Mobile Medical Education Apps as Digital Health Information Resources in China and the United States: Longitudinal App Store Metadata Analysis and Topic Modeling of User Review Narratives

Introduction

With the rapid advancement of mobile internet, AI, and digital technologies, coupled with the growing demands of professional education for continuous learning and self-renewal capabilities, digital learning methods are gradually evolving from traditional teaching aids into essential infrastructure supporting lifelong learning. Medical education, as a field highly dependent on specialized knowledge, clinical skills, and situational decision-making abilities, demands rapid knowledge updates and navigates complex practical scenarios, placing heightened demands on learning flexibility and accessibility [1]. However, the long-standing training model centered on fixed classroom instruction and offline clinical rotations remains constrained in terms of scheduling, learning continuity, and personalized support. This vulnerability becomes particularly pronounced during public health emergencies [2,3]. The COVID-19 pandemic has further accelerated the shift toward digital, self-directed, and contextualized medical education models [4,5]. Against this backdrop, the development of mobile learning and online educational resources has enabled learning activities to be more flexibly integrated into daily teaching and clinical practice [6]. Medical education apps support diverse educational activities including coursework, clinical guideline access [7,8], case reasoning training, and formative assessment [9-13], providing on-demand learning support for medical students, residents, and practicing physicians [14].

Despite the rapid global adoption of medical education apps, their practical implementation faces numerous challenges. Existing research indicates that learners frequently encounter issues such as outdated content, insufficient system stability, and distractions from multifunctional mobile devices when using medical education apps. These factors negatively impact learning outcomes, user experience, and willingness for sustained use [15-18]. Given the high diversity in developers, target users, and application scenarios of medical education apps, quality issues do not manifest uniformly but exhibit distinct characteristics across different markets and cultural contexts [19]. Differences in medical education systems, learning approaches, and cultural backgrounds across countries and regions also lead to significant variations in learners’ perceptions and demands regarding app functionality, content quality, and “user satisfaction” [20]. This implies that evaluations of medical education apps that overlook market and cultural context differences may fail to accurately reflect their true utility and areas for improvement.

These contextual differences are particularly salient in a China-US comparison. Both countries represent large digital app markets, but medical education apps are embedded in different institutional, educational, and sociocultural environments. Established technology adoption theories, including the technology acceptance model (TAM) and the unified theory of acceptance and use of technology, suggest that users’ acceptance and continued use of digital technologies are shaped by perceived usefulness, ease of use, social influence, and facilitating conditions [21,22]. Cross-cultural technology acceptance research further indicates that national and organizational cultures may moderate how users perceive and adopt digital technologies [23]. In China, medical education app use is situated within standardized medical training pathways, national medical education reform, and examination-oriented learning pressures within the medical education system [24]. In the United States, medical education app use occurs within a credentialing environment shaped by the United States Medical Licensing Examination (USMLE) and residency selection processes [25]. These differences provide a rationale for comparing how app store reviewers in the 2 countries evaluate content reliability, technical stability, cost, exam preparation, and personalized learning support. Therefore, this study examines medical education apps not only as digital market products but also as sociotechnical educational tools embedded in distinct policy, institutional, and learning contexts.

Nevertheless, current research on user experience and satisfaction with medical education apps still primarily relies on methods such as questionnaire surveys [26,27], standardized scales, and on-site usability testing [28]. Common tools such as the Mobile App Rating Scale (MARS), developed by Stoyanov et al [29], and the Mobile App Rubric for Learning (MARuL), developed by Gladman et al [28], are mainly used to quantify an app’s functionality, learning support effectiveness, usability, and overall satisfaction [30,31]. Such studies typically rely on participants’ subjective evaluations, feature limited sample sizes, and often focus on single institutions or regions. Research perspectives also tend to be narrow, frequently emphasizing app effectiveness [32], usability [33], or overall user experience [34], while insufficient attention is given to cross-national differences and large-scale user feedback in real-world usage scenarios [35]. A growing body of research indicates that dynamic analysis based on longitudinal data is more effective than static, single-point-in-time assessments for revealing patterns in user experience and satisfaction over time and across app development stages [36,37]. Overall, existing studies remain constrained by limitations in sample size, data types, and temporal dimensions, making it challenging to systematically characterize the evolutionary process of medical education apps and the dynamic shifts in user satisfaction needs across different markets and cultural contexts.

Beyond these methodological limitations, a related conceptual gap is that medical education apps are often evaluated as functional toolkits rather than as sociotechnical interventions. Digital tools in professional education may reshape users’ learning behaviors, daily routines, role expectations, and professional identity. In medical education, mobile technologies are not merely static tools for accessing learning materials; their design and use are shaped by educational theory, preclinical and clinical training contexts, and the needs of institutions, teachers, and learners [38]. Similar role-related effects have been observed in other professional service settings. For example, research on smart technologies in nursing homes showed that digital implementation changed caregivers’ role expectations, work routines, performance norms, and adaptation strategies [39]. This perspective suggests that medical education apps should be examined not only as functional toolkits, but also as digital interventions embedded in professional training environments and institutional workflows.

Given these limitations, it is necessary to examine the evolution of medical education apps and the app-related factors associated with rating-based satisfaction patterns among app store reviewers from a systematic and dynamic perspective. As technology-driven educational products, medical education apps are also closely tied to institutional and cultural contexts in their usage and development. Therefore, understanding the evolution and use of medical education apps requires consideration of multiple dimensions, including changes in app quantity and functionality, as well as user needs and evaluation feedback. This study innovatively proposes a bidirectional “supply-demand” analytical framework. By integrating market evolution with user reviews, it explores functional adjustments, demand evolution, and cross-national differences in rating-based satisfaction patterns among app store reviewers in China and the United States, providing insights for app evaluation and design optimization. Based on the aforementioned research background and analytical framework, this study poses the following research questions (RQs):

• RQ1: What are the temporal evolution characteristics of market supply for medical education apps?
• RQ2: What app-related factors are associated with positive rating deviations (PD) and negative rating deviations (ND) among reviewers of medical education apps, and do these associations exhibit asymmetric patterns?
• RQ3: How can asymmetric patterns in PD and ND be used to classify review topics into Kano functional attributes across country-specific samples and temporal phases?

