the digital divide, AI-enabled blended diabetes care must move beyond standardised models towards configurable, equity-driven service pathways.

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• This study used a discrete-choice experiment with a mixed logit model, enabling quantification of preference weights, willingness-to-pay estimates and simulation of uptake probabilities across subgroups, providing greater inferential depth than attitudinal surveys.

• Prespecified data quality exclusion criteria—including straight-lining, consistency check failure and completion time thresholds—were applied to minimise the influence of inattentive responding on model estimates.

• The inclusion of an opt-out option in each choice task reduces hypothetical bias by allowing respondents to reject all presented alternatives, more closely approximating real-world decision-making.

• Refusals at invitation were not systematically recorded, so the response rate is unknown and non-response bias cannot be excluded; combined with consecutive convenience sampling from two cities in a single province (Henan), the representativeness of the sample to the broader type 2 diabetes mellitus population cannot be verified.

• Attributes and levels were derived from a literature review and a pilot ranking exercise rather than a formal qualitative interview phase, which may not have fully captured all patient-relevant dimensions of artificial intelligence-enabled blended care.

Introduction

The escalating global burden of type 2 diabetes mellitus (T2DM) presents a significant challenge for public health systems, with an estimated 589 million adults affected globally as of 2024.¹ The prevalence has more than quadrupled since 1990,² with the majority of incident cases now occurring in low- and middle-income countries (LMICs).¹ China bears the largest national burden of diabetes: approximately 148 million Chinese adults (11.9% of the adult population) live with diabetes.^{3 4} However, T2DM care capacity remains unevenly distributed. Rural and low-income regions face severe shortages of qualified healthcare providers; in village clinics, only 17.5% of clinicians or village doctors hold a college degree, and treatment and control rates are markedly lower than in urban areas.⁵ These urban–rural disparities in diabetes burden extend well beyond China. According to the 11th edition of the IDF Diabetes Atlas, the global prevalence of diabetes stands at 12.7% in urban areas compared with 8.8% in rural areas—a pattern observed across diverse LMIC settings, from South Asia to Latin America, driven by differential exposure to obesogenic environments, unequal access to quality care and provider shortages in rural communities.⁶ These inequities underscore the urgency of developing scalable, evidence-based approaches to strengthen T2DM management in under-resourced communities.

Artificial intelligence (AI) offers pragmatic opportunities for T2DM care. AI can augment diagnosis, risk stratification and personalised care planning, partly easing workforce constraints.⁷ Reviews suggest that pairing AI with clinician expertise could substantially improve the quality of T2DM care,⁸ and AI is increasingly described as having the potential to transform T2DM management.^{9 10} However, digital health technologies also carry risks. Without careful design, AI may exacerbate digital inequities—disproportionately benefiting younger, affluent and digitally connected patients while excluding those who are older, rural or less digitally literate.¹¹ Indeed, systematic reviews show that while patients are generally receptive to clinical AI, they express strong reservations and prefer expert oversight.¹² Accordingly, if poorly implemented, AI-enabled tools risk widening rather than narrowing existing health disparities.¹³

Despite this, little is known about how patients from different socioeconomic and digital backgrounds value AI-enabled care. Most prior studies have examined general attitudes or adoption intentions,^{12 14} but, to our knowledge, few studies have quantitatively characterised how preference heterogeneity varies by digital experience or socioeconomic status (SES). Understanding these differences is essential for designing blended care models—integrating AI support with in-person follow-up—that are both effective and equitable. Discrete-choice experiments (DCEs) provide a robust stated-preference method to quantify how patients trade off service attributes.15,17 Compared with simpler approaches such as ranking or rating exercises, DCEs more closely mimic real-world decision-making by forcing respondents to choose between fully specified alternatives, thereby yielding preference weights and willingness-to-pay (WTP) estimates that reflect genuine trade-offs rather than abstract importance ratings. By presenting respondents with alternative hypothetical care scenarios, DCEs estimate the relative importance of attributes—such as endorsement source, in-person follow-up interval and pricing—and how these valuations vary across subgroups.^{16 18}

A DCE was conducted in Chinese adults with T2DM to estimate how SES and digital experience relate to preferences for AI-enabled blended care with in-person follow-up. Findings of this study aim to provide evidence to guide equitable, scalable and context-sensitive service design for diverse patient populations.

Methods

Study design

We conducted a cross-sectional study using a DCE to quantify patient preferences for AI-powered T2DM management services.¹⁹ The survey was administered in person by trained enumerators, who explained the study to participants and provided a QR code linking to a web-based questionnaire in Jiaozuo and Puyang, Henan province, China, between June and August 2025.

