Investigating medical YouTubers’ parasocial visual cues in their COVID-related videos: a computer vision approach

Investigating medical YouTubers’ parasocial visual cues in their COVID-related videos: a computer vision approach | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Investigating medical YouTubers’ parasocial visual cues in their COVID-related videos: a computer vision approach Seung Woo Chae, Shujon Naha, Noriko Hara, David J. Crandall This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8101465/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract This study explores the relationship between parasocial visual cues and user engagement with medical YouTubers’ COVID-19-related videos, using a novel approach with computer vision. Based on prior literature, we measured parasocial visual cues through bodily address—where the YouTuber is seen speaking to the audience—and camera gaze—where the YouTuber is looking at the camera. For user engagement, we recorded the numbers of likes and comments, which respectively reflect viewers’ support and engagement with the discussions raised by the YouTubers. Our linear regression analyses revealed contrasting findings: bodily address was positively associated with both the number of likes and comments, whereas camera gaze was not significantly associated with either. In other words, medical YouTubers’ presence in their videos could significantly increase both the number of likes and comments, whereas their camera gaze may have little to do with viewer engagement. Our findings on bodily address align with previous parasocial interaction studies conducted through manual analysis. However, our findings on camera gaze correspond only with literature specific to parasocial interactions on YouTube and diverge from broader arguments in the general parasocial interaction literature. The possible rationales behind these contrasting results are discussed from both theoretical and practical perspectives. Health sciences/Health care Biological sciences/Psychology Social science/Psychology parasocial interaction computer vision automated visual analysis YouTube COVID-19 health communication Figures Figure 1 Figure 2 1. Introduction The rise of social media has significantly changed the traditional role of scientists and medical experts which previously established a boundary between them and the public [ 1 ]. Today, scientists actively engage in science and health communication across diverse social media platforms, from text-based platforms such as X (formerly Twitter) to video-based platforms like TikTok. This novel phenomenon has received a great deal of academic attention, especially during the COVID-19 pandemic [ 2 ]. Of diverse social media platforms, YouTube played a key role in COVID-19 communication during the pandemic [ 3 ]. A host of medical YouTubers—medical professionals who have official licenses or degrees about medicine or other relevant disciplines and upload health-related videos steadily [ 4 ]—operated their channels as reliable COVID-related information sources. This phenomenon is noteworthy given that YouTube was more commonly considered to be an entertainment source in the past rather than a place for discussing health information [ 5 ]. The current study explores medical YouTubers’ health communication during the COVID-19 pandemic and extends prior literature both theoretically and methodologically. Theoretically, this study extends prior literature on parasocial interactions, one of the most frequently employed theoretical concepts in current social media research. Prior literature has investigated diverse aspects of parasocial interactions across different platforms [ 6 – 7 ]. However, none of them have explored how medical YouTubers, particularly in the context of the COVID-19 pandemic, used parasocial visual cues to gain stronger user engagement. Thus, the current study expands the domain of parasocial interaction research by applying the concept to an understudied area. In terms of methods, the present study applies automated visual analysis via computer vision, which is frequently recognized as a promising tool for advancing current communication and media research [ 8 – 9 ]. Peng et al.’s recent article on automated visual analysis for social media research argued that these novel methods can afford a multitude of new opportunities to researchers and reinforce the extant research avenues [ 9 ]. The authors predicted that computer vision would advance current social media research by considerably increasing validity, reliability, and generalizability—for example, through description of visual media content. The current study utilizes computer vision techniques to measure parasocial visual cues in medical YouTubers’ COVID-related videos, focusing on two parasocial visual cues manually investigated by previous studies [ 6 , 10 ]: bodily address and gaze towards the camera. With these novel measurements, we will verify whether these two parasocial visual cues are significantly associated with user engagement, represented as the number of likes and comments. Given that few previous studies on parasocial interactions have linked parasocial visual cues to these two commonly used user engagement metrics on social media, we believe that this study fills an important gap in the literature. 2. Literature review 2.1. Medical experts on YouTube Health communication on YouTube has become a critical component of public health outreach, offering accessible information to a global audience [ 11 ]. In the current social media landscape, where video has risen to be a key media type [ 12 ], YouTube is positioned as a platform in which individuals can access health-related content ranging from wellness tips [ 13 ] to expert medical advice [ 11 ]. Health communication on YouTube is especially impactful because of its visual and interactive nature [ 14 – 15 ], making complex medical information more digestible for a wide range of viewers [ 16 ]. Additionally, the platform allows health/medical experts and organizations to quickly reach a large audience, enabling the rapid dissemination of public health messages, especially during urgent health crises like the COVID-19 pandemic [ 3 ]. Medical YouTubers played a pivotal role in shaping public health responses during the COVID-19 pandemic. Their influence proved to be enormous, with multiple channels accruing a massive number of subscribers and viewers. For instance, one of the most popular medical YouTubers, Doctor Mike , has about 14.27 million subscribers as of August 2025, which is more than 20 times greater than the official YouTube channel of the U.S. Centers for Disease Control and Prevention (CDC; 0.69 million subscribers). This stark contrast represents the growing role of YouTubers in health communication. Building on this influence, many medical YouTubers took steps to provide accurate and timely information by interviewing key experts during the pandemic. Notably, Doctor Mike invited Anthony Fauci, the then-Director of the National Institute of Allergy and Infectious Diseases (NIAID), to discuss the U.S. government’s guidelines regarding the virus on March 29, 2020, only 15 days after the WHO’s declaration of COVID-19 as a pandemic [ 4 ]. Considering that Fauci was the central figure of the entire nation’s COVID-19 countermeasures, this underlines the fact that medical YouTubers’ positions have evolved beyond the roles traditionally expected of them. Such collaborations exemplify how medical YouTubers leveraged their channels to spread critical health information and elevate the voices of relevant experts during the crisis. 2.2. Parasocial interactions between medical YouTubers and viewers The concept of “parasocial interaction” was originally suggested by Horton and Wohl [ 17 ], and they defined parasocial interaction as a “simulacrum of conversational give and take” (p. 215). Building on the seminal work, Perse and Rubin updated this concept particularly for communication and media research [ 18 ]. They reconceptualized a parasocial interaction as “a perceived interpersonal relationship on the part of a television viewer with a mass media persona” (p. 59) and highlighted that this phenomenon is specific to TV, which was undoubtedly one of the most compelling media at the time. It is worth noting that the related concept of “parasocial relationship” is often discussed together in previous literature [ 19 – 20 ]. That said, some studies have clearly recommended that researchers should separate these two concepts [ 10 , 21 – 22 ]. Horton and Wohl’s original study [ 17 ] introduced parasocial interaction as a momentary interaction-like experience with media figures during a single encounter, such as watching a TV show. In contrast, they described parasocial relationships as long-term emotional bonds formed over repeated interactions. In the same vein, Schramm and Hartmann [ 23 ] also underscored the need to differentiate short-term, specific parasocial interactions from long-term, generalizable parasocial relationships and emphasized their distinct methodological implications. To avoid potential confusion, we clarify here that this study is focused on visual factors in parasocial interactions, not relationships. In fact, until the appearance of social media, parasocial interactions via mass media were considered not only illusionary but sometimes even pathological [ 24 ]. This perception is understandable in the context of TV viewership, because TV stars are not a counterpart for interaction in most cases, as the foundational meaning of the word “star” indicates a shining object in the sky which is admired but not approachable. However, the emergence of social media changes both the scope and intensity of parasocial interactions. Many studies explore social media platforms as new venues for parasocial interactions [ 25 ], with a particular focus on video-based platforms [ 6 – 7 ], since, like TV, they can support parasocial visual cues. Nevertheless, one significant difference between TV and video-based social media platforms such as YouTube and Twitch is that social media influencers are much more approachable compared to TV stars. The boundary between parasocial interactions and real interactions on social media is less defined, since influencers often interact with their fans via comment threads or direct messages. As such, parasocial interactions are now “usual social activity” [ 26 ] (p. 280) rather than abnormal or pathological. This revised understanding of parasocial interactions has sparked a multitude of studies, some of which have investigated “parasocial visual cues” that have been considered crucial factors in parasocial interactions. Using TV show clips as stimuli, Cummins and Cui [ 10 ] empirically tested how a performer’s style of address in a video affects viewers’ actual parasocial interaction experience. They divided the style of address into three categories: bodily address , referring to “instances where the viewer both sees and hears the mediated performer speaking to the viewer” (p. 727); verbal