Information completeness, sentiment, and engagement of orthodontic patient experience content on Chinese social media: a cross-sectional study

Abstract

Background

Social media has become an important source of orthodontic information, yet patient experience content varies in completeness and may shape treatment expectations. This study evaluated information completeness, comment sentiment, and engagement patterns in orthodontic experience content on three major Chinese social media platforms.

Methods

A cross-sectional sample of 180 eligible posts was collected from Bilibili, Douyin, and Xiaohongshu using two search terms, “orthodontic experience” and “braces diary.” Information completeness was assessed using a modified 6-point Information Completeness Score (ICS). Comment sentiment and title sentiment were quantified using sentiment analysis scores (SAS). Engagement metrics, including likes, favorites, shares, and comments, were recorded. Platform differences and subgroup differences by appliance type, treatment population, video duration, and video age were analyzed.

Results

Inter-rater reliability for total ICS was excellent (ICC (3,1) = 0.954). The automated SAS classifications showed strong agreement with the human consensus gold standard, with an overall accuracy of 91.6% (κ = 0.848). ICS differed significantly across platforms (H = 40.70, p < 0.001), with Bilibili showing the highest score (3.53 ± 1.46) and Xiaohongshu the lowest (1.78 ± 1.24). Treatment details and treatment duration were the most frequently mentioned dimensions, whereas clinician or institution credentials and cost were least frequently covered. Title sentiment showed a statistically significant but weak positive correlation with comment sentiment in the full sample (rho = 0.247, p < 0.001), with platform-specific variation. Engagement indicators were strongly intercorrelated (rho = 0.69–0.87), but ICS showed no meaningful correlation with engagement. Longer videos, clear aligner content, adult orthodontic content, and older videos had higher ICS, whereas SAS varied only slightly across groups.

Conclusions

Orthodontic patient experience content on Chinese social media shows substantial variation in information completeness. High engagement does not necessarily indicate high information quality. Professional orthodontic content should provide balanced information on treatment benefits, risks, costs, duration, and professional sources to support informed patient decision-making.

Background

Malocclusion is recognized by the World Health Organization as the third most common oral health problem after dental caries and periodontal disease, affecting a large population worldwide [1]. In China, the prevalence of malocclusion among school-aged children is approximately 47.92%, affecting more than 260 million children; its impact extends beyond dental alignment to facial function, esthetics, and psychological well-being [2, 3]. In clinical practice, orthodontic treatment can guide alveolar bone growth and optimize dental alignment, thereby improving facial esthetics and self-confidence [4]. It can also take advantage of skeletal plasticity before the pubertal growth peak to promote more balanced jaw development [5]. However, orthodontic treatment usually lasts a long time (1–3 years), is costly, and may involve invasive procedures such as tooth extraction, temporary anchorage device placement, and appliance wear; consequently, patients often approach treatment decisions with considerable caution [6–9].

With the rapid development of digital media, social media platforms have surpassed traditional search engines as major sources of health information for the public [10–12]. General video platforms such as YouTube continue to grow, with more than one billion users across age groups, while short-video platforms such as TikTok have expanded rapidly since 2021 and now reach approximately one-third of adults in the United States [13]. According to the China Internet Network Information Center, by December 2024, China had 1.101 billion social network users, accounting for 99.3% of all Internet users; online video and short-video users reached 1.070 billion and 1.040 billion, respectively [14]. Because of language barriers, platform governance mechanisms, and user ecology, China has developed a relatively distinct social media environment, including WeChat, Douyin, Xiaohongshu, and Bilibili. In the present study, Douyin refers specifically to the mainland Chinese short-video platform operated separately from the international TikTok application; and Xiaohongshu (RedNote) refers to a lifestyle-oriented platform that integrates text, images, short videos, and social commerce. Bilibili, as a leading Chinese medium- and long-form video community with more than 300 million monthly active users, has also become an important source of orthodontic information [15]. Beyond differences in content format, these platforms also differ in how users encounter and interact with health-related information: Douyin is primarily organized around short-video recommendation and rapid browsing, Xiaohongshu combines search-oriented experience sharing with image-text and short-video posts, and Bilibili supports relatively longer video narratives and community-based interaction. These differences may influence the narrative depth, informational completeness, emotional expression, and audience engagement of orthodontic-related content across platforms [16, 17]. Unlike traditional medical education, health information on social media is decentralized, personalized, and interaction-driven [18]. Algorithmic recommendation systems distribute content according to user interactions such as likes, comments, and shares, giving highly interactive content greater exposure. Although such systems improve the efficiency of information dissemination, their tendency to prioritize engagement over scientific rigor may lead to the excessive spread of sensational or emotionally charged content, while more rigorous but less interactive content may be marginalized [19–22]. Unverified claims, exaggerated treatment effects, and promotion of non-evidence-based interventions may therefore pose systematic risks to public health [23].