Methods

Overall Study Design and Analytical Framework

This study uses a longitudinal, cross-national mixed methods research design to systematically characterize the observable developmental features of medical education apps and rating-based satisfaction patterns among app store reviewers. The overall research approach introduces a bidirectional “supply-demand” analytical framework (Figure 1), integrating structural changes at the app market level with user experience feedback from real-world usage scenarios. This enables simultaneous capture of both the developmental trajectory of medical education apps and the dynamics of user perceptions. The analytical workflow comprised app identification and screening, extraction and preprocessing of ratings and reviews, and supply- and demand-side analyses.

‎

On the supply side, this study analyzes annual cumulative counts and temporal trends of observable and analytically eligible medical education apps from 2009 to 2025 based on app store data from China and the United States. Joinpoint regression models identify data-derived temporal segments and potential inflection points in observable app growth, thereby describing temporal patterns within the 2025-observable analytic cohort.

On the demand side, this study analyzes app store reviewers’ primary concerns and rating patterns using large-scale user-generated review data. First, by combining latent Dirichlet allocation (LDA) topic modeling with large language model–assisted topic labeling using DeepSeek-R1, we identified key themes in app store reviews. Subsequently, Tobit regression models examined the associations between topic probabilities and PD and ND, while Wald tests assessed asymmetry between these associations. Building on these results, the Kano model was used to classify the identified topics into functional attributes according to their observed associations with rating deviations.

By integrating the supply-side analysis of observable app development patterns with the demand-side analysis of review topics and rating deviations, this study provides a unified framework for comparing medical education apps in China and the United States. This framework supports the comparison of data-derived temporal patterns and topic-rating deviation associations across the 2 country-specific samples, directly addressing the research questions posed in this study.

Protocol Registration and Deviations

This study was registered on the Open Science Framework [40]. The overall research objectives, cross-national comparative design, study scope, and supply-demand analytical framework remained consistent with the registered protocol. However, several deviations and methodological refinements were introduced during implementation because the availability, structure, and quality of the app store data could not be fully determined in advance. First, app identification was supplemented by screening related or recommended apps displayed by the app store platforms in addition to retrieval using the prespecified search terms. This refinement was introduced to reduce the risk of missing eligible apps that used atypical titles or descriptions. Second, the eligibility criteria and review preprocessing procedures were operationalized in greater detail, including the requirements concerning recent app updates, minimum review counts, duplicate and noninformative records (including developer replies), suspected inauthentic or spam reviews, and rating-text concordance. These procedures were specified to ensure that the final sample was observable, analytically eligible, and suitable for linking textual content with rating-based outcomes. Third, after LDA topic modeling, AI-assisted candidate topic labels were subjected to consensus-based review by 3 research team members and independent structured assessment by 2 domain experts. This additional human validation step was introduced to improve the interpretability and semantic validity of the topic labels. These deviations and refinements did not alter the study’s primary objectives or overall analytical framework; they were made to improve app coverage, data quality, analytic feasibility, reproducibility, and methodological transparency.

Data Sources and Sample Construction

App Selection

To comprehensively cover medical education apps in China and the United States, this study systematically searched mainstream app stores in both countries in September 2025. In China, the search encompassed the Apple App Store and 7 major Android platforms (Huawei, Xiaomi, OPPO, VIVO, Baidu, 360, and App Treasure). In the United States, it covered the Apple App Store and Google Play. Searches used Chinese and English keywords (eg, “medical education,” “medical examinations,” “USMLE”; detailed strategy in Table S1 in Multimedia Appendix 1), initially identifying 5812 apps. This keyword-based app store search strategy was consistent with prior systematic searches of medical education apps, which used predefined search terms followed by title-and-description screening and eligibility assessment to identify relevant educational apps [31]. To reduce the risk of missing apps that used alternative descriptions, the keyword set was designed to include broad bilingual terms related to medical education, medical examinations, clinical learning, and professional training. In addition to keyword-based retrieval, related or recommended apps displayed by the app-store platforms were reviewed where available and screened using the same eligibility criteria. Retrieved apps were then screened according to app titles, descriptions, categories, and predefined eligibility criteria rather than keywords alone.

The predefined eligibility criteria were as follows. Apps were included if they (1) were identified through the prespecified Chinese or English search terms listed in Table S1 in Multimedia Appendix 1 or through related or recommended apps displayed by the app store platforms during the search process; (2) had an update record within the past year as shown in the app store at the September 2025 search cutoff; (3) were categorized under “Medical” or “Education”; and (4) primarily targeted medical learners, including medical students, residents, medical graduate school applicants, or users preparing for the National Medical Licensing Examination or the USMLE. Apps were excluded if they (1) primarily provided noneducational or nonsystematic medical or health-related functions, such as pharmaceutical electronic commerce, wellness or health preservation, news or information services, medical social networking, or doctor-patient communication; (2) were institution-restricted internal systems of hospitals or medical schools; (3) operated in languages other than Chinese or English; or (4) had fewer than 30 platform reviews as of the search cutoff. These criteria are also summarized in Table S2 in Multimedia Appendix 1. To ensure rigorous and objective screening, 2 researchers (JM and J Liang) independently conducted selection and evaluation based on these predefined criteria [41], with consistency verified via the kappa coefficient (Cohen κ=0.87). Disputed results were adjudicated by a third expert (YC). After removing 1993 duplicate apps, 3819 apps remained for screening. Of these, 2326 did not meet the inclusion criteria, leaving 1493 apps. A further 1026 apps were excluded based on the exclusion criteria. Ultimately, 467 medical education apps were included, comprising 192 apps from China and 275 apps from the United States (refer to Figure 2 for the complete screening process). This rigorous screening process ensured the systematic nature, comparability, and reproducibility of the sample. The final analytic sample was defined based on information available at the September 2025 search cutoff. Therefore, apps that had been discontinued before the search cutoff and were no longer searchable or had no retrievable metadata could not be systematically included in the historical denominator. The update-record and review-count criteria were used to define apps that were observable and analytically eligible at the time of data collection.