address , indicating instances where the performer speaks in the video without showing their appearance; and no address . Their results demonstrated that bodily address aroused more pronounced feelings of interaction than either verbal address or no address. Additionally, emotional contagion, which is a key element of empathy, played a crucial role in enhancing these perceived interactions, particularly in response to bodily address. The current study extends Cummins and Cui’s study [ 10 ] by testing if medical YouTubers’ bodily address is significantly associated with viewer engagement. Although there is an established body of research on parasocial visual cues, there are few studies connecting those cues with user engagement on social media. The present study bridges this gap by investigating how bodily address, a proven parasocial visual cue [ 10 ], is associated with two typical user engagement metrics on YouTube: the number of likes and the number of comments [ 27 – 28 ]. These two metrics have been frequently utilized in prior literature as indicators of YouTuber viewers’ engagement. To illustrate, Munaro et al. [ 29 ] tested how these metrics are associated with language elements, linguistic style, subjectivity, emotion valence, and video category. Therefore, the current study proposes the two research questions below: RQ1-1 How is the presence of medical YouTubers in their COVID-related videos associated with the number of likes those videos receive? RQ1-2 How is the presence of medical YouTubers in their COVID-related videos associated with the number of comments those videos receive? Additionally, the present study investigates another parasocial visual cue, the direction of gaze. Ferchaud et al.’s paper [ 6 ], one of the most cited articles on parasocial interactions on YouTube, explored multiple parasocial visual cues in videos from popular YouTube channels, including direction of gaze. The variable was composed of three categories: “primarily facing towards camera,” “face not on camera,” and “primarily facing away from camera.” The authors tested whether this variable was significantly associated with the two parasocial attributes in the study: realism and authenticity. Their results were mixed; while the three categories did not significantly differ in terms of realism, “face not on camera” was associated with significantly less authenticity than both “primarily facing away from camera” and “primarily facing towards camera.” Essentially, their findings revealed that as long as the YouTuber is present, the direction of their gaze does not significantly relate to the video’s perceived authenticity. Building on this relevant work [ 6 ], the present study investigates how medical YouTubers’ direction of gaze is associated with user engagement. Since our previous two research questions address the presence of the YouTuber, we simplify the categories for the direction of gaze to “looking at the camera” and “not looking at the camera.” Regarding user engagement, again, the number of likes and comments are employed as proxies for the users’ support and participation in the discussion raised by the YouTuber. We suggest another two research questions regarding the direction of gaze: RQ2-1 How is the gaze of medical YouTubers towards the camera associated with the number of likes that their COVID-related videos receive? RQ2-2 How is the gaze of medical YouTubers towards the camera associated with the number of comments that their COVID-related videos receive? 3. Methods 3.1. Computer vision approach Computer vision has made significant strides in advancing social media research by enabling quick and large-scale analysis of visual content, which is crucial for understanding trends and patterns within digital media [ 8 ]. One key development is the integration of machine learning techniques for efficient processing and interpretation of content shared on image- or video-based platforms such as Instagram, YouTube, and Twitch [ 9 ]. These techniques enhance the ability to classify and analyze visual data at a much larger scale, which was previously hindered by manual coding processes [ 30 ]. Peng et al.’s recent review paper [ 9 ] suggested some major topics that could potentially be better investigated with computer vision: visual politics, mis-/disinformation, digital activism, body image, and digital connections. In fact, the topic of the current study—parasocial visual cues on social media—is not mentioned even in Peng et al.’s thorough work. Likewise, to our knowledge, our methodological approach to parasocial interactions is truly novel, with no previous attempts of a similar nature. That said, it is worth noting that state-of-the-art computer vision techniques, not only for parasocial interaction research but for nearly all tasks, are still far from perfect as research methods [ 8 – 9 ]. Even with the reported accuracy of computer vision models approaching 100% in some papers and on some datasets, this does not mean that the results are completely reliable, as modern computer vision algorithms are poor at generalizing to situations and contexts that they have never seen before. Moreover, while computer vision’s ability to recognize obvious visual features—such as faces, objects, and simple actions—has become fairly advanced, it generally struggles to recognize more subtle semantics based on specific context, such as memes, visual analogies, and symbolic images, which are common on social media [ 8 ]. Accordingly, as an initial attempt in this new research avenue, this study tries to avoid the fallacy of placing excessive faith in computational methods. Put differently, the methodological goal of the current study is to serve as a first steppingstone for future research on parasocial interaction using computer vision, and we share the challenges encountered in our data collection and analysis in the Limitations and Future Directions section. 3.2. Sampling To collect our sample of medical YouTubers, three researchers conducted a Google search for articles on medical and science-focused YouTube channels with the keywords “medical YouTuber” and “science YouTuber.” They then expanded their searches and identified five relevant articles recommending YouTubers who primarily discuss health, medicine, or science. From these articles, an initial list of 21 medical YouTubers was compiled. Non-English-speaking YouTubers were excluded due to language barriers, as were those lacking nationally certified medical credentials or licenses. Channels with fewer than 10 COVID-19-related videos were also removed. After applying these criteria, the final channel list consisted of five channels. As of March 1, 2023, these YouTube channels averaged 2.54 million subscribers, with subscriber counts between 24,453 and 10.6 million. The sample includes four Doctors of Medicine (M.D.) and one Doctor of Osteopathic Medicine (D.O.), with three YouTubers based in the U.S. and two in the U.K. We focused on COVID-19-related videos uploaded to these channels between January 2020 and January 2023. Only videos with at least one of the keywords “COVID,” “Corona,” or “SARS-CoV-2” in their title were included in the analysis. This process yielded 194 COVID-related videos. Of these, 39 videos were additionally ruled out because they showed too many variations of the YouTuber’s face; for example, some videos alternated between showing the YouTuber’s face with and without a mask, while others shifted between indoor and outdoor locations. After this exclusion, 155 videos remained in our final sample. These videos were, on average, 10 minutes and 54 seconds long and had 922,183 views, 29,191 likes, and 3,673 comments (including replies) as of March 1, 2023. This indicates that the videos garnered roughly one like for every 32 views and one comment for every 251 views. 3.3. Measures For all 155 videos, we preprocessed each video (24 frames per second) by extracting every other frame to reduce computational cost. From these frames, we measured two parasocial visual cues, bodily address and camera gaze , as independent variables. Regarding the dependent variables, two user engagement metrics—the number of likes and comments—were recorded. 3.3.1. Bodily address Based on Cummins and Cui’s study [ 10 ], bodily address refers to moments when an on-screen performer appears to directly speak to the viewer, which may enhance a feeling of parasocial interaction. To measure this, we used computer vision techniques to identify when the YouTuber is visible in frame. First, we detected all faces in each frame using a face detection model called “RetinaFace” [ 31 ]. RetinaFace is a deep - learning-based Convolutional Neural Network (CNN) model specifically designed for accurate and robust face detection in real-world settings. RetinaFace processes each image by applying convolutional layers that extract learned features such as edges, textures, and patterns relevant to identifying faces, and then estimates bounding boxes (minimally-enclosing rectangles surrounding detected faces) and locations of facial landmarks such as the eyes, nose, and mouth. While not perfect, RetinaFace is able to handle challenging scenarios like partially visible faces, varying lighting conditions, and different facial angles. Our next step was to determine, among every face detected in each frame of a video, which corresponded to the YouTuber. While we could have trained a face recognition algorithm to recognize specific YouTubers, this would have required a significant amount of manual data labeling effort for each YouTuber. Instead, we took a fully automatic and more scalable approach: we clustered every face appearing in the frames across each video to find groups of face images corresponding to the same people. We then assumed that the largest cluster—the person whose face appears most on-camera—was the YouTuber. We used a clustering algorithm called DBSCAN [ 32 ], which groups data points based on their density to identify clusters of closely packed points (within a defined radius, “epsilon”) and separating them from sparser regions. This approach is well-suited for detecting clusters without predefining the number of clusters and can effectively identify outliers. As the input to the clustering, we represented each detected face numerically using another deep convolutional neural network called “FaceNet” [ 33 ]. FaceNet processes facial features to create unique “embeddings,” which are compact numerical representations that capture the distinguishing characteristics of each face. In our study, DBSCAN organized the embeddings into clusters, with the largest cluster assumed to represent the YouTuber’s face. When a frame contained a face from this cluster, we counted the YouTuber as being present in that frame. Finally, we calculated the percentage of frames where the YouTuber was present compared to all extracted video frames. Figure 1 displays the distribution of this variable. —Insert Fig. 1 about here— 3.3.2. Camera gaze Research by Ferchaud et al. [ 6 ] highlighted that the direction of YouTubers’ gaze, including looking directly at the camera, may function as a parasocial visual cue. To assess this, we used a CNN-based gaze detection model [ 34 ] to estimate whether someone is looking at the camera. The model provides a confidence score (ranging from 0 to 1) for each detected face in each image frame, where higher scores indicate a greater likelihood of the performer looking at the camera. A score of 0 indicates that the model believes that the performer is not looking at the camera at all, while a score of 1 means the model is fully confident the performer is looking at the camera. For every frame where the YouTuber was present, we ran the gaze model on the YouTuber’s face and then computed the mean of these scores across the video to generate an overall measure of camera gaze. Figure 2 displays the distribution of this variable. —Insert Fig. 2 about here— 3.3.3. User engagement metrics As dependent variables, the number of likes and comments for each of the 155 videos were recorded by entering the URL information into a website that provides YouTube video data, YouTubeLikeCounter.com . To avoid any possible variation by time, we recorded the like and comments counts within two hours on March 1, 2023. 