Among the many forms of health information on social media, patient testimonial videos constitute a distinctive genre. Presented through first-person narratives, these videos combine emotional expression with experiential sharing and may exert strong persuasive effects on audiences. They commonly cover subjective feelings during treatment, physical and psychological experiences during recovery, before-and-after changes, and practical concerns such as cost and time commitment, thereby offering perspectives that are difficult to obtain from clinical guidelines alone. Previous studies have shown that patient experiences can influence prospective patients’ treatment decisions, particularly for elective treatments such as orthodontic treatment and orthognathic surgery [13, 24–27]. However, existing literature also indicates that the quality of patient testimonial videos varies widely, ranging from useful experience sharing to partial or exaggerated narratives. The completeness, accuracy, and emotional tone of such information, and their potential effects on informed decision-making, have not been systematically evaluated and may pose risks to viewers [11, 17].

In recent years, research evaluating the quality of orthodontic information has increased. Knosel and Jung first proposed an Information Completeness Score (ICS) for orthodontic videos in 2011, providing a framework for subsequent studies of orthodontic information quality on social media [28]. Ustdal and Guney further found that YouTube videos on clear aligners varied in quality and that patient-generated videos commonly lacked important information [29]. In the Chinese social media context, Wu et al. systematically evaluated 200 highly interactive videos on early orthodontic intervention for children on Douyin and Xiaohongshu and found that content produced by orthodontists was of significantly higher quality than content produced by non-professionals [17]. Beyond information quality, researchers have also used content analysis and sentiment analysis to examine how patients express orthodontic experiences on social media. Noll et al. used large-scale Twitter sentiment analysis to compare patient experiences with fixed appliances and clear aligners and found significant differences in emotional expression between the two treatment modalities [25]. Alvarado et al. compared patient testimonial videos on orthognathic surgery and found no significant difference in comment sentiment analysis scores (SAS) between platforms, and no significant correlations between engagement, information quality, or sentiment [13]. Although these studies provide an important foundation for evaluating orthodontic information quality on social media, several gaps remain. First, no previous study has simultaneously compared a medium- to long-form video platform (Bilibili), a short-video platform (Douyin), and a hybrid text/image/short-video platform (Xiaohongshu) in the Chinese social media environment. Second, most existing studies have focused on product information or professional educational content; patient testimonial videos remain insufficiently studied, particularly with respect to quantitative evidence linking engagement metrics with information quality and sentiment.

Therefore, this study aimed to systematically evaluate the information completeness of highly visible patient-generated orthodontic experience content on three major Chinese social media platforms, namely Bilibili, Douyin, and Xiaohongshu, and to compare platform differences. We also analyzed comment sentiment and its consistency with title sentiment, examined the associations of engagement metrics (likes, favorites, shares, and comments) with information completeness and sentiment, and compared ICS and SAS across appliance types, treatment population, video duration, and video age. The findings may provide empirical evidence for orthodontists using social media for patient education and may inform governance of health information on social media.

Materials and methods

Study design

This cross-sectional study examined patient-generated orthodontic experience content on three major Chinese social media platforms: Bilibili, Douyin, and Xiaohongshu. Because the study aimed to characterize the content most likely to be encountered by patients during routine online searches, sampling was based on the default ranking of each platform rather than on random retrieval of all available orthodontic content.

The search terms were determined in three steps. First, broad terms such as “orthodontics,” “teeth straightening,” and “braces” were used for preliminary searches on the three platforms to understand the distribution of relevant content and to extract frequently used hashtags and video tags as candidate terms. Next, each candidate term was tested separately on each platform. The following exclusion criteria were applied: content not presented from a patient perspective, including posts from dentists, dental clinics, orthodontic institutions, commercial accounts, explicit advertisements, or suspected promotional collaborations; content unrelated to orthodontic treatment experience; poor audiovisual quality (severely blurred images or noisy audio); non-Chinese videos; and low-engagement content with fewer than 100 likes. The threshold was used to ensure sufficient comment data for sentiment analysis and to focus on content with a reasonable level of user visibility. Previous research has shown that most users restrict their searches to the first 20 to 30 videos [13]. Therefore, the efficiency of each term was assessed by the proportion of eligible videos among the first 30 results, with an efficiency threshold of 1:3, meaning that at least 10 of the first 30 results had to meet the inclusion criteria. Based on this testing, “orthodontic experience” and “braces diary” were selected as the final search terms. For each platform, the first 30 eligible videos or image-text posts under each search term were collected, resulting in 180 samples in total: 3 platforms × 2 keywords × 30 posts.