‎

User-Generated Data

To obtain real-world user feedback, this study collected user ratings and review data for the 467 selected apps across 10 major app stores in China and the United States. As of October 2025, a total of 498,760 retrievable reviews were collected (China: 321,911; United States: 176,849). All reviews underwent standardized preprocessing, including removing duplicate records and records with missing or noninformative content (including official developer replies), as well as excluding suspected inauthentic or spam reviews [42]. Sentiment polarity was classified as negative, neutral, or positive using the Baidu AI Cloud Sentiment Analysis API [43]. Rating-text discordant reviews were then excluded as a quality-control step to construct a rating-text-concordant analytic subset (refer to Figure 3 and Note S1 in Multimedia Appendix 1). Figure 3 presents 4 platform groups: the Apple App Store in China, the 7 Chinese Android app stores combined, the Apple App Store in the United States, and Google Play in the United States. Rows represent negative, neutral, and positive sentiment, and columns represent ratings of 1-2 stars, 3 stars, and 4-5 stars. Some low-rated reviews in the Chinese Android app stores expressed positive sentiment, whereas some high-rated reviews in the US Apple App Store expressed negative sentiment. Rating-text discordant reviews were not interpreted as definitively invalid records, because prior online review studies have shown that numerical ratings and textual sentiment do not always align [44,45]. Such discordance may reflect mixed evaluations, feature-specific complaints, sarcasm, or rating noise. The detailed exclusion criteria, the number of reviews removed during each quantitative preprocessing step, and the proportions excluded by the rating-text concordance screen are provided in Note S1 in Multimedia Appendix 1. After preprocessing, the final rating-text-concordant analytic subset comprised 232,598 reviews (China: 139,416; United States: 93,182). Drawing from prior research [46], we operationalized the rating-based satisfaction proxy rate as the proportion of reviews with ratings of 4 stars or higher among app store reviewers. This proxy was not interpreted as a direct measure of satisfaction in the full user population. To eliminate rating scale differences across apps, user ratings were converted into deviation values relative to the corresponding app’s overall rating. Based on this, PD and ND were calculated, both ranging from 0 to 4. To ensure analytic consistency, the final review-level dataset retained only records with valid app identifiers, review text, user ratings, and review dates. All subsequent review-level analyses, including topic modeling, Tobit regression, Wald tests, and Kano classification, were conducted using this rating-text-concordant analytic subset.

‎

Sample Comparability

The Chinese and US samples maintained consistency in search time, platform coverage, inclusion criteria, and data processing procedures to ensure comparability across cultures. Through this methodology, this study constructed a 2-sided dataset that links observable app-market supply with app-store review-based demand signals, laying the foundation for subsequent supply-side trend analysis and demand-side rating-deviation modeling. The same search period, screening criteria, preprocessing workflow, and analytic procedures were applied to the Chinese and US samples to improve cross-national comparability. However, individual-level demographic information of app-store reviewers, such as age, sex, education level, professional background, and training stage, was not available from the app stores. Therefore, direct demographic adjustment was not possible. To reduce platform-related heterogeneity, we applied identical eligibility criteria, preprocessing procedures, rating-deviation construction, and analytic workflows across all app stores, and all rating-deviation models were estimated separately by country and developmental phase.

Supply-Side Analysis: Cumulative Observable App Counts and Temporal Trends

To characterize supply-side temporal patterns, this study used the cumulative count of apps that remained observable and analytically eligible at the September 2025 search cutoff as the core metric. In this study, “survival” was operationalized as an app being retrievable from the included app stores or Qimai Apps Data platform (Beijing QiMai Technology Co, Ltd) at the search cutoff, having an update record within the preceding year, and meeting the minimum review-count criterion. Accordingly, the annual cumulative survival counts should be interpreted as temporal patterns within the 2025-observable analytic cohort, rather than as a complete historical census of all medical education apps that had ever existed in each year. Therefore, the terms “survival” and “inflection points” in this study refer to app availability and trend shifts within the 2025-observable analytic cohort, not to the survival history of all medical education apps ever released. In time-series research, relevant literature has used joinpoint regression analysis to model trends in publicly available data (such as app counts, user interest, or technological attention) to identify significant inflection points in developmental trajectories and reveal phased market shifts [47]. This method avoids predefined time segmentation by identifying trend structure shifts directly from the data, effectively mitigating subjective classification biases. Drawing on this research paradigm and considering the potential phased development characteristics of the medical education app market, this study uses joinpoint regression to analyze the temporal trends in cumulative counts of observable and analytically eligible apps. This approach identifies data-derived temporal segments, providing methodological support for subsequent phase-based analysis grounded in user reviews.

Specifically, we used Joinpoint software (version 5.4.0; National Cancer Institute, November 2025) to construct regression models on a logarithmic scale. Using annual cumulative counts of observable and analytically eligible medical education apps as the dependent variable, we estimated annual percentage changes (APCs) and average annual percentage changes (AAPCs) for each time segment, supplemented by pairwise tests to assess trend parallelism. The model allowed 0 to 3 joinpoints, selected based on data-driven Bayesian information criterion (BIC) with a global significance level of 0.05. Poisson SEs were used to estimate variability under the assumption of Poisson-distributed count data. Finally, the AAPC and its 95% CI for each trend segment were calculated from the slope of the log-transformed regression line, systematically revealing temporal dynamics.