DCE attributes and levels

In DCE terminology, attributes refer to the defining characteristics of a care service (eg, follow-up frequency), while levels are the specific values each attribute can take (eg, weekly, bi-weekly or monthly visits). Qualitative interviews with patients were not conducted as a formal attribute-elicitation step. Instead, the initial attribute pool was derived from a systematic review of existing literature on digital health and T2DM care preferences,820,25 drawing on studies from multiple international settings, including the USA,⁸ Europe and Asia, to ensure the selected attributes reflected both global evidence and local applicability. From the literature, 11 candidate attributes were identified and subsequently screened for policy relevance and non-redundancy, then prioritised through a pilot ranking exercise with 50 patients in which low-priority or contextually infeasible attributes were excluded; a full list of candidate attributes and the rationale for their selection or exclusion is provided in online supplemental table S1. While qualitative interviews would represent best practice for attribute development, the literature base was considered sufficiently mature and the pilot exercise adequately representative of patient priorities for this population; the absence of a formal qualitative phase is acknowledged as a limitation. The final set of five attributes reflected key policy- and practice-relevant features of emerging AI-based blended care: monthly out-of-pocket subscription fee, basis for recommendation, feedback modality for food diary entries, in-person follow-up frequency and AI-generated outputs subject to expert review (hereafter, ‘expert oversight’). For example, the attribute ‘in-person follow-up frequency’ was operationalised across three levels: every 2 weeks, once a month or once every 3 months. Full attribute levels were specified to represent realistic and policy-relevant variations and are presented in table 1.

Table 1. Attribute-level selection frequencies across all choice tasks.

Attributes and levels	Total number of appearances	Total number of selection	Percentage of selection
Basis for recommendation
Clinical trial evidence	2277	842	37.8
Guideline recommendation	2259	854	37.0
Physician recommendation	2232	797	35.7
In-person follow-up frequency
Every 3 months	2281	910	39.9
Every month	2226	810	36.4
Every 2 weeks	2261	773	34.2
Expert oversight
No	3384	1259	37.2
Yes	3384	1234	36.5
Food diary feedback modality
Instant AI feedback	3384	1292	38.2
Regular dietitian feedback	3384	1201	35.5
Subscription fee
¥0 per month	2243	1897	84.6
¥100 per month	2276	419	18.4
¥200 per month	2249	177	7.9

AI, artificial intelligence.

Monthly out-of-pocket subscription fee was included as a key attribute, given the substantial financial burden T2DM management imposes on Chinese patients. Although China’s basic medical insurance system covers the majority of the population, out-of-pocket expenditures for T2DM-related outpatient care remain substantial: studies in urban tertiary hospitals found that out-of-pocket costs accounted for approximately 46% of total annual outpatient medical expenses.²⁶ For rural patients, reimbursement from the New Rural Cooperative Medical Scheme is predominantly directed towards inpatient costs, with outpatient chronic disease coverage limited to no more than 30% of pooled insurance funds,²⁷ leaving patients with considerable residual costs. In this context, any additional monthly subscription fee for AI-enabled blended care represents a meaningful financial consideration, particularly for lower-income and rural populations.

These three levels of basis for recommendation represent distinct trust signals in the Chinese primary care context: ‘clinical trial evidence’ indicates the recommendation is derived from published clinical research; ‘guideline recommendation’ refers to endorsement by a national or professional body (eg, the Chinese Diabetes Society guidelines); and ‘physician recommendation’ refers to an individual physician’s personal endorsement of the AI service, independent of whether it is guideline-endorsed. In China’s primary care setting, where patient–physician relational trust is strong, but patients may not always be aware of whether individual physician advice aligns with national guidelines, these represent meaningfully distinct trust sources from the patient’s perspective.

Experimental design and sample size

We generated a Bayesian D-efficient design using the Dcreate package in Stata.²⁸ The design was optimised to minimise the standard errors of the parameter estimates. The final design comprised 48 choice tasks, which were divided into six blocks of eight tasks each. Each participant was randomly assigned to one block. Block assignment was randomised at the point of survey entry via Wenjuanxing.

The minimum sample size requirement was determined using the rule of thumb proposed by Johnson and Orme²⁹ for DCEs: N≥500 c/(t×a), where c represents the largest number of levels for any attribute (three levels, as seen in four of the five attributes; see table 1), t is the number of choice tasks (8) and a is the number of alternatives per task (3). This calculation indicated a minimum required sample size of approximately 63 participants for main effects estimation. Our final analytical sample of 423 participants substantially exceeds this threshold, providing sufficient statistical power for the primary mixed logit analysis.

Procedures and participants

Participants were recruited using consecutive convenience sampling from 12 community health centres in Jiaozuo and Puyang, Henan Province. Patients attending the study centres for their routine T2DM follow-up appointments or regular prescription refills were approached by trained enumerators in the waiting areas, and all patients who met the eligibility criteria during the study period were invited to participate. Eligible participants were adults diagnosed with T2DM at least 6 months prior, free from serious complications that could impair independent decision-making capacity or life expectancy. Trained enumerators conducted in-person surveys, explained the study details and provided a QR code for a web-based questionnaire. The questionnaire consisted of demographic questions followed by the DCE section. Participants answered a total of nine choice tasks: eight tasks drawn from their assigned block, plus one repeated task (identical to the first task) appended at the end to serve as a consistency check.³⁰ The repeated task allows identification of respondents whose answers are inconsistent, indicating inattentive or random responding. No financial incentive was provided for participation. Since enumerators approached patients during routine clinic flow and refusals were not systematically logged, we are unable to calculate a response rate at the invitation stage. Consequently, the number and characteristics of eligible patients who declined are unknown. Prior to data collection, the questionnaire was reviewed by all participating nurses, who confirmed that the attribute descriptions and levels were clearly comprehensible to the target patient population.