Searches were conducted using newly registered accounts to reduce algorithmic preference, and browser history and cache were cleared before each search. Data were collected between March 1 and March 16, 2026, with all searches performed according to the default ranking of each platform. The study used only publicly available video information and comment data and did not involve human participants; therefore, it was exempt from ethics review.

Data collection and assessment

For each video, the following information was recorded: title, link, uploader ID, number of likes, favorites, shares, and comments, publication date, video duration, platform, and search keyword. View counts were recorded only when publicly displayed by the platform, either as exact values or approximate ranges. For each video, the top 30% of comments by ranking were collected for analysis, with a maximum of 50 comments per video. Top comments were defined as comments displayed under the platform-default sorting order.

The ICS was adapted from the framework proposed by Hunsaker et al. [30]. The modified score included six dimensions, each assigned 0 or 1 point, yielding a total score ranging from 0 to 6. Two researchers independently scored all videos after standardized training. The six dimensions were as follows: (1) treatment details, including whether the video described the treatment plan, extraction, temporary anchorage devices, or appliance type; (2) side effects and complications, including pain, oral ulcers, facial swelling, or other discomfort; (3) clinician or institution credentials, including whether the professional background of the treating clinician or the institution was mentioned; (4) cost mention, including total cost or cost components; (5) visual aids, including before-and-after comparisons, dental casts, or imaging materials; and (6) treatment duration, including expected or actual treatment length and recall frequency. Inter-rater reliability was assessed using the intraclass correlation coefficient (ICC (3,1), two-way mixed-effects model, single measurement, consistency) for total ICS and Cohen’s kappa for each binary dimension. Disagreements were resolved by discussion between the two raters.

Collected comments were preprocessed by excluding emoji-only comments, advertisements, creator self-replies, and replies to other comments, yielding the final set of valid comments. The Baidu AI Open Platform sentiment analysis API was used to classify comment sentiment and generate a continuous score ranging from 0 to 1, where 0 indicated negative sentiment, 0.5 neutral sentiment, and 1 positive sentiment. For each video, the mean SAS of all valid comments was calculated as the comment sentiment score; the title text was analyzed separately to obtain a title sentiment score. To validate the automated sentiment analysis, 500 platform-stratified randomly sampled comments (Bilibili, n = 159; Douyin, n = 182; Xiaohongshu, n = 159) were independently classified as negative, neutral, or positive by two researchers blinded to the API output. Inter-rater agreement was assessed using Cohen’s kappa, and disagreements were resolved by consensus. The API classifications were compared with the human consensus gold standard using accuracy, class-specific precision, recall, and F1 scores.

Statistical analysis

Statistical analyses were performed using Python. Engagement metrics (likes, favorites, shares, and comments) were reported as medians and interquartile ranges (IQRs). Normality of continuous variables was assessed using the Shapiro-Wilk test. Because total ICS and dimensional scores were right-skewed, group comparisons were performed using the Mann-Whitney U test for two groups and the Kruskal-Wallis test for three groups, with Dunn’s test for post hoc pairwise comparisons. Categorical variables were compared using the chi-square test. Spearman rank correlation coefficients were used for correlations involving non-normally distributed or ordinal variables. All tests were two-sided, with a significance level of alpha = 0.05.

Stratified analysis

To further explore the effects of video characteristics on ICS and SAS, samples were stratified by appliance type (fixed appliances, clear aligners, lingual appliances, and functional appliances), treatment population (adolescent and adult orthodontic treatment), video duration (divided by ranking into short, medium-duration, and long videos), and video age (divided by ranking into new, intermediate, and old videos). ICS and SAS were compared across strata. Videos for which the appliance type or treatment population could not be determined were coded as unclear and excluded from subgroup comparisons.

Results

Reliability and validation of ICS and SAS assessment

Inter-rater reliability results for ICS scoring are shown in Table 1. The ICC (3,1) for total ICS was 0.954 (p < 0.001), and Cohen’s kappa values for the six dimensions ranged from 0.874 to 0.977 (all p < 0.001), indicating excellent agreement. Manual validation of the automated SAS classification was performed using a platform-stratified random sample of 500 comments. The two human raters agreed on 462 comments (92.4%; Cohen’s kappa = 0.863). Against the final human consensus labels, the API achieved an overall accuracy of 91.6% and a Cohen’s kappa of 0.848. Class-specific performance is shown in Table 2.

Table 1.