Based on critical inflection points identified through joinpoint regression analysis, user review data underwent temporal segmentation. Analysis revealed key inflection points occurring in both China and the United States in 2017, dividing the data into 2 developmental phases: “2009-2017” and “2018-2025.” Although the first included Chinese app appeared in 2010, the first phase was labeled 2009-2017 to maintain a unified cross-national comparison window; the Chinese sample had no included apps in 2009. APC estimates that did not reach statistical significance were also considered in each phase to ensure that the subsequent analysis covered the entire temporal trajectory. Through this segmentation strategy, the study examined app-related factors associated with PD and ND and assessed how these associations varied across temporal phases.

Demand-Side Analysis: User Reviews and Rating-Deviation Modeling

Review Topic Identification (Text Processing and LDA Modeling)

To extract core themes from user reviews of medical education apps, the study first performed text preprocessing on Chinese and English reviews. This included tokenization using the jieba library [48] and the Natural Language Toolkit (NLTK), stopword removal using Chinese and English stopword lists, and cleaning numerical values, punctuation, and emoticons. Subsequently, the LDA topic model [49] was applied to analyze review data across different phases, identifying core themes and their distribution. The number of themes was determined comprehensively by balancing perplexity and coherence metrics with the interpretability of thematic clusters, ensuring themes were both statistically robust and understandable (Figures S1-S4 in Multimedia Appendix 1). This step structured review texts into quantifiable topic probability distributions, providing a quantitative foundation for the subsequent rating-deviation analysis. Visualization results, detailed text processing workflows, parameter settings, and thematic modeling specifics are documented in Figures S5-S8 and Note S2 in Multimedia Appendix 1. To address cross-linguistic differences, topic modeling was conducted separately for Chinese and English review corpora rather than by pooling translated texts. Chinese and English reviews were preprocessed using the same standardized preprocessing workflow, and topic labels were generated and validated separately within each country-phase corpus. For reporting and cross-national comparison, Chinese topic labels were translated into English after modeling and checked by bilingual researchers against the original high-probability keywords. Therefore, topics were compared at the level of conceptual domains, such as exam preparation, cost, technical stability, and learning support, rather than being treated as strictly one-to-one equivalent linguistic categories across languages.

Topic Naming and Semantic Validation (Assisted by DeepSeek-R1)

To enhance topic interpretability, DeepSeek-R1 was used to assist with the semantic labeling of topics generated by the LDA models. For each topic, the top 30 high-probability keywords were entered into the model to generate candidate topic names and brief descriptions (Table S3 in Multimedia Appendix 1). The model was accessed on October 26, 2025, with the temperature set to 0.6 and all other parameters kept at their default settings. After AI-assisted label generation, 3 members of the research team conducted a consensus-based semantic review by comparing each candidate topic name and description with the corresponding high-probability keywords. Disagreements were resolved through discussion, and the final topic labels were determined by human consensus. To provide additional verification, 2 domain experts independently assessed the finalized topic labels using a structured 5-point rubric for name accuracy and description interpretability (Table S4 in Multimedia Appendix 1). Because this validation task involved a small number of rubric-based semantic assessments of finalized topic names and descriptions, rather than independent categorical coding of individual reviews into topic categories, Cohen κ was not calculated. Instead, exact agreement between the two experts and expert validity scores were reported as descriptive indicators of topic-label validation quality. This human-in-the-loop approach was used to combine the efficiency of AI-assisted semantic labeling with the contextual judgment of human reviewers [50]. Generative AI was not used for data collection, review preprocessing, topic modeling, statistical analysis, result interpretation, or conclusion generation. To strengthen the computational validity and robustness of the text-mining procedure, we adopted a staged human-in-the-loop workflow. LDA was used as an unsupervised topic modeling approach to reduce high-dimensional review text into latent topic probability distributions, rather than to generate final conceptual categories directly. The number of topics was determined by combining perplexity, coherence, and human interpretability. AI-assisted candidate labels were subsequently reviewed through the consensus-based and expert-validation process described above. Similar natural language processing approaches have been used to analyze unstructured user-generated digital text in health-related contexts [51]. Therefore, the resulting topics should be interpreted as structured approximations of recurrent user-review themes rather than exhaustive or definitive categories of all user needs.

Asymmetric Modeling of Rating Deviations (Tobit Regression and Wald Test)

Guided by Herzberg two-factor theory, which distinguishes positive and negative evaluations, this study modeled PD and ND. PD and ND were used as dependent variables to represent the extent to which an individual review rating deviated above or below the corresponding app’s overall rating, while LDA topic probability distributions were used as independent variables to quantify the thematic composition of each review. Because PD and ND are bounded dependent variables with a lower bound of zero, Tobit regression was used to examine the associations between review topics and rating deviations. Before model estimation, multicollinearity was assessed using variance inflation factors (VIFs); all VIF values were below 5, indicating no substantial multicollinearity [52] among the independent variables (Tables S5-S12 in Multimedia Appendix 1). Tobit regression is commonly used for limited dependent variables with restricted distributions [53,54]. Wald tests were subsequently conducted to test the null hypothesis that the PD coefficient for each topic equaled the sign-reversed coefficient from the corresponding ND model (H₀: β_PD,k = –β_ND,k), thereby evaluating asymmetry between the PD and ND associations. Detailed procedures for constructing PD and ND, model specification, and statistical testing are provided in Note S3 in Multimedia Appendix 1.

Functional Attribute Classification Using the Kano Model

Based on the asymmetric associations identified in the PD and ND models, each review topic was further classified into a functional attribute using the Kano model. The resulting categories included essential factors, expectation factors, charm factors, no differential factors, and reverse factors. The classification criteria integrated statistical characteristics from the PD and ND models, including coefficient significance, coefficient direction, and the significance of coefficient differences [55]. This classification was used to translate the observed rating-deviation patterns into potentially actionable categories for app design and optimization. Detailed classification criteria and procedures are provided in Figure S9, Table S13, and Note S4 in Multimedia Appendix 1.