A total of 769 participants completed the survey. To ensure data quality, we applied prespecified exclusion criteria, removing respondents who exhibited ‘straight-lining’ (selecting the same option for all choice tasks),³¹ failed the consistency check, or completed the survey in less than 40% of the median completion time.³² After exclusions, the final analytical sample comprised 423 participants, a sample size sufficient for robust estimation of main effects and planned interaction analyses based on prestudy simulations. A detailed breakdown of effective questionnaire rates by collection site is provided in online supplemental table S2.

A sensitivity analysis was conducted to assess potential biases arising from heterogeneity in survey effectiveness across sites. The high-effectiveness group was disproportionately composed of urban residents and individuals with higher education and income levels (online supplemental tables S3 and S4). To confirm robustness, we re-estimated the primary model separately in the high- and low-effectiveness subsamples and obtained consistent results (online supplemental table S5 and note 1).

Survey instrument

The survey was administered in Chinese via a web-based questionnaire (Wenjuanxing platform) accessed by participants through a QR code provided by trained enumerators. Before completing the DCE tasks, participants were introduced to the concept of the exercise through a standardised preamble explaining that they would be asked to choose between hypothetical AI-enabled blended care services, that these scenarios were not real service offerings and that there were no right or wrong answers. Each of the five attributes and their corresponding levels was explained in plain language within the questionnaire prior to the choice tasks.

Each choice task presented two hypothetical service profiles and an opt-out option (‘neither of these’), with attributes and levels displayed in a structured table format (see online supplemental figure S1, for an example task). Prior to data collection, the questionnaire was reviewed by all participating nurses at the 12 community health centres, who confirmed that the attribute descriptions and levels were clearly comprehensible to the target patient population. A full copy of the survey questionnaire (original Chinese with English translation) is provided as online supplemental material.

Statistical analysis

Participant preferences were analysed using a mixed logit (MXL) model, which accommodates both observed and unobserved preference heterogeneity.³³ Seven parameters were specified as random, each following a normal distribution: the opt-out alternative specific constant (none), and the effects-coded levels for all four non-price attributes (guideline recommendation, physician recommendation, bi-weekly follow-up, monthly follow-up, expert oversight and AI instant feedback). The subscription fee coefficient and the left-alternative specific constant were specified as fixed (non-random) parameters, consistent with standard practice to ensure a stable price scalar for WTP calculation. Analyses were conducted in R (V.4.5.0; R Foundation for Statistical Computing, Vienna, Austria)³⁴ using the mlogit package.³⁵

The primary model included main effects for all attribute levels, from which the left-and-right bias in choice selection was calculated. To explore heterogeneity, we systematically tested for interactions between attributes and participant characteristics (age, sex, income, marital status, employment, education, residence, AI use, diabetes app use). Following prior DCE methodology,³⁶ we applied a two-stage procedure: (1) testing each demographic variable’s full interaction with all attributes separately, retaining those with p<0.10 (consistent with the threshold used in that reference) and (2) entering all retained terms into a final comprehensive MXL model.

From the final model, we calculated WTP with 95% CIs and simulated uptake probabilities under different market scenarios across four key user profiles: (1) a baseline profile, representing the reference user; (2) a female profile; (3) a rural profile; and (4) an AI-savvy profile (used AI before). WTP for each service attribute was calculated as the negative ratio of the attribute coefficient to the price coefficient. WTP values are presented in Chinese Yuan (¥). The Krinsky and Robb simulation method,³⁷ with 1000 draws from the estimated coefficient distribution, was used to generate the 95% CIs for all WTP estimates. In addition to the MXL model, we conducted a latent class analysis (LCA) as a complementary approach to further investigate the structure of preference heterogeneity. Unlike the MXL model, which assumes continuous preference distributions, LCA identifies discrete latent subgroups of respondents with distinct preference profiles without requiring a priori subgroup definitions. Full LCA results, including model selection criteria and class-specific preference estimates, are reported in online supplemental tables S6, S7, S8 and S9 and note 2. Reporting of this DCE followed the DIRECT checklist³⁸; the completed checklist is provided as online supplemental table S10.

Patient and public involvement

Patients were involved in the design phase of this research. Specifically, a pilot ranking exercise was conducted with 50 patients with T2DM to prioritise and refine the attributes and levels used in the DCE, ensuring the survey content reflected real-world patient priorities. Patients were not involved in the recruitment to or conduct of the study, nor were they involved in the writing or editing of this manuscript. The results of the study will be disseminated to the participating community health centres to be shared with the study participants.