3.4. Analysis Four different multiple linear regression analyses were conducted to address the four research questions with the statistical analysis software R . For RQ1-1 and RQ1-2, we built two multiple regression models where the percentage of the YouTuber’s presence (hereinafter YouTuber presence ) was the independent variable, and the number of likes and comments were the dependent variables, respectively. For RQ2-1 and RQ2-2, we created another two multiple regression models where the confidence score of the YouTuber’s camera gaze (hereinafter camera gaze ) was the independent variable, and the number of likes and comments were the dependent variables, respectively. In these two models, the scale of camera gaze was adjusted for ease of presentation by multiplying it by 10,000. Additionally, the models for RQ1-1 and RQ2-1, where the dependent variable was the number of likes, analyzed 153 videos instead of all the 155 videos because two videos did not publicly display their like counts. Across the four regression models, there were three control variables: number of views , days since the upload (as of March 1, 2024), and video length (in seconds). The number of views was a likely covariate, as it could be positively correlated with both the number of likes and comments. It is rather intuitive that the more people watch a video, the more likes and comments it receives. Especially given the high variance in the number of views ( M = 766539.76, SD = 1659155.97), the variable was included as a control variable. Additionally, the days since upload ( M = 851.69, SD = 237.14) and the video length ( M = 693.46, SD = 479.30) were included as control variables, due to their possible influence on the number of likes and comments. For all four regression models, the VIF analysis was conducted to check multicollinearity issues. As a result, no issues with multicollinearity were found; all the VIF values were under the conventional high multicollinearity threshold of 5, as suggested by Shrestha [ 35 ]. 4. Results 4.1. Descriptives The mean value of YouTuber presence is 70.76% ( SD = 21.68). Thus, on average, around 70% of each video showed the YouTuber, with the range spanning from about 18% to 100%. Regarding camera gaze , the mean value was 0.76 ( SD = 0.19), and the range spanned between 0.25 to 0.98. This indicates that the model recognized that the YouTubers were looking at the camera for an average of 76% of the total time they appeared on screen. When it comes to the dependent variables, the mean of number of likes was 29191.38 ( SD = 46405.57), and the mean of number of comments was 3672.51 ( SD = 5134.51). 4.2. Bodily address (RQ1) The regression model for RQ1-1 that addresses how bodily address is associated with number of likes showed that bodily address ( B = 384.53, SE = 114.69, p = .001) was positively associated with number of likes when controlling for number of views , video length , and days since the upload (see Table 1 for details). These results indicate that for every 1% increase in the appearance of medical YouTubers in their videos, those videos received an average of 385 additional likes from the viewers. The R 2 value of the model indicates that about 68% of the variance in number of likes can be explained by this regression model. Table 1 Regression analysis between bodily address and number of likes. † p < 0.1; * p < 0.05; ** p < 0.01; *** p < 0.001. Coefficient SE P value Predictor Bodily address ** 384.531 114.686 0.001 Control variables Number of views *** 0.022 0.001 < 0.001 Video length (seconds) † -12.956 7.130 0.071 Days since the upload † 19.88 10.480 0.060 Constant * -26585.796 10466.498 0.012 Adjusted R 2 0.677 —Insert Table 1 about here— The regression model for RQ1-2 that addresses how bodily address is associated with number of comments demonstrated that bodily address ( B = 30.47, SE = 14.53, p = 0.038) was also positively associated with number of comments when controlling number of views , video length , and days since the upload (see Table 2 for details). These results indicate that for every 1% increase in the appearance of medical YouTubers in their videos, those videos received an average of 30 additional viewer comments. The R 2 value of the model indicates that about 56% of the variance in number of comments can be explained by this regression model. Table 2 Regression analysis between bodily address and number of comments. * p < 0.05; *** p < 0.001. Coefficient SE P value Predictor Bodily address * 30.473 14.526 0.038 Control variables Number of views *** 0.002 0.0002 < 0.001 Video length (seconds) -1.274 0.905 NS Days since the upload 1.256 1.329 NS Constant -748.673 1330.120 NS Adjusted R 2 0.558 —Insert Table 2 about here— 4.3. Camera gaze (RQ2) The regression model for RQ2-1 that addresses how camera gaze is associated with number of likes showed that camera gaze ( B = -2.29, SE = 1.33, p = 0.087) was marginally associated with number of likes when controlling number of views , video length , and days since the upload (see Table 3 for details). For the RQ2-2 model that addresses how camera gaze is associated with number of comments , there was no significant association between the two variables. To summarize, how much a medical YouTuber looks at the camera does not have a significant association with either the number of likes or comments. These findings contrast with those of the previous models regarding bodily address and user engagement, both of which displayed significant relationships. Table 3 Regression analysis between camera gaze and number of likes. The camera gaze scale was adjusted for presentation purposes by multiplying it by 10,000. † p < 0.1; * p < 0.05; *** p < 0.001. Coefficient SE P value Predictor Camera Gaze † -2.292 1.330 0.087 Control variables Number of views *** 0.022 0.001 < 0.001 Video length (seconds) -10.946 7.289 NS Days since the upload * 25.248 10.599 0.018 Constant 12573.063 14114.982 NS Adjusted R 2 0.659 —Insert Table 3 about here— 5. Discussion This study explores the relationship between parasocial visual cues and user engagement with medical YouTubers’ videos, with a particular focus on COVID-related health communication. Interestingly, the results for the two parasocial cues showed clear contrasts: bodily address was positively associated with both the number of likes and comments, whereas camera gaze was not significantly associated with either. In other words, medical YouTubers’ presence in their videos could significantly help increase both the number of likes and comments, while their camera gaze may have nothing to do with viewer engagement. As mentioned, our results showed that bodily address has a positive association with users’ support for the video, which was measured by the number of likes. Considering that our sample is composed of medical YouTubers’ COVID-related videos, these results may correspond to typical patterns in our offline health communication in everyday life; if the situation is serious, we might prefer to see medical professionals in person rather than relying on voice communication over the phone. The unique characteristics of the pandemic are also noteworthy. During this period, we experienced a lockdown situation where physical contact was restricted for safety reasons. In the wake of the lockdown, offline social interactions decreased significantly, which resulted in people using social media as an alternative venue for social interaction [ 36 ]. Studies on computer-mediated communication during the pandemic suggest that at the time, social media users were drawn to more active and dynamic forms of communication, as evidenced by the soaring popularity of TikTok [ 37 ]. Medical YouTubers’ videos might not be an exception in this regard, with viewers preferring more dynamic visual cues from the YouTuber’s presence on screen as opposed to “B-roll” footage. Our findings about bodily address and the number of likes may also relate to the findings of prior literature on the affective aspect of parasocial interactions. Horton and Wohl’s initial work [ 17 ] suggested that onscreen performers talking to the audience can increase an “illusion of intimacy” (p. 217). Extending the literature, Cummins and Cui [ 10 ] empirically confirmed that bodily address can significantly enhance viewers’ affective empathy towards a performer, particularly in the context of TV viewership. Although these studies did not explore the connection between parasocial cues and viewer engagement with the content, they demonstrated that parasocial cues could influence viewers’ affective aspect, as opposed to their cognitive aspect. The present study adds a novel finding to this body of research: the influence of parasocial cues on viewers’ affective aspect remains valid in video-based social media, particularly in the context of health communication. Our findings also show that medical YouTubers’ bodily address encouraged their viewers to engage more in the comment threads. Writing a comment is, no doubt, a more active form of user engagement and commitment than simply clicking the “like” button. In fact, prior literature on health communication has extensively examined social media comment threads as venues for user interactivity. Heldman et al.’s work [ 38 ], one of the earliest studies investigating health communication and user engagement in social media, underscores interactivity as one of the key features of social media, which distinguishes these platforms as avenues for new types of communication, rather than simple sources of information or news. This importance was further emphasized during the COVID-19 pandemic [ 39 ]; participation in comment threads allowed users to position themselves as active discussants of the COVID-19 situation, as opposed to passive observers. For example, the comment thread under one of our sample