Methodological Integration and Analytical Advantages

This study systematically analyzed reviews of medical education apps by integrating phase-specific topic modeling, asymmetric rating-deviation modeling, and Kano functional attribute classification. This methodological framework aligns the review analysis with the data-derived temporal segments identified from the observable app cohort, thereby supporting longitudinal comparisons. By modeling PD and ND separately, the analysis avoided collapsing the 2 directions of rating deviation into a single outcome. The Kano model then translated the observed topic-rating deviation associations into functional attribute categories that may inform the design and optimization of medical education apps.

Statistical Analysis

Data entry and organization for this study were performed using Microsoft Excel 2024. Data preprocessing, LDA topic modeling, and visualization were implemented using R 4.3.3 (R Core Team) and Python (Gensim, PyLDAvis, and other packages). The temporal trend analysis of cumulative counts of observable and analytically eligible medical education apps used the joinpoint regression method, executed using Joinpoint software (version 5.4.0, National Cancer Institute). Tobit regression and Wald tests of the associations between review topics and PD and ND were performed using Stata 18.0 (StataCorp LLC). All statistical tests were 2-tailed, with the significance level set at α=.05.

Ethical Considerations

All data used in this study were sourced from publicly accessible websites, and all user information was anonymized. Sensitive data were not collected for this study, and no compensation was provided for participants. All data relevant to this study were stored on a password-encrypted computer, and only the researchers had access to the data. Ethics approval was not required according to the ethical review by the ethics committee of the Second Affiliated Hospital Zhejiang University School of Medicine.

Results

Market Evolution and Phased Changes of Medical Education Apps in China and the United States

The final analytic sample included 467 medical education apps, comprising 192 apps from China and 275 apps from the United States. Among all reviews, 219,105 were rated 4 stars or higher, corresponding to an overall rating-based satisfaction proxy rate of 94.20%. By region, the rating-based satisfaction proxy rates for medical education apps in China and the United States were 91.82% (128,018/139,416) and 97.75% (91,087/93,182), respectively. This baseline difference in the rating-based satisfaction proxy rate was interpreted as a descriptive cross-national contrast rather than as direct evidence of higher or lower educational quality in either market. Because app store ratings are generated by self-selected reviewers, the observed difference may reflect multiple factors, including app quality, platform review moderation practices, reviewer composition, and culturally shaped rating behaviors. Therefore, we further examined whether the baseline rating difference was accompanied by distinct review topics and asymmetric rating-deviation patterns across countries and temporal phases. Cumulative downloads reached 361,367,013 for Chinese apps and 28,189,607 for US apps. Figure 4 presents the annual numbers of app releases (Figure 4A), annual user review counts (Figure 4B), and joinpoint regression analyses of annual cumulative observable app counts for China (Figure 4C) and the United States (Figure 4D) within the 2025-observable analytic cohort. Apps in the Chinese market first appeared in 2010, reaching an annual release peak of 48 in 2016, followed by a gradual decline in new app releases each year thereafter. The US market began in 2009, peaking in 2017 (43 apps), and also showed a declining trend thereafter. Therefore, the 2009-2017 phase for China should be understood as a harmonized comparative period, with no included Chinese apps observed in 2009. Regarding user reviews, the Chinese market exhibited sustained overall growth; the US market reached a review peak in 2020 (14,341 reviews) before gradually declining.

‎

Joinpoint results indicate that, within the 2025-observable analytic cohort, the annual cumulative counts of observable and analytically eligible medical education apps in both China and the United States showed significant growth (P<.001) throughout the observation period. The AAPCs were 41.69% (95% CI 31.69-52.45) and 41.17% (95% CI 31.61-51.42) for China and the United States, respectively, indicating similar overall growth rates. Specifically, China’s cumulative observable app counts exhibited distinct inflection points in 2017 and 2022, revealing a 3-phase evolutionary pattern: rapid growth from 2009 to 2017 (APC=91.30%, 95% CI 64.36-122.66; P<.001), decelerating growth from 2017 to 2022 (APC=10.11%, 95% CI 0.68-20.43; P=.04), and a nonsignificant decline after 2022 (APC=–3.12%, P=.59). The cumulative observable app counts in the United States exhibited turning points in 2012, 2017, and 2021, following a 4-phase trajectory: a steep rise from 2009 to 2012 (APC=170.86%, 95% CI 72.86-324.44; P=.002), followed by sustained high growth from 2012 to 2017 (APC=47.50%, 95% CI 38.78-56.77; P<.001), a further deceleration in growth from 2017 to 2021 (APC=14.95%, 95% CI 9.36-20.83; P<.001), and no significant change after 2021 (P=.58).

Topic Distribution in User Reviews of Medical Education Apps

Tables 1 and 2 illustrate the thematic distribution of user reviews for medical education apps in China and the United States across different developmental stages. Overall, user-focused topics exhibit phased evolution over time. In China, from 2009 to 2017, users primarily emphasized “Clinical practicality” and “Technical stability,” whereas from 2018 to 2025, the focus shifted toward “Practice test function.” In the United States, users primarily focused on “Functional availability” from 2009 to 2017, while from 2018 to 2025, attention shifted toward “Cost” and “Exam preparation and practice questions.” Detailed analysis results are provided in Note S5 in Multimedia Appendix 1.

Associations Between Review Topics and Rating Deviations

Tobit Regression Results for PD and ND

The Tobit models revealed country-specific and phase-specific patterns in the associations between review topics and rating deviations. In China, reviews from the early phase were more strongly characterized by clinical practicality, content reliability, technical stability, cost, and learning support, whereas reviews from the later phase were more concentrated around examination preparation, question banks, user service, technical stability, and cost-related experiences. In the United States, reviews from the early phase were more strongly characterized by functional availability, interface clarity, content reliability, and usability, whereas reviews from the later phase showed greater attention to cost, exam preparation, convenience, and learning support. These patterns indicate that rating deviations were associated with different app-related topics across countries and temporal phases. Tables 3 and 4 present the Tobit regression results for medical education apps in China and the United States. These tables report complete estimates for all LDA-derived topics before subsequent factor discrimination and Kano classification. Overall, both the PD and ND models showed that multiple review topics were significantly associated with rating deviations. In China, during Phase I (2009-2017), content reliability showed the strongest positive association with PD (regression coefficient β=0.488, P<.001), whereas cost showed the strongest negative association with PD (β=–0.459, P<.001). In the ND model, cost showed the strongest positive association with ND (β=2.704, P<.001), whereas technical stability showed the strongest negative association with ND (β=–4.464, P<.001). During Phase II (2018-2025), user service showed the strongest positive association with PD (β=0.196, P<.001), whereas user-perceived value showed the strongest negative association with PD (β=–0.134, P<.001). In the ND model, user-perceived value showed the strongest positive association with ND (β=3.368, P<.001), whereas the nurse qualification examination showed the strongest negative association with ND (β=–1.774, P<.001).