Results

Of the 769 participants who completed the survey, 346 were excluded following prespecified quality criteria, yielding a final analytic sample of 423 (retention rate: 55.0%) (table 2). The sample was predominantly female (53.0%), with a significant proportion aged 60 years or older (44.0%), a common retirement-age benchmark in China, and the majority were married (94.0%). Socioeconomically, the cohort was characterised by lower to middle educational attainment, with 80.9% having a high school education or less. The income distribution was concentrated in the lower brackets, with 83.3% of participants earning a monthly income of ¥4000 or less. Most respondents were urban residents (81.5%) and were not in active employment (67.4%). Familiarity with digital technologies was low; 80.6% had never used AI tools and 92.0% did not use a mobile app for T2DM management.

Table 2. Baseline demographic of participants.

	Frequency	Percentage
Gender
Male	199	47.0
Female	224	53.0
Age group
Below 60 years	237	56.0
60 years and above	186	44.0
Marital status
Married	398	94.0
Single, divorced or widowed	25	6.0
Education
Below high school	172	40.7
High school or equivalent	170	40.2
College and above	81	19.1
Monthly income (¥)*
≤2000	145	34.3
2000–4000	207	49.0
≥4000	71	16.7
Residence
Urban	345	81.5
Rural	78	18.5
Employment
Employed	138	32.6
Unemployed or retired	285	67.4
AI familiarity
Used AI before	82	19.4
Never used AI before	341	80.6
Diabetes app usage
Using	34	8.0
Not using	389	92.0

^*US$1=¥6.90. The demographic information of the 423 participants in the study is presented as a frequency and percentage of the total number of participants.

AI, artificial intelligence.

A sensitivity analysis was conducted to assess potential bias arising from site-level heterogeneity in survey effectiveness. We found that the high-effectiveness group was disproportionately composed of urban residents, individuals with higher education and income levels (online supplemental table S2). To further investigate whether these demographic differences systematically biased our preference estimates, we re-estimated our primary model on the high- and low-effectiveness subsamples separately. The results showed high consistency in the direction, statistical significance and relative magnitude of all key attribute coefficients, confirming the robustness of our main findings in online supplemental note 1.

In unadjusted selection frequencies, price showed a steep negative gradient: the alternative priced at ¥0/month was selected in 84.6% of its appearances (1897/2243), falling to 18.4% at ¥100 (419/2276) and 7.9% at ¥200 (177/2249). Preferences for in-person follow-up frequency favoured less frequent contact: every 3 months was chosen 39.9% of the time (910/2281), compared with 36.4% for every month (810/2226) and 34.2% for every 2 weeks (773/2261). It is worth noting that the current Chinese National Basic Public Health Service guidelines stipulate a minimum of four follow-up visits per year (approximately once every 3 months) as the standard of care for T2DM at the primary care level. For feedback modality, instant AI feedback was selected somewhat more often (38.2%, 1292/3384) than regular dietitian feedback (35.5%, 1201/3384). Expert oversight showed little difference (36.5% with oversight vs 37.2% without; 1234/3384 vs 1259/3384). Within the basis for recommendation, selection proportions were similar (clinical trial evidence 37.8% (842/2277), guideline recommendation 37.0% (854/2259), physician recommendation 35.7% (797/2232)).

A comprehensive mixed logit model incorporating interactions between service attributes and participant characteristics was estimated (table 3). The main effect coefficients in this model represent the preferences of a base reference population (eg, male, under 60, rural residents with a high school education and low income; see table 3 notes for full details), while interaction terms capture how these preferences shift across different demographic groups.

Table 3. Mixed logit model results.