videos, titled “Why we need LOCKDOWN // UK DOCTOR // Covid-19 Vlog #6” from Dr Hope’s Sick Notes , is filled with viewer comments about the lockdown, where some clearly expressed their opinions. Thus, the commenters were not merely seeking COVID-19 information but were also voicing their perspectives. Corresponding to this importance of user participation in comment threads during the pandemic, the role of bodily address may be significant in encouraging COVID-19 discussions. While some previous studies frame bodily address as a production technique that may influence cognitive processes [ 40 ], our findings suggest that effectively deploying this parasocial cue could yield more affective results than typically expected from such a production technique. We recommend that future research further explore how different production techniques can possibly function as parasocial cues, ultimately fostering viewer engagement not only with the video itself but also with the broader topic it addresses. The present study did not find a significant association between camera gaze and either the number of likes or comments. Interestingly, this finding is not aligned with previous studies on interpersonal communication, which found that mutual gaze is a key component for engagement [ 41 – 42 ]. Since YouTubers’ gaze towards the camera can create a mediated mutual gaze [ 10 ], it would be reasonable to expect a positive correlation between the extent of camera gaze and the number of likes or comments. However, it is also true that our findings are somewhat in line with Ferchaud et al.’s study regarding parasocial cues on YouTube [ 6 ]. They found that while the presence of YouTubers on screen is significantly associated with perceived authenticity, whether they look at the camera or somewhere else is not. Both their study and the current study, which examine parasocial interactions on the same platform, found camera gaze to be a non-significant factor, while YouTuber presence was significant. Since these findings differ from those from previous studies on communication in other settings [ 43 ], this may indicate that the reason behind the results is related to the affordances of the platform rather than the variable itself [ 2 ]. Although the exact reason behind this phenomenon is beyond the scope of the current study, we present one possible explanation based on prior literature. Several studies in communication science have highlighted the importance of screen size in video consumption as a potential factor for viewer engagement and emotional response [ 44 – 45 ]. In traditional TV viewing, content is presented on a large screen, typically located in the living room. In contrast, YouTube videos are often consumed on mobile devices, tablets, or laptops, which have smaller screens in comparison to typical TV sets. As more details in the mise-en-scène of a video are visible on larger screens than on smaller ones, it is possible that smaller screen sizes render the YouTuber’s gaze a less significant factor. We hope future research will probe into this possibility or other potential explanations for this phenomenon. The current study employs a novel method, automated visual analysis via computer vision, to explore a long-researched topic, parasocial interactions. While this study has non-negligible limitations as an exploratory attempt, which we will outline in the Limitations and Future Directions section, it also has significant methodological contributions. First, this study proves that computer vision can be a valid research tool for communication and media research by confirming the findings of prior literature based on manual analysis. It is worth noting that our results align well with those of our two key references [ 6 , 10 ], from which we derived our chosen parasocial visual cues. In fact, few social media studies using computer vision have focused on confirming previous findings [ 46 ]; most have been directed towards exploring new topics rather than validating results from manual analyses [ 47 ]. While new explorations are undoubtedly worthwhile, we believe that connecting traditional, manual research methods and computer vision is essential to help create a continuum in our research, rather than separating it into distinct avenues based on methodology. The second methodological contribution of this study is that by employing computer vision, the current study significantly enhances the granularity of observation in parasocial interaction research. Previous studies on parasocial visual cues typically chose a video clip as the unit of their observation [ 6 ]. In contrast, the current study measured every other frame, which brought remarkably higher granularity. For instance, one of the videos in our sample from Doctor Mike Hansen ’s channel is 10 minutes and 26 seconds long. If this video were explored using the traditional approach, only one data point would represent this video. Meanwhile, our computer vision approach analyzed the video as 7,512 data points, demonstrating a significant improvement in granularity. Lastly, this study, which validates computer vision as a research tool for YouTube, offers valuable insights for future research on parasocial interactions on other video-based social media platforms, where manual analysis is less feasible. For example, the platform Twitch, frequently studied for parasocial interactions [ 7 ], presents challenges for manual analysis due to the length of its content; a recent study shows that the average length of popular Twitch streamers’ videos exceeds six hours [ 48 ]. This lengthy content may not be amenable to manual research methods, but can be analyzed using computer vision, which likely maintains validity regardless of video duration. 6. Limitations and future directions Once again, we clarify that current computer vision techniques are far from perfect. Thus, in any study, they should be regarded as research tools that can provide one of multiple ways to look at a phenomenon, rather than an absolute pathway to the truth. The current study is no exception, and we additionally share specific methodological limitations in our case. First, in employing computer vision, our sample had to include only videos suitable for our automated approach, which resulted in the exclusion of some valid videos. After our initial sampling, we ruled out 39 videos that were not amenable to computer vision analysis. Those excluded videos that displayed too many variations of the YouTuber’s face, either alternating between showing their face with and without a mask or between indoor and outdoor locations, which made it impossible to identify the YouTuber using our current computational approach. Second, the method we used to decide on the YouTuber’s facial cluster has a limited level of validity. Our model selected the cluster containing the most frequently shown face throughout the video as the YouTuber. However, it is possible that some faces, closely relevant to the specific video topic, may appear more often in the video than the YouTuber. Hypothetically, a video about the U.S. lockdown during the early pandemic might include images of President Trump more frequently than the YouTuber. Although our manual check revealed no such cases in our sample, future research with large-scale samples should consider this as a possible hurdle. Additionally, we suggest possible directions for future research investigating parasocial interactions via computer vision. Ferchaud et al. [ 6 ] suggested posture of the YouTuber and camera angle as potential parasocial visual cues. The authors suggested specific categories for posture of the YouTuber (e.g., “primarily seated,” “primarily standing,” “primarily in action”) and for camera angle (e.g., “straight on,” “looking up at personality/ies,” “looking down at personality/ies”), which could be useful clues for future research. While the current study, as an initial attempt in this research avenue, did not include these variables, we hope that future research will consider these visual cues with computer vision models specifically designed for this purpose. This study also has non-methodological limitations. First, the scope of this study is limited, which constrains the generalizability of our findings. We focused on a very specific topic, medical YouTubers’ health communication during the pandemic; although this is a topic worthy of academic attention, our findings may not apply to other cases involving parasocial interactions with different video topics or on other social media platforms. Our limited sample size may not reflect broader patterns on social media, which future research could consider. This study also did not explore the mechanisms behind the significant correlations between bodily address and the number of likes and comments. Examining the mechanisms through experimental research could substantially enhance our understanding of social media engagement. In such experimental research, key external variables, such as the viewing device or the YouTuber’s verbal style, could be controlled, and necessary control variables, such as demographic factors, could be incorporated into the analysis model to yield more accurate results. Specifically, investigating how parasocial cues might encourage viewers to comment on social media videos could be a meaningful topic, not only for health communication but also for understanding broader participation culture in online spaces. 7. Conclusions This study investigates an underexplored topic, the relationship between parasocial visual cues and user engagement with medical YouTubers’ COVID-related videos, using an emerging research method, computer vision. Based on prior literature [ 6 , 10 ], bodily address and camera gaze were measured as parasocial visual cues, while the number of likes and comments, which reflect viewers’ support and engagement with the discussion raised by the YouTuber, respectively, were selected as user engagement metrics. Our linear regression analyses revealed contrasting findings: bodily address was positively associated with both the number of likes and comments, whereas camera gaze was not significantly associated with either. This indicates that medical YouTubers’ presence in their videos could significantly increase both the number of likes and comments, whereas their camera gaze may have little to do with viewer engagement. While the study has non-negligible limitations, especially in its methods, we believe that these novel findings contribute to the literature on parasocial interactions on social media and that this study serves as a meaningful first step in parasocial interaction research using computer vision, offering suggestions for future research directions. Declarations Data availability The dataset analyzed in this study is available at https://osf.io/efmgn/overview?view_only=0de589ba3d6145ce84d070ad1dc4ac28. Competing interests The authors declare no competing interest. Funding This research was funded by the U.S. National Science Foundation (award number 2152423). References Gieryn, T. F. Boundary-work and the demarcation of science from non-science: strains and interests in professional ideologies of scientists. Am. Sociol. Rev. 48 , 781–795 (1983). Hara, N. & Chae, S. W. Social media affordances for mediated science communication during the COVID-19 pandemic. In Social Informatics (eds Hara, N. & Fichman, P.) (Taylor & Francis, 2025). Shukla, A. COVID-19 pandemic: an analysis of popular YouTube videos as an alternative health information platform. Health Inform. J. 27 , 1460458221994878 (2021). Chae, S. W., Hara, N., Shiroiya, H. R., Chen, J. & Ogihara, E. Being vulnerable with viewers: exploring how medical YouTubers communicated about COVID-19 with the public. PLoS ONE 19 , e0313857 (2024). Miller, K. Playing along: Digital games, YouTube, and virtual performance (Oxford Univ. Press, 2012). Ferchaud, A., Grzeslo, J., Orme, S. & LaGroue, J. Parasocial attributes and YouTube personalities: exploring content trends across the most subscribed YouTube channels. Comput. Hum. Behav. 