In the United States, during Phase I (2009-2017), interface design and visual clarity showed the strongest positive association with PD (β=0.216, P<.001), whereas the practice test function showed the strongest negative association with PD (β=–0.883, P<.001). In the ND model, cost showed the strongest positive association with ND (β=1.537, P<.001), whereas content accuracy and reliability showed the strongest negative association with ND (β=–0.605, P<.001). During Phase II (2018-2025), convenience showed the strongest positive association with PD (β=0.168, P<.001), whereas the nursing qualification exam topic showed the strongest negative association with PD (β=–0.188, P<.001). In the ND model, the nursing qualification exam topic showed the strongest positive association with ND (β=1.157, P<.001), whereas learning outcomes and confidence building showed the strongest negative association with ND (β=–0.888, P<.001). Complete model results are provided in Note S6 in Multimedia Appendix 1.

Asymmetry Between PD and ND Coefficients

The Wald tests compared the PD coefficient for each review topic with the sign-reversed coefficient from the corresponding ND model. A significant difference in this comparison indicated asymmetry between the positive and negative rating-deviation associations. In the Chinese sample, during Phase I (2009-2017), all topics included in the Wald comparison except “Clinical practicality” showed significant differences. During Phase II (2018-2025), all included topics showed significant differences. In the US sample, all topics included in the Wald comparison showed significant differences during Phase I (2009-2017). During Phase II (2018-2025), all included topics except “Personalized learning services” showed significant differences. Detailed results are presented in Tables S14 and S15 and Note S7 in Multimedia Appendix 1.

Kano Classification of Review Topics

Furthermore, the Kano classifications of review topics differed across countries and temporal phases. In the Chinese sample, Phase I (2009-2017) included 1 essential factor (infant medication inquiry) and 6 charm factors, whereas Phase II (2018-2025) included 1 essential factor (user-perceived value) and 10 charm factors. In the US sample, Phase I (2009-2017) included 2 essential factors (clinical teaching practice and technical stability) and 7 charm factors, whereas Phase II (2018-2025) included 2 essential factors (teaching quality and clinical decision support) and 7 charm factors. Classification results and the corresponding classification indicators are presented in Figure S10 and Tables S16-S19 in Multimedia Appendix 1, respectively. Detailed analysis is provided in Note S8 in Multimedia Appendix 1. Taken together, the Kano classification suggested that app store reviewers in both countries increasingly distinguished between baseline requirements and value-added features. Essential factors reflected functions whose absence was more likely to generate dissatisfaction, such as technical stability, clinical teaching practice, teaching quality, or clinical decision support, depending on country and phase. Charm factors reflected features that could increase satisfaction when present, particularly exam preparation, learning management, personalized learning, convenience, user service, and cross-screen adaptation. This distinction helped translate the Tobit regression results into more actionable insights for app design by separating foundational quality expectations from features that may create additional perceived value.

Discussion

Principal Findings

This study aimed to characterize the longitudinal evolution of medical education apps in China and the United States and to identify app-related factors associated with PD and ND among app store reviewers. Overall, within the 2025-observable analytic cohort, the findings showed phased development patterns, with notable data-derived inflection points around 2017 and 2021-2022, while country-specific trajectories differed in observable app availability, review activity, and user-evaluation patterns. On the demand side, Chinese app store reviewers showed stronger associations between rating deviations and exam preparation, cost, question banks, personalized learning, and user service, whereas US app store reviewers placed greater emphasis on technical stability, functional availability, content reliability, and clinical decision support. These findings address the study objectives by linking supply-side observable app development patterns with demand-side user review patterns in 2 distinct medical education contexts.

Phased Evolutionary Characteristics of Medical Education Apps in China and the United States

Phased Inflection Points and Synchronization

Based on joinpoint regression results, the 2025-observable medical education app cohorts in China and the United States exhibited similar data-derived temporal inflection points during the study period, primarily concentrated in 2017 and 2021-2022. The apparent similarity in these temporal patterns may be interpreted in relation to shared contextual developments, including the global diffusion of mobile internet technology, changes in digital medical education platforms, and shifts in the external environment. However, these factors were not directly measured, and the observed similarity should not be interpreted as evidence that they caused the temporal patterns. From a technological perspective, policy updates on the iOS platform may provide contextual support for interpreting the observed temporal patterns. Starting with iOS 11, all 32-bit apps were no longer supported, requiring developers to update their apps to a 64-bit architecture for operation [56]. As early as 2015, Apple mandated that new apps and updates support 64-bit architecture [57]. This policy may have contributed to the withdrawal of outdated or unmaintained applications by increasing technical requirements for app development and maintenance, and may therefore have influenced the composition of the observable app cohort during the study period. From a technology adoption perspective, while the TAM was introduced to medical education in 2010, it remained in an exploratory phase until 2016, with research primarily focused on electronic learning and mobile learning [58]. These developments in technology adoption may have created contextual conditions that supported growing awareness and use of mobile learning tools around 2017. Institutionally, China formally advanced its integrated clinical medical education system (“5+3” [5-year undergraduate + 3-year residency] and “3+2” [3-year specialty + 2-year assistant general practice training]) alongside medical education coordination reforms in 2017, emphasizing seamless integration between undergraduate education, residency training, and clinical practice [59]. This reform may have increased attention to standardized and practice-oriented clinical skills training and may have created an institutional context supportive of digital and mobile learning tools.