	Coefficient	95% CI	P value
Left	−0.09	(−0.26 to 0.08)	0.298
Base reference population
Opt out	−2.98		<0.001
Subscription fee (per ¥100)	−3.57	(−4.01 to −3.13)	<0.001
Guideline recommendation	−0.14	(−0.53 to 0.25)	0.50
Physician recommendation	0.05	(−0.20 to 0.29)	0.69
Every 2 weeks	−0.37	(−0.68 to −0.05)	0.02
Every month	−0.29	(−0.55 to −0.03)	0.03
Expert oversight	−0.29	(−0.51 to −0.06)	0.01
Instant feedback via AI	0.17	(−0.07 to 0.41)	0.17
Impact of demographic factors
Subscription fee (per ¥100)×age≥60	−0.53	(−0.81 to −0.26)	<0.001
Subscription fee (per ¥100)×unemployed/retired	0.26	(−0.07 to 0.59)	0.12
Subscription fee (per ¥100)×income 2k–4k	0.03	(−0.24 to 0.30)	0.84
Subscription fee (per ¥100)×income>4k	0.35	(−0.03 to 0.72)	0.07
Subscription fee (per ¥100)×rural	−1.51	(−1.91 to −1.12)	<0.001
Subscription fee (per ¥100)×college+	0.63	(0.27 to 0.98)	<0.001
Subscription fee (per ¥100)×below high school	−0.40	(−0.68 to −0.11)	0.007
Subscription fee (per ¥100)×used AI before	−0.89	(−1.23 to −0.55)	<0.001
Subscription fee (per ¥100)×using diabetes app	2.52	(2.15 to 2.89)	<0.001
Opt out×income 2k–4k	−0.76	(−1.21 to −0.31)	<0.001
Opt out×income>4k	−2.07	(−2.82 to −1.31)	<0.001
Opt out×college+	0.83	(0.23 to 1.43)	0.007
Opt out×below high school	−0.13	(−0.66 to 0.39)	0.62
Left×age≥60	−0.48	(−0.88 to −0.08)	0.02
Left×unemployed/retired	−0.21	(−0.60 to 0.18)	0.30
Guideline recommendation×female	0.79	(0.40 to 1.19)	<0.001
Guideline recommendation×age≥60	−0.18	(−0.63 to 0.27)	0.43
Guideline recommendation×college+	−0.36	(−0.87 to 0.14)	0.16
Physician recommendation×rural	0.84	(0.23 to 1.46)	0.007
Every 2 weeks×income>4k	−0.76	(−1.30 to −0.23)	0.005
Every 2 weeks×rural	0.70	(0.10 to 1.30)	0.02
Every 2 weeks×used AI before	0.65	(0.05 to 1.25)	0.03
Every 2 weeks×using diabetes app	0.31	(−0.38 to 1.01)	0.38
Every month×used AI before	0.42	(−0.12 to 0.97)	0.13
Every month×using diabetes app	0.43	(−0.18 to 1.05)	0.17
Expert oversight×below high school	0.29	(−0.08 to 0.65)	0.12
Instant feedback via AI×below high school	0.37	(−0.01 to 0.74)	0.06
Instant feedback via AI×using diabetes app	−0.49	(−0.98 to −0.00)	0.05
Random preference heterogeneity at the individual level
Opt out	2.13	(1.82 to 2.43)	<0.001
Guideline recommendation	1.35	(1.08 to 1.63)	<0.001
Physician recommendation	0.98	(0.68 to 1.27)	<0.001
Every 2 weeks	1.31	(1.01 to 1.61)	<0.001
Every month	0.06	(−1.21 to 1.33)	0.92
Expert oversight	0.58	(0.27 to 0.89)	<0.001
Instant feedback via AI	1.75	(1.55 to 1.94)	<0.001

This table presents the findings from a mixed logit analysis that has been adjusted for participant demographics. Each estimated coefficient is provided with its corresponding 95% CI. The statistical significance of these coefficients was evaluated using a t-test, with the resulting p values displayed. Subscription fee coefficient is scaled per ¥100 increase. The base reference population consisted of individuals who were male, married, employed, urban residents, had a high school education or equivalent, earned an income lower than 2000, were under 60 years of age, had no AI experience and did not use diabetes apps. Multiplication sign (×) denotes interactions between these demographic characteristics and the choice attributes.

AI, artificial intelligence.

There was no evidence of left-and-right bias in choice selection. Price had a large negative effect on choice probability (subscription fee per ¥100: coef=−3.57; 95% CI −4.01 to −3.13; p<0.001). More frequent in-person follow-up reduced utility (every 2 weeks: coef=−0.37; 95% CI −0.68 to −0.05; p=0.02; every month: coef=−0.29; 95% CI −0.55 to −0.03; p=0.03). Expert oversight was also negative (coef=−0.29; 95% CI −0.51 to −0.06; p=0.01), whereas instant AI feedback showed a small, non-significant positive mean effect (coef=0.17; 95% CI −0.07 to 0.41; p=0.17). Guideline recommendation (coef=−0.14; 95% CI −0.53 to 0.25; p=0.50) and physician recommendation (coef=0.05; 95% CI −0.20 to 0.29; p=0.69) were not significant on average.

Price sensitivity was strongly moderated by demographics. Participants aged ≥60 (n=186; coef=−0.53; 95% CI −0.81 to −0.26; p<0.001), those in rural areas (n=78; coef=−1.51; 95% CI −1.91 to −1.12; p<0.001) and those with education below high school (n=172; coef=−0.40; −0.68 to −0.11; p=0.007) were more price-sensitive. In contrast, college-educated participants were less price-sensitive (n=81; coef=0.63; 95% CI 0.27 to 0.98; p<0.001), as were diabetes-app users (n=34; coef=2.52; 95% CI 2.15 to 2.89; p<0.001). Prior AI users were more price-sensitive (n=82; coef=−0.89; 95% CI −1.23 to −0.55; p<0.001). The interaction for higher income >¥4000 trended toward lower sensitivity (n=71; coef=0.35; 95% CI −0.03 to 0.72; p=0.07), whereas ¥2000–¥4000 was null (n=207; coef=0.03; 95% CI −0.24 to 0.30; p=0.84).