Inter-rater reliability for ICS scoring

Reliability metric	Value	p
Total ICS (ICC)	0.954	< 0.001
Treatment details (kappa)	0.933	< 0.001
Side effects/complications (kappa)	0.874	< 0.001
Clinician/institution credentials (kappa)	0.879	< 0.001
Cost mention (kappa)	0.935	< 0.001
Visual aids (kappa)	0.977	< 0.001
Treatment duration (kappa)	0.890	< 0.001

ICC was calculated using the intraclass correlation coefficient (3,1) model. The p value for ICC was based on the F test (H0: ICC = 0), and p values for Cohen’s kappa were based on z tests (H0: kappa = 0). All p values were < 0.001, indicating excellent inter-rater reliability

Table 2.

Performance of automated sentiment classification against human consensus

Sentiment class	Precision	Recall	F1-score
Negative	0.911	0.940	0.925
Neutral	0.905	0.487	0.633
Positive	0.923	0.967	0.944

Platform differences in ICS, SAS, and engagement metrics

Table 3 presents descriptive statistics for engagement metrics, ICS, and SAS across the three platforms. Engagement metrics (likes, favorites, shares, and comments) were markedly right-skewed and are reported as medians (IQRs). Likes, favorites, shares, and comments also differed significantly across platforms (all Kruskal-Wallis p < 0.001). Total ICS and mean SAS are reported as means ± standard deviations. Total ICS differed significantly across platforms (Kruskal-Wallis H = 40.70, p < 0.001), with Bilibili showing the highest score (3.53 ± 1.46) and Xiaohongshu the lowest (1.78 ± 1.24). Dunn’s post hoc tests showed that Bilibili had significantly higher total ICS than Douyin (z = 4.95, p < 0.001) and Xiaohongshu (z = 5.79, p < 0.001), whereas Douyin and Xiaohongshu did not differ significantly (z = 0.84, p = 0.403). Mean SAS showed a statistically significant but small global difference across platforms (Kruskal-Wallis H = 6.47, p = 0.039). However, no pairwise comparison remained significant after Holm adjustment (all adjusted p > 0.05), and the platform means were numerically close (0.43–0.49). Therefore, platform-level differences in SAS should be interpreted cautiously. Figure 1 further shows that the ICS distribution for Bilibili was generally higher than those for Douyin and Xiaohongshu, whereas SAS distributions were relatively similar across platforms.

Table 3.

Descriptive statistics for engagement metrics, ICS, and SAS across platforms

	Bilibili (n = 60)	Douyin (n = 60)	Xiaohongshu (n = 60)	P value (K-W)
Likes	2764 (4361)	19,654 (41727)	1930 (3774)	p < 0.001
Favorites	633 (1210)	3862 (12159)	348 (464)	p < 0.001
Shares	144 (427)	3236 (5900)	95 (174)	p < 0.001
Comments	349 (698)	1596 (3462)	190 (222)	p < 0.001
Total ICS	3.53 ± 1.46	2.07 ± 1.60	1.78 ± 1.24	p < 0.001
Mean SAS	0.47 ± 0.12	0.43 ± 0.10	0.49 ± 0.12	p = 0.039

Engagement metrics are reported as medians (IQRs); total ICS and mean SAS are reported as means ± standard deviations. Kruskal-Wallis tests were used among the three platforms

Fig. 1.

Distribution of Information Completeness Score (ICS) and sentiment analysis score (SAS) across Bilibili, Douyin, and Xiaohongshu

Mention rates of ICS dimensions

Bilibili had the highest, or nearly the highest, mention rates across all dimensions. Xiaohongshu ranked second only to Bilibili in the visual aid dimension (46.7% vs. 55.0%) but had the lowest mention rates in all other dimensions. Douyin generally showed intermediate mention rates. Platform differences were particularly evident for clinician or institution credentials, cost mention, and visual aids (Fig. 2). Overall, treatment details had the highest mention rate, whereas clinician/institution credentials and cost mention had the lowest coverage (Table 4). All chi-square tests met the expected-frequency assumption, with a minimum expected frequency of 18.67.

Fig. 2.

Mention rates of the six Information Completeness Score (ICS) dimensions across platforms. The assessed dimensions were treatment details, side effects or complications, clinician or institution credentials, cost mention, visual aids, and treatment duration

Table 4.