Changes in the external environment may also provide contextual explanations for the observed temporal patterns. Previous research has reported increased use of mobile medical education apps among medical students during COVID-19 lockdowns, with use remaining elevated after the pandemic [60]. This evidence may provide contextual support for interpreting the data-derived temporal changes observed around 2021-2022. More broadly, the increasing use of mobile learning may reflect its transition from a supplementary resource toward a more established educational approach.

Divergent Market Trajectories in China and the United States

Despite similar data-derived inflection points, the observable cohort-based trajectories in China and the United States showed notable differences. Within the 2025-observable analytic cohort, China showed rapid early expansion followed by slower growth, while user reviews continued to grow. This may suggest that review activity and expressed engagement increased among app-store reviewers even as observable app growth slowed. In contrast, the US sample showed declining app listings after 2017, with review volumes peaking around 2020 before gradually declining. This pattern may be consistent with product consolidation within the observable app cohort and changing review activity in a more mature platform environment. Overall, the similarity in inflection points may reflect shared pressures in the global digital transformation of medical education, whereas the divergent trajectories may be related to differences in market maturity, institutional environments, and platform competition dynamics.

Dynamic Differences in Rating-Based Satisfaction Patterns for Medical Education Apps in China and the United States

The observed pattern of rating-based satisfaction among Chinese app store reviewers showed distinct phased changes. During the early stage (2009-2017), reviewers emphasized clinical practicality, technical stability, and professional learning community. These patterns indicate that the directions and magnitudes of associations with PD and ND varied across review topics. In the later phase, reviewer attention was more frequently directed toward exam preparation drills, question banks, personalized learning, user services, and cross-screen compatibility. These findings indicate that greater positive rating deviations were most strongly associated with user service, technical stability, and cross-screen adaptation, whereas greater NDs were most strongly associated with user-perceived value and audio-visual teaching function. This demand shift may be consistent with an assessment-driven learning context in high-stakes medical education. Because this study did not measure learning processes or outcomes, it cannot determine whether greater attention to question banks reflects rote memorization or is associated with superficial learning. As a broader educational design consideration, app-based exam preparation could be combined with activities such as problem-based learning (PBL) case discussions to support critical thinking and deep learning.

The observed pattern of rating-based satisfaction among US app store reviewers showed a somewhat different trajectory. In the early phase (2009-2017), reviewers appeared to prioritize functional availability, interface design, and content reliability, suggesting that these topics were associated with smaller negative rating deviations. In the later phase (2018-2025), greater PDs were associated with exam preparation and practice questions, error analysis and memorization, personalized learning services, and convenience, whereas greater NDs were associated with cost, nursing qualification exam, clinical teaching practice, teaching quality, and clinical decision support. These observed associations may be considered through self-determination theory as a possible interpretive lens [61]. From this perspective, stable technology and authoritative content could be perceived as supporting a trustworthy learning environment and a sense of competence, while access to clinical resources and guideline-search functions may be consistent with learner autonomy.

Cross-national comparisons indicated that app store reviewers in both countries showed a gradual shift in focus from core learning and technical functions toward exam preparation, question banks, and personalized learning. This observed pattern may be considered in relation to product iterations and the broader institutional contexts in which medical education apps are used. Country-specific patterns were also observed: among Chinese app store reviewers, rating deviations were more closely associated with service experience, cost, and exam-related and personalized features [62], whereas US app store reviewers showed stronger associations with the stability and professional reliability of core functions [63]. The shared attention to “personalized learning” among app store reviewers in both countries may suggest growing interest in adaptive learning models. Such systems could customize learning paths by adjusting content difficulty to learners’ ability levels, potentially helping manage cognitive load and learning efficiency. Future medical education apps could move beyond content aggregation toward intelligent tutoring systems that are responsive to learners’ real-time states, although such designs require further validation using learning process and outcome data. Developers could also consider optimizing exam support, personalized learning, and user service functions for different educational contexts, with the effects of such design changes requiring further evaluation.

From a functional classification perspective, the Kano model results indicated that review topics were primarily classified as essential or charm factors across country-phase models, whereas expectation factors showed no stable pattern. These classifications suggest that the associations between review topics and rating deviations varied across countries and temporal phases rather than following a single pattern. Essential factors differed across country-phase models, while exam support, learning management, and personalized features were more frequently classified as charm factors in particular models. This pattern suggests that functional innovation may contribute to competitive differentiation among medical education apps.

This study conducted a cross-national comparison of the observable development and rating-based satisfaction patterns of medical education apps in China and the United States. Within the 2025-observable analytic cohorts, data-derived temporal changes were identified around 2017 and 2021-2022. Although the timing of these changes was similar, the country-specific trajectories differed: the Chinese cohort showed rapid early expansion followed by slower growth while review activity continued to increase, whereas the US cohort showed declining app listings after 2017 and changing review activity after 2020. The similarity in timing may be considered in relation to shared contextual developments, including the global diffusion of mobile internet technology, institutional reforms, and shifts in the external environment, whereas the divergent trajectories may be considered in relation to differences in institutional environments, market maturity, and platform competition dynamics.

Regarding rating-based satisfaction patterns, this study identified phased shifts in app store reviewers’ priorities between China and the United States. In the Chinese sample, reviewers’ attention shifted from clinical utility and interactive learning in the early phase toward exam preparation, question banks, personalized learning, user service, and cross-screen compatibility in the later phase, while the directions of topic-rating-deviation associations varied across topics and phases. In the US sample, reviewers’ attention shifted from functional availability, interface design, and content reliability in the early phase toward cost, exam preparation, and practice questions, and convenience-related features in the later phase. Cross-national comparisons suggest that while functional evolution paths differed between countries, reviewers in both samples showed increasing attention to personalized learning, which may be consistent with growing interest in adaptive learning models in educational technology.