We also observed systematic subgroup shifts in non-price attributes. Females (n=224) showed a stronger preference for guideline endorsement (coef=0.79; 95% CI 0.40 to 1.19; p<0.001). Rural residents (n=78) preferred physician recommendation (coef=0.84; 95% CI 0.23 to 1.46; p=0.007) and were more tolerant of bi-weekly visits (coef=0.70; 95% CI 0.10 to 1.30; p=0.02). High-income participants (n=71) disliked bi-weekly visits (coef=−0.76; 95% CI −1.30 to −0.23; p=0.005). Prior AI users (n=82) favoured higher visit frequency (coef=0.65; 95% CI 0.05 to 1.25; p=0.03). App users (n=34) showed a lower preference for instant AI feedback (coef=−0.49; 95% CI −0.98 to −0.00; p=0.05) and a suggestive positive shift among those below high school education (n=172; coef=0.37; 95% CI −0.01 to 0.74; p=0.05).

Attitudes toward opting out also varied by SES: the opt-out alternative was less attractive among middle- and high-income groups (opt-out×income 2k–4k: coef=−0.76; 95% CI −1.21 to −0.31; p<0.001; opt-out×income >4k: coef=−2.07; 95% CI −2.82 to −1.31; p<0.001), but more attractive among college-educated participants (coef=0.83; 95% CI 0.23 to 1.43; p=0.007).

Finally, there was substantial unobserved heterogeneity: the random-coefficient SDs were significant for most attributes (all p<0.001). The variance for ‘every month’ was small and not significant (p=0.92), indicating limited between-person spread for that attribute.

WTP estimates

WTP estimates, with 95% CIs, for each service attribute across different user profiles are presented in figure 1. These values represent the additional amount (in Chinese Yuan, ¥) participants were willing to pay for an attribute level compared with its reference level. In the baseline profile, respondents showed a clear disutility for more frequent in-person follow-up: WTP was −10.31 (95% CI −19.14 to −1.43) for every 2 weeks and −8.10 (95% CI −15.87 to −1.15) for every month. Mean valuations for guideline recommendation (−3.79, 95% CI −15.35 to 7.08) and physician recommendation (1.39, 95% CI −5.58 to 8.10) were small with wide CIs overlapping zero.

Figure 1. Willingness-to-pay for service attributes across user profiles. The forest plot displays the willingness-to-pay (WTP) point estimates and 95% CIs in Chinese Yuan (¥) for selected service attributes, segmented by five distinct user profiles. WTP was calculated based on the final mixed logit model with interaction terms. The reference population for each profile (baseline, high SES, AI-savvy, female, rural) is detailed in the Methods section and table 3 notes. The dashed line at zero indicates the threshold for statistical significance. AI, artificial intelligence; SES, socioeconomic status.

WTP varied systematically by profile. High-SES participants exhibited the strongest aversion to high-frequency visits, with WTP −35.16 (95% CI −52.09 to −18.91) for every 2 weeks and −8.98 (95% CI −17.76 to −1.29) for every month. In contrast, the AI-savvy profile showed higher tolerance (or modest preference) for frequent follow-up, with WTP 6.29 (95% CI −6.88 to 19.42) for every 2 weeks and 2.98 (95% CI −9.15 to 14.77) for every month. For authority-based attributes, female participants valued guideline recommendation positively—WTP 18.45 (95% CI 7.81 to 29.40)—whereas rural participants valued physician recommendation—WTP 17.58 (95% CI 5.98 to 29.17). WTP for guideline endorsement among rural participants centred near zero (−2.66, 95% CI −10.90 to 5.00).

Full WTP estimates for all attributes and profiles, including Expert Oversight and AI Instant Feedback, are reported in online supplemental figure S2.

To translate preferences into potential market behaviour, we simulated the uptake probability for three distinct AI health service packages across the predefined user profiles (figure 2). We considered the following programme profiles: AI-user-focused=AI instant feedback+bi-weekly in-person follow-up; female-focused=guideline recommendation+AI instant feedback; rural-focused=physician recommendation+bi-weekly in-person follow-up+AI instant feedback.

Figure 2. Predicted uptake at decision prices by scenario and subgroup. Bars show the mean probability of choosing the programme over the opt-out alternative (in %) from a mixed-logit Monte Carlo simulation; error bars are 95% simulation intervals. Simulations drew parameter vectors from the asymptotic sampling distribution and, for each, drew random tastes for the normally distributed coefficients (400 parameter draws×200 taste draws). AI, artificial intelligence.