Mention rates of ICS dimensions across platforms

Dimension	Bilibili (n = 60)	Douyin (n = 60)	Xiaohongshu (n = 60)	Full sample (n = 180)	p
Treatment details	43 (71.7%)	28 (46.7%)	27 (45.0%)	98 (54.4%)	0.0045
Side effects	35 (58.3%)	28 (46.7%)	14 (23.3%)	77 (42.8%)	< 0.001
Clinician credentials	31 (51.7%)	20 (33.3%)	5 (8.3%)	56 (31.1%)	< 0.001
Cost mention	31 (51.7%)	15 (25.0%)	10 (16.7%)	56 (31.1%)	< 0.001
Visual aids	33 (55.0%)	9 (15.0%)	28 (46.7%)	70 (38.9%)	< 0.001
Treatment duration	39 (65.0%)	24 (40.0%)	23 (38.3%)	86 (47.8%)	0.0047

Mention rate = number of videos mentioning the dimension / number of videos on the platform. Chi-square tests were used to compare mention rates among the three platforms. The minimum expected frequency across the six 3 × 2 tables was 18.67

Consistency between title and comment sentiment

Across the full sample, title SAS was significantly positively correlated with comment SAS (Spearman rho = 0.247, p < 0.001). In platform-stratified analyses, significant positive correlations were observed for Bilibili (rho = 0.372, p = 0.003) and Xiaohongshu (rho = 0.300, p = 0.020), whereas no significant correlation was found for Douyin (rho = 0.041, p = 0.753). The mismatch between title and comment sentiment on Douyin may be related to its algorithmic recommendation mechanism (Table 5).

Table 5.

Correlations between title SAS and comment SAS (Spearman rho)

	Bilibili	Douyin	Xiaohongshu	Full sample
Spearman rho	rho = 0.372	rho = 0.041	rho = 0.300	rho = 0.247
p	0.003	0.753	0.020	< 0.001

Correlations among ICS, SAS, and engagement metrics

The four engagement indicators (likes, favorites, shares, and comments) were highly positively correlated with each other (Spearman rho = 0.69–0.87, p < 0.001), indicating clustering of user engagement behaviors. Total ICS showed no meaningful correlation with any of the four engagement indicators (|rho|<0.3), suggesting that information completeness was not meaningfully associated with engagement. SAS was weakly negatively correlated with favorites (rho=-0.240, p < 0.01) and shares (rho=-0.223, p < 0.01), indicating that content with slightly more negative comment sentiment tended to have higher engagement, possibly reflecting bidirectional interaction around controversial content. This pattern is also visible in the correlation heatmap in Fig. 3: engagement indicators clustered strongly, whereas ICS did not increase in parallel with engagement.

Fig. 3.

Spearman correlation heatmap of Information Completeness Score (ICS), sentiment analysis score (SAS), and engagement metrics

Comparison of ICS and SAS across appliance type, treatment population, video duration, and video age

Comparisons of ICS and SAS across appliance type, treatment population, video duration, and video age are shown in Table 6. Videos that did not explicitly mention appliance type (n = 31) or treatment population (n = 1) were excluded from the corresponding analyses. Lingual appliances and functional appliances were not included in between-group statistical comparisons because of small sample sizes (n = 7 each). ICS for clear aligners (3.36 ± 1.64) was significantly higher than that for fixed appliances (2.33 ± 1.60; Mann-Whitney U = 1153, p = 0.001). Regarding treatment population, adult orthodontic content had significantly higher ICS (2.66 ± 1.62) than adolescent orthodontic content (1.28 ± 1.14; U = 2887, p < 0.001). SAS did not differ significantly in either subgroup comparison.

Table 6.

Comparison of ICS and SAS across appliance type, treatment population, video duration, and video age

	ICS mean ± SD	SAS mean ± SD
Appliance type
Fixed appliances (n = 99)	2.33 ± 1.60	0.47 ± 0.12
Clear aligners (n = 36)	3.36 ± 1.64	0.47 ± 0.12
Lingual appliances (n = 7)†	3.43 ± 1.62	0.51 ± 0.11
Functional appliances (n = 7)†	1.57 ± 1.27	0.45 ± 0.07
Treatment population
Adult orthodontic treatment (n = 154)	2.66 ± 1.62	0.46 ± 0.11
Adolescent orthodontic treatment (n = 25)	1.28 ± 1.14	0.47 ± 0.15
Video duration (s)
Short videos (0–78) (n = 60)	1.42 ± 1.03	0.46 ± 0.12
Medium-duration videos (78–186) (n = 60)	2.22 ± 1.38	0.47 ± 0.11
Long videos (188–1409) (n = 60)	3.75 ± 1.48	0.46 ± 0.12
Video age (days)
New videos (4-202) (n = 60)	1.93 ± 1.51	0.44 ± 0.13
Intermediate videos (206–780) (n = 60)	2.28 ± 1.51	0.48 ± 0.11
Old videos (781–2920) (n = 60)	3.17 ± 1.63	0.47 ± 0.12

Posts with unclear appliance type (n = 31) or treatment population (n = 1) were excluded from the corresponding subgroup analyses

Fixed vs. clear aligners: Mann-Whitney U test, ICS U = 1153, p = 0.001; SAS U = 1814, p = 0.876