80 , 88–96 (2018). Sherrick, B., Smith, C., Jia, Y., Thomas, B. & Franklin, S. B. How parasocial phenomena contribute to sense of community on Twitch. J. Broadcast. Electron. Media 67 , 47–67 (2022). Kim, M. & Bas, O. Seeing the Black Lives Matter movement through computer vision? an automated visual analysis of news media images on Facebook. Soc. Media Soc. 9 , 20563051231195582 (2023). Peng, Y., Lock, I. & Salah, A. A. Automated visual analysis for the study of social media effects: opportunities, approaches, and challenges. Commun. Methods Meas. 18 , 163–185 (2024). Cummins, R. G. & Cui, B. Reconceptualizing address in television programming: the effect of address and affective empathy on viewer experience of parasocial interaction. J. Commun. 64 , 723–742 (2014). Madathil, K. C., Rivera-Rodriguez, A. J., Greenstein, J. S. & Gramopadhye, A. K. Healthcare information on YouTube: a systematic review. Health Inform. J. 21 , 173–194 (2015). Schellewald, A. Understanding the popularity and affordances of TikTok through user experiences. Media Cult. Soc. 45 , 1568–1582 (2023). Herrmann, L. K. & Elutilo, B. B. Exercise with YouTube: an evaluation of YouTube’s workout #withme playlist. Int. J. Health Wellness Soc. 14 , 21–36 (2023). Kohler, S. & Dietrich, T. C. Potentials and limitations of educational videos on YouTube for science communication. Front. Commun. 6 , 581302 (2021). O’Mara, B. Social media, digital video and health promotion in a culturally and linguistically diverse Australia. Health Promot. Int. 28 , 466–476 (2013). Osman, W., Mohamed, F., Elhassan, M. & Shoufan, A. Is YouTube a reliable source of health-related information? a systematic review. BMC Med. Educ. 22 , 382 (2022). Horton, D. & Wohl, R. R. Mass communication and para-social interaction: observations on intimacy at a distance. Psychiatry 19 , 215–229 (1956). Perse, E. M. & Rubin, R. B. Attribution in social and parasocial relationships. Commun. Res. 16 , 59–77 (1989). Dibble, J. L., Hartmann, T. & Rosaen, S. F. Parasocial interaction and parasocial relationship: conceptual clarification and a critical assessment of measures. Hum. Commun. Res. 42 , 21–44 (2016). Liebers, N. & Schramm, H. Parasocial interactions and relationships with media characters: an inventory of 60 years of research. Commun. Res. Trends 38 , 4–31 (2019). Dibble, J. L. & Rosaen, S. F. Parasocial interaction as more than friendship. J. Media Psychol. (2011). Hartmann, T. & Goldhoorn, C. Horton and Wohl revisited: exploring viewers’ experience of parasocial interaction. J. Commun. 61 , 1104–1121 (2011). Schramm, H. & Hartmann, T. The PSI-process scales: a new measure to capture the intensity and breadth of parasocial processes. Communications 33 , 385–401 (2008). Kowert, R. & Daniel, E. Jr. The one-and-a-half sided parasocial relationship: the curious case of live streaming. Comput. Hum. Behav. Rep. 4 , 100150 (2021). Labrecque, L. I. Fostering consumer–brand relationships in social media environments: the role of parasocial interaction. J. Interact. Mark. 28 , 134–148 (2014). Giles, D. C. Parasocial interaction: a review of the literature and a model for future research. Media Psychol. 4 , 279–305 (2002). Kabadayi, S. & Price, K. Consumer–brand engagement on Facebook: liking and commenting behaviors. J. Res. Interact. Mark. 8 , 203–223 (2014). Kim, J. The meaning of numbers: effect of social media engagement metrics in risk communication. Commun. Stud. 72 , 195–213 (2021). Munaro, A. C., Barcelos, R. H., Maffezzolli, E. C. F., Rodrigues, J. P. S. & Paraiso, E. C. To engage or not engage? the features of video content on YouTube affecting digital consumer engagement. J. Consum. Behav. 20 , 1336–1352 (2021). Araujo, T., Lock, I. & van de Velde, B. Automated visual content analysis (AVCA) in communication research: a protocol for large scale image classification with pre-trained computer vision models. Commun. Methods Meas. 14 , 239–265 (2020). Deng, J. et al. Retinaface: single-stage dense face localisation in the wild. arXiv (2019). Hahsler, M., Piekenbrock, M. & Doran, D. dbscan: fast density-based clustering with R. J. Stat. Softw. 91 , 1–30 (2019). Schroff, F., Kalenichenko, D. & Philbin, J. Facenet: a unified embedding for face recognition and clustering. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit. 815–823 (2015). Chong, E. et al. Detection of eye contact with deep neural networks is as accurate as human experts. Nat. Commun. 11 , 6386 (2020). Shrestha, N. Detecting multicollinearity in regression analysis. Am. J. Appl. Math. Stat. 8 , 39–42 (2020). Osler, L. & Zahavi, D. Sociality and embodiment: online communication during and after COVID-19. Found. Sci. 28 , 1125–1142 (2023). Gottfried, J. Americans’ social media use: YouTube and Facebook are by far the most used online platforms among U.S. adults; TikTok’s user base has grown since 2021. Pew Res. Cent. (31 Jan 2024). Heldman, A. B., Schindelar, J. & Weaver, J. B. Social media engagement and public health communication: implications for public health organizations being truly “social”. Public Health Rev. 35 , 1–18 (2013). Chae, S. W. & Hara, N. Exploring how a YouTube channel’s political stance is associated with early COVID-19 communication on YouTube. Inf. Commun. Soc. 27 , 618–644 (2024). Hutson, J. P., Smith, T. J., Magliano, J. P. & Loschky, L. C. What is the role of the film viewer? the effects of narrative comprehension and viewing task on gaze control in film. Cogn. Res. Princ. Implic. 2 , 1–30 (2017). Goffman, E. Behavior in public places (Free Press, 1963). Kleinke, C. L. Gaze and eye contact: a research review. Psychol. Bull. 100 , 78–100 (1986). Burgoon, J. K., Coker, D. A. & Coker, R. A. Communicative effects of gaze behavior. Hum. Commun. Res. 12 , 495–524 (1986). Dunaway, J. & Soroka, S. Smartphone-size screens constrain cognitive access to video news stories. Inf. Commun. Soc. 24 , 69–84 (2021). Kim, K. J. & Sundar, S. S. Mobile persuasion: can screen size and presentation mode make a difference to trust? Hum. Commun. Res. 42 , 45–70 (2016). Peng, Y. What makes politicians’ Instagram posts popular? analyzing social media strategies of candidates and office holders with computer vision. Int. J. Press/Polit. 26 , 143–166 (2021). d’Andrea, C. & Mintz, A. Studying the live cross-platform circulation of images with computer vision API: an experiment based on a sports media event. Int. J. Commun. 13 , 21 (2019). Chae, S. W. Twitch aggression profile: exploring aggression on a live mixed-media platform. Inf. Commun. Soc. 1–19 (2025). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8101465","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":561890903,"identity":"0c03ff31-293e-48de-99ea-4f6ad7b1974b","order_by":0,"name":"Seung Woo Chae","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAqklEQVRIiWNgGAWjYJACZiCWgzDZSNBizMNGqpbEHqK1yLf3GH4uqLFJ3y/fY8DwoewwYS0GZ84YS884lpbbw8ZjwDjjHDFaJHLMmHnYDoO1MPO2EaFFfv4boJZ//9N5QFr+EqOF4QaPGdDwAwlgLYzEaDE4k1YszduXbNhzLK3gYM+5dCIc1n5442eeb3by7M2HNz74UWZNhMOQwQES1Y+CUTAKRsEowAUAa64xgNTWT8gAAAAASUVORK5CYII=","orcid":"","institution":"Texas Tech University","correspondingAuthor":true,"prefix":"","firstName":"Seung","middleName":"Woo","lastName":"Chae","suffix":""},{"id":561890905,"identity":"081f7350-5498-4d8d-838b-fb2264c68b5e","order_by":1,"name":"Shujon Naha","email":"","orcid":"","institution":"Waymo","correspondingAuthor":false,"prefix":"","firstName":"Shujon","middleName":"","lastName":"Naha","suffix":""},{"id":561890907,"identity":"bf15ef1d-0e46-44b7-aae9-d34227296f47","order_by":2,"name":"Noriko Hara","email":"","orcid":"","institution":"Indiana University","correspondingAuthor":false,"prefix":"","firstName":"Noriko","middleName":"","lastName":"Hara","suffix":""},{"id":561890908,"identity":"6a09e0e0-5c7a-477e-9c50-878814d9e96b","order_by":3,"name":"David J. Crandall","email":"","orcid":"","institution":"Indiana University","correspondingAuthor":false,"prefix":"","firstName":"David","middleName":"J.","lastName":"Crandall","suffix":""}],"badges":[],"createdAt":"2025-11-13 04:53:11","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8101465/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8101465/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":98627593,"identity":"7ada0fe4-2770-483f-b8c5-940f335aa278","added_by":"auto","created_at":"2025-12-19 17:10:29","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":205658,"visible":true,"origin":"","legend":"","description":"","filename":"Manuscript11142025.docx","url":"https://assets-eu.researchsquare.com/files/rs-8101465/v1/2622c58f3693825aedb8a697.docx"},{"id":98582167,"identity":"67e69f31-a94a-45a0-b209-0090a2b38519","added_by":"auto","created_at":"2025-12-19 08:40:47","extension":"jpg","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":106668,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8101465/v1/a941e3f75b0e55ec8b5f6b69.jpg"},{"id":98627042,"identity":"8a80ea19-22f8-43b2-9923-8216c64bbfc7","added_by":"auto","created_at":"2025-12-19 17:10:06","extension":"jpg","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":107084,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8101465/v1/b876f88992cc0233a0d74b09.jpg"},{"id":98582170,"identity":"4f4cdecc-2e4c-4de4-b261-a151586a9ea8","added_by":"auto","created_at":"2025-12-19 08:40:47","extension":"json","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5930,"visible":true,"origin":"","legend":"","description":"","filename":"a068ba84baa042b1965800a39b3a19a1.json","url":"https://assets-eu.researchsquare.com/files/rs-8101465/v1/82b2d2b698ad51fe5dab3430.json"},{"id":98627839,"identity":"3aa5e6b4-130e-4c75-88ce-dd750bab7df6","added_by":"auto","created_at":"2025-12-19 17:10:42","extension":"xml","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":121219,"visible":true,"origin":"","legend":"","description":"","filename":"a068ba84baa042b1965800a39b3a19a11enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8101465/v1/eb9df7559d1a8ec1cbef5315.xml"},{"id":98582175,"identity":"f0a7f9b2-f266-4b3e-8539-e04aa156501d","added_by":"auto","created_at":"2025-12-19 