At the two decision prices, mean uptake was substantially higher at ¥50 than at ¥100 across all scenarios and subgroups (figure 2). The absolute gain from the discount ranged from ~18 to 28 percentage points (pp) depending on the subgroup. Targeting the bundle to the intended audience produced the clearest lifts: in the female-focused (guideline+AI) bundle, female participants had the highest uptake (71% at ¥50; 53% at ¥100), exceeding other subgroups by ~8–16 pp at ¥50. In the rural-focused (Physician+bi-weekly+AI) bundle, rural participants led (69% at ¥50; 42% at ¥100), outperforming others by ~6–8 pp at ¥50. For the AI-user-focused (AI+bi-weekly) bundle, AI-savvy respondents led modestly at ¥50. Across scenarios, baseline uptake at ¥50 clustered around 58%–63%, and at ¥100 around 34%–44%, while rural groups tended to be lowest within bundles that did not explicitly include physician endorsement. Online supplemental figure S2 shows the full price–uptake curves (¥0–¥100) with 95% simulation bands. Uptake declines approximately linearly with price in all scenarios.

Discussion

This study extends a growing body of evidence on patient preferences for AI-enabled diabetes care by providing quantified, subgroup-specific estimates in a low- to middle-income setting. Prior research has documented broad openness to AI for data-intensive tasks—such as glucose tracking and dietary monitoring—alongside consistent reservations about AI involvement in clinical judgement and relational care. Martini et al, in a cross-sectional survey of New Zealand adults with diabetes, found that perceived usefulness of AI predicted preference for AI across data-driven tasks, while stronger patient-clinician relationships reduced preference for AI in treatment decision-making.³⁹ Similarly, Ryan et al, in a qualitative study of mHealth users, showed that physician endorsement was among the strongest drivers of trust in AI tools, and that users widely accepted physician oversight as a prerequisite for adoption.⁴⁰ Yet these studies largely treat patients as a homogeneous group. A recent systematic review⁴¹ by Thanthrige and Wickramasinghe synthesising 37 digital health intervention studies identified inadequate personalisation and cultural mismatch as the most prevalent barriers to sustained engagement with AI tools, while Rishaug et al found that T2DM patients in qualitative focus groups consistently demanded tailored, human-coaching-integrated solutions over generic one-size-fits-all approaches.⁴² The present study both corroborates and complicates these findings. Consistent with prior evidence on the primacy of cost barriers—particularly in resource-constrained settings—price emerged as the dominant determinant of adoption across all subgroups. However, this study reveals a more nuanced picture beneath the price threshold: contrary to the assumption that greater digital experience uniformly increases AI acceptance, established diabetes app users exhibited a negative preference for AI-generated dietary feedback, instead valuing professional dietitian input—a pattern we term the ‘digital experience paradox’. This finding stands in contrast to the perceived usefulness–preference relationship documented by Martini et al and suggests that direct, repeated exposure to existing AI tools may raise rather than lower the threshold for AI-generated recommendations. Meanwhile, rural residents’ preference for physician endorsement—rather than guideline or clinical evidence—aligns with Ryan et al’s emphasis on interpersonal trust as the primary mediator of AI acceptance, and underscores what Thanthrige and Wickramasinghe identified as the inadequacy of culturally non-adaptive AI systems in diverse populations.

Previous research in non-diabetes digital health contexts showed that higher digital literacy was associated with increased price sensitivity, likely because they are exposed to more free services online.³⁶ However, the data show that while AI users follow this trend, routine app users are the least price-sensitive group.

Diabetes app users are highly engaged managers of their chronic conditions.⁴³ We hypothesise that their experience with often repetitive, templated ‘AI responses’ from existing apps likely fuels a preference for high-quality, personalised mentoring from a professional dietitian; however, this interpretation would require further qualitative investigation to confirm. Consequently, they are willing to pay more for tangible expertise, viewing the service as a crucial health investment. In contrast, the AI user’s expectations are logically shaped by the consumer technology market, where powerful tools are often free or low-cost. Their higher price sensitivity is therefore not an aversion to spending, but a reflection of market-driven value expectations. Consequently, their statistically significant preference for more frequent follow-ups should be interpreted as their definition of what constitutes a premium service. For this group, high-touch, continuous engagement is a primary, tangible feature that differentiates a paid product from standard free offerings. This pattern supports a preference for blended care rather than substitution, although with fundamentally different motivations.

Rural participants in our study show significant preferences for both bi-weekly appointments and physician endorsement. These preferences are arguably intertwined, reflecting a context where in-person, physician-led visits are the primary source of trusted medical care.⁴⁴ In such settings, frequent contact serves to build crucial relational trust,⁴⁵ while the physician’s endorsement acts as the ultimate validation, reframing travel burdens⁴⁶ as a worthwhile investment in a valued relationship. It is also worth noting that the hypothetical scenarios were not explicitly framed for a specific stage of the T2DM care journey; bi-weekly contact may be more appropriate during early management or treatment intensification than for long-term stable disease.