This study holds significant theoretical and practical implications. Theoretically, this study extends the cross-national evidence on mobile learning and rating-based satisfaction and situates the observed review patterns in relation to the TAM and educational evaluation theories. Educational systems and cultural contexts may provide useful interpretive context for the observed differences in user review patterns. Practically, the findings identify potential design priorities for medical education app developers, including core functional stability, professional reliability, exam preparation support, personalized learning, and user service features. For educational administrators and policymakers, the observed patterns may inform the planning and evaluation of digital teaching resources in different institutional and educational contexts.

Design Implications for AI-Supported Medical Education Platforms

Before translating these findings into a design framework, it is important to interpret the observed cross-national differences within their institutional and educational contexts. In China, the stronger association of rating deviations with exam preparation, question banks, cost, and user service may be considered in relation to a standardized and assessment-oriented medical training environment, in which affordable and efficient exam-related support may be particularly salient. In the United States, the stronger associations with technical stability, functional availability, content reliability, and clinical decision support may be considered in relation to a credentialing and self-directed learning context, in which stable access to trustworthy resources may be particularly salient across clinical and educational settings.

On the basis of these empirically observed and contextually interpreted review and rating-deviation patterns, we proposed a conceptual design framework, termed “MEdu Copilot,” to illustrate potential design implications for future AI-supported medical education platforms (Figure 5). This framework conceptually translates the user review patterns and topic-rating deviation associations identified in China and the United States into potential design strategies for intelligent recommendation. Building on these findings, it illustrates a goal-oriented design approach that could inform the development of intelligent assistance for knowledge acquisition, clinical reasoning, and professional competency development.

‎

Figure 5 presents a proposed layered architecture for a future AI-supported medical education platform. The computing layer represents the computational infrastructure; the data layer represents a “Medical Knowledge Base” comprising multimodal and structured resources, such as textbooks, guidelines, clinical cases, 3D models, videos, images, and examination questions. The capability layer centers on a large language model that could support functions such as terminology assistance, clinical reasoning support, intelligent question answering, classroom assistance, and automated examination generation and grading. The application layer maps these potential capabilities to educational scenarios, including self-directed learning, instructional activities, assessment, and quality monitoring. The framework illustrates possible relationships among infrastructure, data, AI-supported capabilities, and educational apps; it does not represent a validated or implemented platform.

This conceptual framework follows a user-centered, goal-oriented approach that connects the user and technical design space through an iterative feedback mechanism. In a future implementation, user interaction data and learning outcomes could inform the adaptive refinement of learning resources, instructional support tools, and assessment strategies. Privacy protection, trust, and cultural sensitivity are proposed as guiding design principles that may help align future platform development with expectations for accountability and inclusivity in digital health interventions. Overall, the framework illustrates a potential pathway for the responsible integration of AI-supported capabilities into medical education.

Limitations

This study has several limitations. First, the analytic sample was defined at the September 2025 search cutoff. Although broad bilingual keywords, related-app screening where available, and title-description screening were used to improve app identification, apps discontinued earlier, no longer searchable, lacking retrievable metadata, or using atypical titles or descriptions may still have been missed. Therefore, supply-side trends should be interpreted as patterns within the 2025-observable analytic cohort, rather than as a complete historical census of medical education apps. Because the temporal segmentation was derived from cumulative counts within the 2025-observable analytic cohort, phase-based demand-side comparisons should be interpreted as exploratory comparisons across data-derived temporal segments rather than as definitive comparisons across complete historical market stages.

Second, app store reviews provide a self-selected and rating-based source of user feedback. These reviews may underrepresent “silent users” who used apps but did not leave ratings or comments. They may also overrepresent users with particularly positive experiences or strong technical frustrations. Because reviewer-level demographic, professional, training-stage, and app-use metadata were unavailable, we could not directly adjust for demographic differences across app stores or assess subgroup-specific satisfaction patterns. Accordingly, rating-based satisfaction metrics and topic patterns should be interpreted as rating- and review-based indicators among app store reviewers, rather than population-level estimates for all medical education app users or learners. In addition, all review-level analyses were conducted using the rating-text-concordant analytic subset. Although this concordance criterion was intended to improve consistency between numerical ratings and review text, it may also have excluded meaningful mixed, sarcastic, or feature-specific evaluations. Accordingly, the possibility of selection bias should be considered when interpreting the review-level findings.

Third, topic modeling and cross-linguistic comparison involve interpretive uncertainty. Although Chinese and English reviews were modeled separately and topic labels were checked through bilingual semantic validation, differences in sentiment expression and topic wording may still affect comparability across countries. The AI-assisted topic labels should therefore be understood as structured representations of review themes, rather than definitive categories of user needs.

Finally, the observational design limits causal interpretation. Reviews reflect subjective perceptions rather than actual learning outcomes, and the data do not contain individual-level information on educational status, examination participation, institutional affiliation, or cultural orientation. Interpretations related to national educational systems, institutional reforms, cultural learning behaviors, or learner psychology should therefore be viewed as contextual explanations for observed associations rather than causal inferences. Future studies should integrate learning logs, examination outcomes, user interviews, subgroup analyses, and longitudinal or experimental designs to evaluate how app features influence learning processes and professional competency development.

Conclusions

In summary, this longitudinal app store analysis identified phased evolution and country-specific rating-based satisfaction patterns among medical education apps in China and the United States. The findings suggest that app store reviewers’ concerns shifted from basic functionality, clinical practicality, and technical stability toward exam preparation, cost, and personalized learning support. More broadly, this study highlights the need to understand medical education apps not only as digital products but also as sociotechnical educational tools situated within platform ecosystems, institutional contexts, and learner needs. The MEdu Copilot framework should therefore be interpreted as a conceptual synthesis of the empirical findings and contextual policy interpretations, rather than as a validated platform, and future studies should evaluate such designs using learning process and outcome data.