08:40:47","extension":"jpg","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":106668,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8101465/v1/81aad4620752f86067b8adc7.jpg"},{"id":98628815,"identity":"9234f367-d72d-4511-82d0-a8aeb3759b3f","added_by":"auto","created_at":"2025-12-19 17:12:34","extension":"jpg","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":107084,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8101465/v1/ee953fb8e3b84abc65714c99.jpg"},{"id":98582171,"identity":"ed93fdc4-70fb-4380-a6c8-b7fb4e091971","added_by":"auto","created_at":"2025-12-19 08:40:47","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":61129,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8101465/v1/e63b91561e1f6114e23e85a9.png"},{"id":98628498,"identity":"2d35cd8a-8816-4066-b845-74d2cb86a32a","added_by":"auto","created_at":"2025-12-19 17:11:37","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":64081,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8101465/v1/1090fad0dbc5160579bafdba.png"},{"id":98628691,"identity":"3d605f65-4419-4b4b-8e7e-a53b6812a384","added_by":"auto","created_at":"2025-12-19 17:12:03","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":33236,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure1.png","url":"https://assets-eu.researchsquare.com/files/rs-8101465/v1/a87f93a4f186f8629eb034ba.png"},{"id":98627084,"identity":"99274368-8078-4c08-a953-672df5440398","added_by":"auto","created_at":"2025-12-19 17:10:07","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":34043,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure2.png","url":"https://assets-eu.researchsquare.com/files/rs-8101465/v1/be6eb2830ba1a67b42b41af6.png"},{"id":98628389,"identity":"11669523-a900-4ea9-83fe-d1329ae7d027","added_by":"auto","created_at":"2025-12-19 17:11:23","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":20811,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8101465/v1/15d5a644f373f5c39208f7cc.png"},{"id":98627611,"identity":"497f6141-522e-43a1-a949-b29d311ce372","added_by":"auto","created_at":"2025-12-19 17:10:30","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":21686,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8101465/v1/8c8e7a63f6c233eac6bf76e2.png"},{"id":98627025,"identity":"01048746-e47f-465e-9269-f70763aad7c5","added_by":"auto","created_at":"2025-12-19 17:10:05","extension":"xml","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":122020,"visible":true,"origin":"","legend":"","description":"","filename":"a068ba84baa042b1965800a39b3a19a11structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8101465/v1/10287b6bd53d2025f9a3f229.xml"},{"id":98627264,"identity":"f0eddb23-d6b5-420f-8293-9b21c84c2676","added_by":"auto","created_at":"2025-12-19 17:10:15","extension":"html","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":133990,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8101465/v1/ebd5008f93f000c6b26ab368.html"},{"id":98582166,"identity":"404ffe32-1a90-44df-a9b8-6dd963e2119b","added_by":"auto","created_at":"2025-12-19 08:40:47","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":41853,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of percentage of YouTuber face presence.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8101465/v1/3de67a78a8e65f0834481f44.png"},{"id":98629046,"identity":"2116bc2a-4ab6-4e67-afdb-3b997a5ef7f3","added_by":"auto","created_at":"2025-12-19 17:13:08","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":46799,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of average confidence scores of camera gaze.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-8101465/v1/2b12224187831450cd61d8b2.png"},{"id":102025536,"identity":"1a214f3b-3384-45b6-a254-23e956a76b08","added_by":"auto","created_at":"2026-02-06 09:42:05","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":892326,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8101465/v1/b42f463f-b2fb-4a00-8e23-d533a7d7193b.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Investigating medical YouTubers’ parasocial visual cues in their COVID-related videos: a computer vision approach","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eThe rise of social media has significantly changed the traditional role of scientists and medical experts which previously established a boundary between them and the public [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Today, scientists actively engage in science and health communication across diverse social media platforms, from text-based platforms such as X (formerly Twitter) to video-based platforms like TikTok. This novel phenomenon has received a great deal of academic attention, especially during the COVID-19 pandemic [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Of diverse social media platforms, YouTube played a key role in COVID-19 communication during the pandemic [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. A host of medical YouTubers\u0026mdash;medical professionals who have official licenses or degrees about medicine or other relevant disciplines and upload health-related videos steadily [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]\u0026mdash;operated their channels as reliable COVID-related information sources. This phenomenon is noteworthy given that YouTube was more commonly considered to be an entertainment source in the past rather than a place for discussing health information [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe current study explores medical YouTubers\u0026rsquo; health communication during the COVID-19 pandemic and extends prior literature both theoretically and methodologically. Theoretically, this study extends prior literature on parasocial interactions, one of the most frequently employed theoretical concepts in current social media research. Prior literature has investigated diverse aspects of parasocial interactions across different platforms [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. However, none of them have explored how medical YouTubers, particularly in the context of the COVID-19 pandemic, used parasocial visual cues to gain stronger user engagement. Thus, the current study expands the domain of parasocial interaction research by applying the concept to an understudied area.\u003c/p\u003e \u003cp\u003eIn terms of methods, the present study applies automated visual analysis via computer vision, which is frequently recognized as a promising tool for advancing current communication and media research [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Peng et al.\u0026rsquo;s recent article on automated visual analysis for social media research argued that these novel methods can afford a multitude of new opportunities to researchers and reinforce the extant research avenues [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. The authors predicted that computer vision would advance current social media research by considerably increasing validity, reliability, and generalizability\u0026mdash;for example, through description of visual media content. The current study utilizes computer vision techniques to measure parasocial visual cues in medical YouTubers\u0026rsquo; COVID-related videos, focusing on two parasocial visual cues manually investigated by previous studies [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]: bodily address and gaze towards the camera.\u003c/p\u003e \u003cp\u003eWith these novel measurements, we will verify whether these two parasocial visual cues are significantly associated with user engagement, represented as the number of likes and comments. Given that few previous studies on parasocial interactions have linked parasocial visual cues to these two commonly used user engagement metrics on social media, we believe that this study fills an important gap in the literature.\u003c/p\u003e"},{"header":"2. Literature review","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1. Medical experts on YouTube\u003c/h2\u003e \u003cp\u003eHealth communication on YouTube has become a critical component of public health outreach, offering accessible information to a global audience [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. In the current social media landscape, where video has risen to be a key media type [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e], YouTube is positioned as a platform in which individuals can access health-related content ranging from wellness tips [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] to expert medical advice [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Health communication on YouTube is especially impactful because of its visual and interactive nature [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], making complex medical information more digestible for a wide range of viewers [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Additionally, the platform allows health/medical experts and organizations to quickly reach a large audience, enabling the rapid dissemination of public health messages, especially during urgent health crises like the COVID-19 pandemic [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eMedical YouTubers played a pivotal role in shaping public health responses during the COVID-19 pandemic. Their influence proved to be enormous, with multiple channels accruing a massive number of subscribers and viewers. For instance, one of the most popular medical YouTubers, \u003cem\u003eDoctor Mike\u003c/em\u003e, has about 14.27\u0026nbsp;million subscribers as of August 2025, which is more than 20 times greater than the official YouTube channel of the U.S. Centers for Disease Control and Prevention (CDC; 0.69\u0026nbsp;million subscribers). This stark contrast represents the growing role of YouTubers in health communication. Building on this influence, many medical YouTubers took steps to provide accurate and timely information by interviewing key experts during the pandemic. Notably, Doctor Mike invited Anthony Fauci, the then-Director of the National Institute of Allergy and Infectious Diseases (NIAID), to discuss the U.S. government\u0026rsquo;s guidelines regarding the virus on March 29, 2020, only 15 days after the WHO\u0026rsquo;s declaration of COVID-19 as a pandemic [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Considering that Fauci was the central figure of the entire nation\u0026rsquo;s COVID-19 countermeasures, this underlines the fact that medical YouTubers\u0026rsquo; positions have evolved beyond the roles traditionally expected of them. Such collaborations exemplify how medical YouTubers leveraged their channels to spread critical health information and elevate the voices of relevant experts during the crisis.