Women in our sample prefer guideline endorsement, reversing the reference group’s ranking. This divergence suggests two complementary explanations. First, it may reflect gendered experiences within the healthcare system; women’s trust in individual physicians can be undermined by experiences of symptom discounting or scepticism.⁴⁷ In this context, a guideline may be perceived as an impartial, evidence-based safeguard against potential interpersonal biases. Additionally, the finding aligns with extensive research showing that women are often more proactive health-information seekers, motivated by a desire to be well-informed.⁴⁸ From this perspective, guidelines are not merely a source of authority, but a tool for empowerment, used to facilitate more informed, shared dialogue with their physician.

While patients generally report greater comfort with AI when a human-in-the-loop is present,^{49 50} which is validated by the pilot ranking exercise, the estimates, in contrast, indicate that expert oversight carried consistent negative WTP across profiles. One plausible interpretation lies in Herzberg’s two-factor theory: patients may perceive human oversight as a critical ‘hygiene factor’—essential for safety and trust but viewed as a fundamental component of the service rather than a premium, value-added feature for which they should pay extra.⁵¹ Furthermore, the negative WTP for oversight may also reflect a substitution effect in the experimental design. The utility derived from physician recommendation and the option for more frequent in-person follow-ups may have already fulfilled patients’ need for human assurance, thus diminishing the perceived marginal value of an additional, separately priced oversight feature. These findings suggest a potential commercial implication: if human oversight is perceived as a hygiene factor rather than a value-added premium, integrating it as a core, non-priced feature may support adoption better than pricing it separately. However, given that this interpretation is inferential and based on revealed preferences in a stated-choice experiment, further research—including qualitative work with patients and implementation studies—would be needed to validate this as a design principle.

A major implication from these findings is that a ‘one-size-fits-all’ approach to AI-enabled blended care is suboptimal and potentially inequitable—a principle long recognised in the broader T2DM management literature, where both the American Diabetes Association/European Association for the Study of Diabetes (EASD) consensus guidelines and international integrated care frameworks have increasingly called for patient-centred, individualised management.^{52 53} The population of patients with T2DM is not homogenous, with clear preferential differences for key attributes like follow-up frequency and sources of trust emerging across demographic profiles, such as gender, residence and digital experience. Even though this study elucidates optimised service bundles that would maximise uptake for specific subgroups, the reality is that variations must exist to cater to those with different needs. Furthermore, the random preference heterogeneities at the individual level were significant for most attributes. These individual-level heterogeneities could capture other factors affecting people’s preferences but were not controlled in this study. While identifying all these factors is not feasible, a practical solution is to design a programme which allows for personal customisation. This could be implemented by offering configurable pathways for follow-up intensity and a multi-channel trust architecture, empowering patients to select a blend of care that aligns with their individual needs and values.

This study has several strengths. First, the DCE methodology provides quantitative preference weights and WTP estimates that go beyond attitudinal surveys, enabling direct comparison across subgroups and informing pricing decisions. Second, the large analytical sample (n=423) with prespecified quality controls ensures robust estimation. Third, the inclusion of an opt-out option reduces hypothetical bias by allowing respondents to reject all presented alternatives. Fourth, the systematic two-stage interaction analysis provides a comprehensive characterisation of preference heterogeneity across multiple socioeconomic and digital experience dimensions simultaneously.

This study has several limitations. First, this DCE elicits stated preferences; real-world uptake may differ as patients encounter barriers or facilitators outside the experimental scenarios. Additionally, the scenarios were not explicitly framed for a particular stage of the T2DM care journey, which may affect the interpretation of preferences for visit frequency. Second, because refusals at the point of invitation were not systematically recorded, we could not calculate a response rate, and we cannot rule out non-response bias. Third, the sample—two cities in Henan Province (Jiaozuo, Puyang)—is not nationally representative. However, Henan’s large mixed urban–rural population and pronounced workforce constraints make it a relevant test case for many middle-income settings where opportunities and digital divides coexist. Replication in wealthier coastal cities and resource-constrained western provinces would strengthen the generalisability. Fourth, despite modelling age, income, residence, education, AI experience and app use, residual unobserved heterogeneity remains. Fifth, the absence of a formal qualitative interview phase in attribute development is acknowledged; future studies should consider this step to ensure comprehensive capture of patient-relevant attributes. Finally, an exploratory LCA is reported to triangulate evidence on heterogeneity; given imprecision at the segment level, class-specific estimates were not interpreted in the main text. Going forward, validation across diverse provinces and pragmatic, segment-tailored implementation studies will be needed to determine how these stated preferences for blended care translate into real-world uptake, sustained engagement and equitable outcomes.

In summary, this DCE illustrates that Chinese adults with T2DM made complex trade-offs between AI-enabled care attributes, identifying the subscription fee as the most critical determinant of choice. Distinct subgroups exhibited specific valuations for human-centric features—such as digitally experienced users willing to pay for frequent follow-ups and rural residents valuing physician endorsement—demonstrating the significance of designing ‘equity bundles’ to overcome the digital experience paradox and trust barriers. Uptake probabilities of these tailored service packages are also presented to provide strategic suggestions to policymakers on developing and pricing inclusive, configurable AI-enabled care models.

Data availability statement

Data are available upon reasonable request.