Adult vs. adolescent treatment: Mann-Whitney U test, ICS U = 2887, p < 0.001; SAS U = 1961, p = 0.883

Short vs. medium-duration vs. long videos: Kruskal-Wallis test, ICS H = 62.65, p < 0.001; SAS H = 0.67, p = 0.717

New vs. intermediate vs. old videos: Kruskal-Wallis test, ICS H = 18.75, p < 0.001; SAS H = 6.25, p = 0.044

†Lingual appliances and functional appliances were not included in between-group comparisons because of small sample sizes (n = 7 each)

Video duration was divided into tertiles: short videos (0–78 s), medium-duration videos (78–186 s), and long videos (188–1409 s). ICS differed significantly among the three groups (Kruskal-Wallis H = 62.65, p < 0.001), with the highest score for long videos (3.75 ± 1.48) and the lowest for short videos (1.42 ± 1.03). Dunn’s post hoc tests showed that long videos had significantly higher ICS than medium-duration videos (z = 4.77, p < 0.001) and short videos (z = 7.69, p < 0.001), and medium-duration videos also had significantly higher ICS than short videos (z = 2.91, p < 0.001). SAS values were similar across the three duration groups and did not differ significantly. Figure 4 shows platform-stratified scatterplots and fitted trend lines for the association between video duration and ICS. The correlation was strongest for Douyin (rho = 0.551, p < 0.001), followed by Bilibili (rho = 0.445, p < 0.001), and was not significant for Xiaohongshu (rho = 0.208, p = 0.111).

Fig. 4.

Association between video duration and Information Completeness Score (ICS) by platform. Spearman correlations and fitted trend lines are shown separately for Bilibili, Douyin, and Xiaohongshu

Video age was divided into tertiles: new videos (4-202 days), intermediate videos (206–780 days), and old videos (781–2920 days). ICS differed significantly among the three groups (Kruskal-Wallis H = 18.75, p < 0.001), with the highest score for old videos (3.17 ± 1.63) and the lowest for new videos (1.93 ± 1.51). Dunn’s post hoc tests showed that old videos had significantly higher ICS than new videos (z = 4.14, p < 0.001) and intermediate videos (z = 2.88, p = 0.004), whereas new and intermediate videos did not differ significantly (z = 1.26, p = 0.207). Although SAS differed significantly across the three groups (H = 6.25, p = 0.044), no pairwise comparison remained significant after adjustment for multiple comparisons (all adjusted p > 0.05), and the mean range was narrow (0.44–0.48), suggesting limited practical significance.

Discussion

This study evaluated information completeness, comment sentiment, and engagement characteristics of orthodontic experience content on three major Chinese social media platforms: Bilibili, Douyin, and Xiaohongshu. The results indicate that orthodontic experience content on Chinese social media serves both as experiential sharing and as a source of health information, but its information completeness is inconsistent. ICS was significantly higher on Bilibili than on Douyin and Xiaohongshu; engagement metrics showed no meaningful correlation with ICS; consistency between title and comment sentiment varied by platform; and video duration, appliance type, treatment population, and video age were all associated with ICS. Overall, orthodontic experience content on Chinese social media showed strong dissemination and emotional expression characteristics, while platform form, content structure, and algorithmic distribution may jointly influence the quality of information available to patients.

Differences in information completeness were first reflected in platform form. Orthodontic experience content on Bilibili, a medium- to long-form video platform, had the highest ICS, echoing findings from research on patient testimonials of orthognathic surgery in which YouTube provided more complete information than TikTok [13]. Wu et al. also showed that platform ecology may influence the quality and engagement patterns of early orthodontic education content on Douyin and Xiaohongshu [17]. The positive association between video duration and ICS further supports the interpretation that longer formats provide more space for multidimensional treatment narratives. However, the higher ICS on Bilibili should not be interpreted as evidence of greater medical accuracy; rather, it indicates that the content format is more conducive to presenting multiple dimensions of the treatment process. By contrast, Douyin, as a short-video platform, and Xiaohongshu, as a hybrid text/image/short-video platform, are more likely to emphasize outcome display, visual changes, and emotional expression. Without supplementary text, subtitles, pinned comments, or serial content, such formats may create information gaps.

At the ICS dimensions level, treatment details and treatment duration were mentioned relatively often, whereas clinician or institution credentials, costs, and side effects or complications were less frequently covered. This structural gap is consistent with previous orthodontic social media studies showing that patient-generated content often lacks important information [28, 29], and it is also linked to concerns that social media health information may affect public judgment [19, 22, 23]. Orthodontic treatment is lengthy and costly and may involve pain, oral ulcers, tooth extraction, temporary anchorage devices, recall compliance, and uncertainty of outcomes [6]. If experiential content presents only esthetic improvement or successful results while providing little information on indications, risks, alternative options, and professional sources, it may weaken prospective patients’ informed judgment.