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2. Parasocial interactions between medical YouTubers and viewers\u003c/h2\u003e \u003cp\u003eThe concept of \u0026ldquo;parasocial interaction\u0026rdquo; was originally suggested by Horton and Wohl [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e], and they defined parasocial interaction as a \u0026ldquo;simulacrum of conversational give and take\u0026rdquo; (p. 215). Building on the seminal work, Perse and Rubin updated this concept particularly for communication and media research [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. They reconceptualized a parasocial interaction as \u0026ldquo;a perceived interpersonal relationship on the part of a television viewer with a mass media persona\u0026rdquo; (p. 59) and highlighted that this phenomenon is specific to TV, which was undoubtedly one of the most compelling media at the time. It is worth noting that the related concept of \u0026ldquo;parasocial relationship\u0026rdquo; is often discussed together in previous literature [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. That said, some studies have clearly recommended that researchers should separate these two concepts [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Horton and Wohl\u0026rsquo;s original study [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] introduced parasocial interaction as a momentary interaction-like experience with media figures during a single encounter, such as watching a TV show. In contrast, they described parasocial \u003cem\u003erelationships\u003c/em\u003e as long-term emotional bonds formed over repeated interactions. In the same vein, Schramm and Hartmann [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e] also underscored the need to differentiate short-term, specific parasocial interactions from long-term, generalizable parasocial relationships and emphasized their distinct methodological implications. To avoid potential confusion, we clarify here that this study is focused on visual factors in parasocial interactions, not relationships.\u003c/p\u003e \u003cp\u003eIn fact, until the appearance of social media, parasocial interactions via mass media were considered not only illusionary but sometimes even pathological [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. This perception is understandable in the context of TV viewership, because TV stars are not a counterpart for interaction in most cases, as the foundational meaning of the word \u0026ldquo;star\u0026rdquo; indicates a shining object in the sky which is admired but not approachable. However, the emergence of social media changes both the scope and intensity of parasocial interactions. Many studies explore social media platforms as new venues for parasocial interactions [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e], with a particular focus on video-based platforms [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e], since, like TV, they can support parasocial visual cues. Nevertheless, one significant difference between TV and video-based social media platforms such as YouTube and Twitch is that social media influencers are much more approachable compared to TV stars. The boundary between parasocial interactions and real interactions on social media is less defined, since influencers often interact with their fans via comment threads or direct messages. As such, parasocial interactions are now \u0026ldquo;usual social activity\u0026rdquo; [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e] (p. 280) rather than abnormal or pathological.\u003c/p\u003e \u003cp\u003eThis revised understanding of parasocial interactions has sparked a multitude of studies, some of which have investigated \u0026ldquo;parasocial visual cues\u0026rdquo; that have been considered crucial factors in parasocial interactions. Using TV show clips as stimuli, Cummins and Cui [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] empirically tested how a performer\u0026rsquo;s style of address in a video affects viewers\u0026rsquo; actual parasocial interaction experience. They divided the style of address into three categories: \u003cem\u003ebodily address\u003c/em\u003e, referring to \u0026ldquo;instances where the viewer both sees and hears the mediated performer speaking to the viewer\u0026rdquo; (p. 727); \u003cem\u003everbal address\u003c/em\u003e, indicating instances where the performer speaks in the video without showing their appearance; and \u003cem\u003eno address\u003c/em\u003e. Their results demonstrated that bodily address aroused more pronounced feelings of interaction than either verbal address or no address. Additionally, emotional contagion, which is a key element of empathy, played a crucial role in enhancing these perceived interactions, particularly in response to bodily address.\u003c/p\u003e \u003cp\u003eThe current study extends Cummins and Cui\u0026rsquo;s study [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] by testing if medical YouTubers\u0026rsquo; bodily address is significantly associated with viewer engagement. Although there is an established body of research on parasocial visual cues, there are few studies connecting those cues with user engagement on social media. The present study bridges this gap by investigating how bodily address, a proven parasocial visual cue [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], is associated with two typical user engagement metrics on YouTube: the number of likes and the number of comments [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. These two metrics have been frequently utilized in prior literature as indicators of YouTuber viewers\u0026rsquo; engagement. To illustrate, Munaro et al. [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e] tested how these metrics are associated with language elements, linguistic style, subjectivity, emotion valence, and video category. Therefore, the current study proposes the two research questions below:\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eRQ1-1\u003c/strong\u003e \u003cp\u003eHow is the presence of medical YouTubers in their COVID-related videos associated with the number of likes those videos receive?\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eRQ1-2\u003c/strong\u003e \u003cp\u003eHow is the presence of medical YouTubers in their COVID-related videos associated with the number of comments those videos receive?\u003c/p\u003e \u003c/p\u003e \u003cp\u003eAdditionally, the present study investigates another parasocial visual cue, the direction of gaze. Ferchaud et al.\u0026rsquo;s paper [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], one of the most cited articles on parasocial interactions on YouTube, explored multiple parasocial visual cues in videos from popular YouTube channels, including direction of gaze. The variable was composed of three categories: \u0026ldquo;primarily facing towards camera,\u0026rdquo; \u0026ldquo;face not on camera,\u0026rdquo; and \u0026ldquo;primarily facing away from camera.\u0026rdquo; The authors tested whether this variable was significantly associated with the two parasocial attributes in the study: realism and authenticity. Their results were mixed; while the three categories did not significantly differ in terms of realism, \u0026ldquo;face not on camera\u0026rdquo; was associated with significantly less authenticity than both \u0026ldquo;primarily facing away from camera\u0026rdquo; and \u0026ldquo;primarily facing towards camera.\u0026rdquo; Essentially, their findings revealed that as long as the YouTuber is present, the direction of their gaze does not significantly relate to the video\u0026rsquo;s perceived authenticity.\u003c/p\u003e \u003cp\u003eBuilding on this relevant work [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], the present study investigates how medical YouTubers\u0026rsquo; direction of gaze is associated with user engagement. Since our previous two research questions address the presence of the YouTuber, we simplify the categories for the direction of gaze to \u0026ldquo;looking at the camera\u0026rdquo; and \u0026ldquo;not looking at the camera.\u0026rdquo; Regarding user engagement, again, the number of likes and comments are employed as proxies for the users\u0026rsquo; support and participation in the discussion raised by the YouTuber. We suggest another two research questions regarding the direction of gaze:\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eRQ2-1\u003c/strong\u003e \u003cp\u003eHow is the gaze of medical YouTubers towards the camera associated with the number of likes that their COVID-related videos receive?\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eRQ2-2\u003c/strong\u003e \u003cp\u003eHow is the gaze of medical YouTubers towards the camera associated with the number of comments that their COVID-related videos receive?\u003c/p\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"3. Methods","content":"\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e3.1. Computer vision approach\u003c/h2\u003e \u003cp\u003eComputer vision has made significant strides in advancing social media research by enabling quick and large-scale analysis of visual content, which is crucial for understanding trends and patterns within digital media [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. One key development is the integration of machine learning techniques for efficient processing and interpretation of content shared on image- or video-based platforms such as Instagram, YouTube, and Twitch [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. These techniques enhance the ability to classify and analyze visual data at a much larger scale, which was previously hindered by manual coding processes [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. Peng et al.\u0026rsquo;s recent review paper [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] suggested some major topics that could potentially be better investigated with computer vision: visual politics, mis-/disinformation, digital activism, body image, and digital connections. In fact, the topic of the current study\u0026mdash;parasocial visual cues on social media\u0026mdash;is not mentioned even in Peng et al.\u0026rsquo;s thorough work. Likewise, to our knowledge, our methodological approach to parasocial interactions is truly novel, with no previous attempts of a similar nature.\u003c/p\u003e \u003cp\u003eThat said, it is worth noting that state-of-the-art computer vision techniques, not only for parasocial interaction research but for nearly all tasks, are still far from perfect as research methods [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Even with the reported accuracy of computer vision models approaching 100% in some papers and on some datasets, this does not mean that the results are completely reliable, as modern computer vision algorithms are poor at generalizing to situations and contexts that they have never seen before. Moreover, while computer vision\u0026rsquo;s ability to recognize obvious visual features\u0026mdash;such as faces, objects, and simple actions\u0026mdash;has become fairly advanced, it generally struggles to recognize more subtle semantics based on specific context, such as memes, visual analogies, and symbolic images, which are common on social media [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAccordingly, as an initial attempt in this new research avenue, this study tries to avoid the fallacy of placing excessive faith in computational methods. Put differently, the methodological goal of the current study is to serve as a first steppingstone for future research on parasocial interaction using computer vision, and we share the challenges encountered in our data collection and analysis in the Limitations and Future Directions section.