These information gaps differed across content subgroups. Content about clear aligners had higher ICS than content about fixed appliances, possibly because patients using clear aligners more often document esthetics, cost, compliance, and lifestyle effects over time. Adult orthodontic content had higher ICS than adolescent orthodontic content, which may reflect greater involvement of adult patients in treatment decisions, cost bearing, and self-narration. The insufficiency of information in adolescent orthodontic content deserves particular attention because its intended audience often includes parents. If platform content overemphasizes early intervention or appearance improvement without explaining risks, indications, and alternatives, it may amplify anxiety or create inappropriate expectations. Future research and platform governance should therefore move beyond platform comparisons and further distinguish appliance type, age group, creator identity, and commercial attributes.

The engagement and sentiment results highlight a key public health implication of this study: visibility cannot be used as a proxy for information quality. Likes, favorites, shares, and comments were highly correlated, indicating clustering of user engagement behaviors; however, ICS showed no meaningful correlation with any of the four engagement indicators, consistent with Alvarado et al., who found no significant associations of engagement with ICS or SAS [13]. This finding suggests a structural mismatch between what users interact with and what is informationally complete. In algorithm-mediated social media environments, where ranking and recommendation are strongly shaped by user interaction, highly visible orthodontic testimonials may not necessarily provide balanced information about indications, risks, costs, duration, or professional sources. Although this cross-sectional finding cannot prove that algorithms directly suppress medical accuracy, it does indicate a systemic risk: patients may mistake popularity for credibility, even when engagement is largely independent of information completeness. For expectation-sensitive treatments such as orthodontics, this mismatch may contribute to selective optimism and misaligned expectations.

The weak negative correlations between SAS and favorites and shares also deserve further interpretation. Content with lower sentiment scores may attract saving and sharing not simply because it is controversial, but because it contains information that prospective patients perceive as practically useful. Negative comments often reflect concerns about orthodontic pain, mucosal ulceration, extraction or temporary anchorage devices, unexpected financial burden, prolonged treatment duration, compliance stress, or dissatisfaction with outcomes. From this perspective, negative sentiment may function as a form of peer-to-peer risk disclosure. Patient testimonials are valued not only because they display successful esthetic changes, but also because they expose treatment burdens and uncertainties that may be less visible in formal patient education materials [24–27]. Therefore, negative narratives should not be dismissed merely as sensational or harmful content. Instead, they may reveal unmet informational needs that orthodontists should address more proactively in informed consent, chairside communication, and online patient education.

Title SAS was significantly but weakly positively correlated with comment SAS in the full sample, and the correlation remained significant only for Bilibili and Xiaohongshu but not for Douyin. Titles may provide an emotional frame that shapes viewers’ initial interpretation and comment expression. However, on recommendation-feed platforms, exposure is not entirely driven by active search; comment sentiment may also be influenced by the video narrative, existing comments, algorithmic recommendation context, and users’ immediate emotional states. Therefore, consistency between title and comment sentiment is better understood as platform-specific emotional resonance rather than evidence that titles directly cause changes in comment sentiment.

Regarding video age, old videos had higher ICS than new and intermediate videos. Because this study was cross-sectional, this finding does not indicate that videos naturally become more complete over time. One possible explanation is that older posts remaining visible in platform search results may represent a survivorship-like selection process, in which content with greater practical reference value or accumulated user interaction is more likely to continue being retrieved, saved, or discussed. In contrast, newer posts may rely more on immediate traffic and short-term visibility. However, the present study did not directly measure platform retrieval mechanisms, longitudinal exposure, or changes in content visibility over time; therefore, this interpretation should be regarded as exploratory. Future longitudinal studies are needed to clarify how information completeness and platform visibility change after publication.

At the same time, although patient experience content can feel authentic and emotionally compelling, it may still be influenced by clinic promotion, creator commercial partnerships, platform traffic strategies, and self-presentation. Therefore, evaluation of such content should consider not only whether it appears authentic or highly interactive, but also its source, conflicts of interest, narrative purpose, and medical accuracy.

Based on these findings, orthodontists and professional institutions should not merely act as passive correctors of misinformation on social media; they can also become active producers of high-quality content. Short-content platforms such as Douyin and Xiaohongshu may be useful for reaching patients and responding to common concerns. However, such short-form content should be linked to more complete information through pinned comments, playlists, serial posts, or follow-up videos. Medium- and long-form video platforms such as Bilibili may be more appropriate for serial explanations of diagnosis, treatment options, risks, follow-up requirements, and expected variability. Such strategies respond to the need for audience-oriented, quality-focused, and action-guided health communication on social media [10, 18] and may help support more informed decision-making in a high-engagement but quality-uneven content environment. These content strategies should be interpreted as cautious practical implications rather than tested interventions, because the present cross-sectional study did not evaluate whether exposure to incomplete or emotionally biased content changed patients’ understanding, treatment expectations, treatment decisions, or clinical outcomes.

Limitations

This study has several limitations. First, samples were obtained from three Chinese platforms using two search terms; therefore, the findings may not represent all orthodontic-related social media content. Second, platform search results are affected by algorithms, time, account status, geographic location, and content popularity. Although newly registered accounts were used and browsing history and cache were cleared, personalized recommendation bias could not be completely eliminated. Third, posts with fewer than 100 likes were excluded and only the top-ranked comments were collected, which may be more reflective of relatively visible patient-experience content than of all orthodontic-related content on these platforms. Fourth, SAS was based on a sentiment analysis API. Although suitable for processing large-scale text, this approach has limited ability to recognize sarcasm, online vernacular, metaphors, and complex emotions. Although the sentiment API showed high overall agreement with human coding, recall for neutral comments was lower, indicating residual classification uncertainty. Fifth, ICS primarily evaluates whether information is mentioned and does not directly assess medical accuracy, level of evidence, or potential misleadingness. Sixth, although posts from dentists, clinics, institutions, commercial accounts, explicit advertisements, and suspected promotional collaborations were excluded during screening, the remaining patient-generated content was not further classified according to creator characteristics or potential undisclosed commercial attributes. Future studies could combine professional review, source classification, and longitudinal tracking to more comprehensively evaluate the influence of orthodontic information on social media on patient decision-making. In addition, this study did not directly assess viewers’ understanding, treatment expectations, treatment choices, clinical outcomes, or patient harm. Therefore, the observed information gaps and engagement patterns should be interpreted as potential risks to informed decision-making rather than evidence of actual adverse decisions or clinical harm.

Conclusions

This cross-sectional study provides an overview of patient-generated orthodontic experience content on three major Chinese social media platforms. The findings showed that information completeness, sentiment, and engagement varied across platforms, and that highly engaged content was not necessarily more informative. Douyin showed the highest engagement, whereas Bilibili tended to contain more complete information. Video duration was positively associated with information completeness on some platforms, suggesting that longer-form content may provide more space for patients to describe treatment processes, risks, costs, and outcomes. In contrast, sentiment was only weakly associated with engagement, indicating that emotional tone alone may not fully explain how orthodontic experience content circulates online. These findings suggest that clinicians and public health communicators should pay closer attention to the informational gaps in patient-generated content and provide balanced, accessible orthodontic information that complements patients’ online experiences. Future studies should further examine how such content influences patient understanding, treatment expectations, and decision-making.

Abbreviations

ICS

Information Completeness Score

API

Application programming interface

ICC

Intraclass correlation coefficient

SAS

Sentiment analysis score

AI

Artificial intelligence

IQR

Interquartile range

K-W

Kruskal-Wallis

SD

Standard deviation

Authors’ contributions

ZHJ, XYX and FZ contributed to the conception and design of the study. ZHJ, JEZ and ZJZ performed data collection and data curation. JEZ and ZJZ independently assessed the Information Completeness Score, and disagreements were resolved through discussion with ZHJ and XYX. XYX contributed to the statistical analysis, interpretation of results and methodological supervision. ZHJ drafted the initial manuscript. JEZ, ZJZ and XYX contributed to manuscript revision. FZ obtained funding, supervised the project and critically revised the manuscript. All authors read and approved the final manuscript.

Funding

This work was supported by the Young Physician New Technology Incubation Fund of Shanghai East Hospital (Grant No. 2025-DFQNNT-003) and the Youth Scientific Research Cultivation Fund of Shanghai East Hospital (Grant No. DFPY2025002), both awarded to Fan Zhang. The funding bodies had no role in the design of the study; collection, analysis or interpretation of data; or writing of the manuscript.

Data availability

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request. Because the data were derived from publicly available social media content and online comments, raw links, user identifiers and verbatim comment-level data are not publicly deposited in order to reduce potential privacy risks.

Declarations

Ethics approval and consent to participate

This study analyzed publicly available online human-derived data from social media posts and comments. The study did not involve direct interaction or intervention with individuals, collection of private identifiable information, or presentation of identifiable individual-level data. Formal ethics approval and individual informed consent were not required.

Consent for publication

Not applicable. The manuscript does not contain identifiable individual person’s data, images or videos.

Competing interests

The authors declare no competing interests.