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e3.2. Sampling\u003c/h2\u003e \u003cp\u003eTo collect our sample of medical YouTubers, three researchers conducted a Google search for articles on medical and science-focused YouTube channels with the keywords \u0026ldquo;medical YouTuber\u0026rdquo; and \u0026ldquo;science YouTuber.\u0026rdquo; They then expanded their searches and identified five relevant articles recommending YouTubers who primarily discuss health, medicine, or science. From these articles, an initial list of 21 medical YouTubers was compiled. Non-English-speaking YouTubers were excluded due to language barriers, as were those lacking nationally certified medical credentials or licenses. Channels with fewer than 10 COVID-19-related videos were also removed. After applying these criteria, the final channel list consisted of five channels. As of March 1, 2023, these YouTube channels averaged 2.54\u0026nbsp;million subscribers, with subscriber counts between 24,453 and 10.6\u0026nbsp;million. The sample includes four Doctors of Medicine (M.D.) and one Doctor of Osteopathic Medicine (D.O.), with three YouTubers based in the U.S. and two in the U.K.\u003c/p\u003e \u003cp\u003eWe focused on COVID-19-related videos uploaded to these channels between January 2020 and January 2023. Only videos with at least one of the keywords \u0026ldquo;COVID,\u0026rdquo; \u0026ldquo;Corona,\u0026rdquo; or \u0026ldquo;SARS-CoV-2\u0026rdquo; in their title were included in the analysis. This process yielded 194 COVID-related videos. Of these, 39 videos were additionally ruled out because they showed too many variations of the YouTuber\u0026rsquo;s face; for example, some videos alternated between showing the YouTuber\u0026rsquo;s face with and without a mask, while others shifted between indoor and outdoor locations. After this exclusion, 155 videos remained in our final sample. These videos were, on average, 10 minutes and 54 seconds long and had 922,183 views, 29,191 likes, and 3,673 comments (including replies) as of March 1, 2023. This indicates that the videos garnered roughly one like for every 32 views and one comment for every 251 views.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e3.3. Measures\u003c/h2\u003e \u003cp\u003eFor all 155 videos, we preprocessed each video (24 frames per second) by extracting every other frame to reduce computational cost. From these frames, we measured two parasocial visual cues, \u003cem\u003ebodily address\u003c/em\u003e and \u003cem\u003ecamera gaze\u003c/em\u003e, as independent variables. Regarding the dependent variables, two user engagement metrics\u0026mdash;the number of likes and comments\u0026mdash;were recorded.\u003c/p\u003e \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e \u003ch2\u003e3.3.1. Bodily address\u003c/h2\u003e \u003cp\u003eBased on Cummins and Cui\u0026rsquo;s study [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], bodily address refers to moments when an on-screen performer appears to directly speak to the viewer, which may enhance a feeling of parasocial interaction. To measure this, we used computer vision techniques to identify when the YouTuber is visible in frame. First, we detected all faces in each frame using a face detection model called \u0026ldquo;RetinaFace\u0026rdquo; [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. RetinaFace is a deep\u003cb\u003e-\u003c/b\u003elearning-based Convolutional Neural Network (CNN) model specifically designed for accurate and robust face detection in real-world settings. RetinaFace processes each image by applying convolutional layers that extract learned features such as edges, textures, and patterns relevant to identifying faces, and then estimates bounding boxes (minimally-enclosing rectangles surrounding detected faces) and locations of facial landmarks such as the eyes, nose, and mouth. While not perfect, RetinaFace is able to handle challenging scenarios like partially visible faces, varying lighting conditions, and different facial angles.\u003c/p\u003e \u003cp\u003eOur next step was to determine, among every face detected in each frame of a video, which corresponded to the YouTuber. While we could have trained a face recognition algorithm to recognize specific YouTubers, this would have required a significant amount of manual data labeling effort for each YouTuber. Instead, we took a fully automatic and more scalable approach: we clustered every face appearing in the frames across each video to find groups of face images corresponding to the same people. We then assumed that the largest cluster\u0026mdash;the person whose face appears most on-camera\u0026mdash;was the YouTuber.\u003c/p\u003e \u003cp\u003eWe used a clustering algorithm called DBSCAN [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e], which groups data points based on their density to identify clusters of closely packed points (within a defined radius, \u0026ldquo;epsilon\u0026rdquo;) and separating them from sparser regions. This approach is well-suited for detecting clusters without predefining the number of clusters and can effectively identify outliers. As the input to the clustering, we represented each detected face numerically using another deep convolutional neural network called \u0026ldquo;FaceNet\u0026rdquo; [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. FaceNet processes facial features to create unique \u0026ldquo;embeddings,\u0026rdquo; which are compact numerical representations that capture the distinguishing characteristics of each face. In our study, DBSCAN organized the embeddings into clusters, with the largest cluster assumed to represent the YouTuber\u0026rsquo;s face. When a frame contained a face from this cluster, we counted the YouTuber as being present in that frame. Finally, we calculated the percentage of frames where the YouTuber was present compared to all extracted video frames. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e displays the distribution of this variable.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e\u0026mdash;Insert Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e about here\u0026mdash;\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section3\"\u003e \u003ch2\u003e3.3.2. Camera gaze\u003c/h2\u003e \u003cp\u003eResearch by Ferchaud et al. [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e] highlighted that the direction of YouTubers\u0026rsquo; gaze, including looking directly at the camera, may function as a parasocial visual cue. To assess this, we used a CNN-based gaze detection model [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e] to estimate whether someone is looking at the camera. The model provides a confidence score (ranging from 0 to 1) for each detected face in each image frame, where higher scores indicate a greater likelihood of the performer looking at the camera. A score of 0 indicates that the model believes that the performer is not looking at the camera at all, while a score of 1 means the model is fully confident the performer is looking at the camera. For every frame where the YouTuber was present, we ran the gaze model on the YouTuber\u0026rsquo;s face and then computed the mean of these scores across the video to generate an overall measure of camera gaze. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e displays the distribution of this variable.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e\u0026mdash;Insert Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e about here\u0026mdash;\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section3\"\u003e \u003ch2\u003e3.3.3. User engagement metrics\u003c/h2\u003e \u003cp\u003eAs dependent variables, the number of likes and comments for each of the 155 videos were recorded by entering the URL information into a website that provides YouTube video data, \u003cem\u003eYouTubeLikeCounter.com\u003c/em\u003e. To avoid any possible variation by time, we recorded the like and comments counts within two hours on March 1, 2023.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.4. Analysis\u003c/h2\u003e \u003cp\u003eFour different multiple linear regression analyses were conducted to address the four research questions with the statistical analysis software \u003cem\u003eR\u003c/em\u003e. For RQ1-1 and RQ1-2, we built two multiple regression models where the percentage of the YouTuber\u0026rsquo;s presence (hereinafter \u003cem\u003eYouTuber presence\u003c/em\u003e) was the independent variable, and the number of likes and comments were the dependent variables, respectively. For RQ2-1 and RQ2-2, we created another two multiple regression models where the confidence score of the YouTuber\u0026rsquo;s camera gaze (hereinafter \u003cem\u003ecamera gaze\u003c/em\u003e) was the independent variable, and the number of likes and comments were the dependent variables, respectively. In these two models, the scale of \u003cem\u003ecamera gaze\u003c/em\u003e was adjusted for ease of presentation by multiplying it by 10,000. Additionally, the models for RQ1-1 and RQ2-1, where the dependent variable was the number of likes, analyzed 153 videos instead of all the 155 videos because two videos did not publicly display their like counts.\u003c/p\u003e \u003cp\u003eAcross the four regression models, there were three control variables: \u003cem\u003enumber of views\u003c/em\u003e, \u003cem\u003edays since the upload\u003c/em\u003e (as of March 1, 2024), and \u003cem\u003evideo length\u003c/em\u003e (in seconds). The number of views was a likely covariate, as it could be positively correlated with both the number of likes and comments. It is rather intuitive that the more people watch a video, the more likes and comments it receives. Especially given the high variance in the number of views (\u003cem\u003eM\u003c/em\u003e\u0026thinsp;=\u0026thinsp;766539.76, \u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;1659155.97), the variable was included as a control variable. Additionally, the days since upload (\u003cem\u003eM\u003c/em\u003e\u0026thinsp;=\u0026thinsp;851.69, \u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;237.14) and the video length (\u003cem\u003eM\u003c/em\u003e\u0026thinsp;=\u0026thinsp;693.46, \u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;479.30) were included as control variables, due to their possible influence on the number of likes and comments. For all four regression models, the VIF analysis was conducted to check multicollinearity issues. As a result, no issues with multicollinearity were found; all the VIF values were under the conventional high multicollinearity threshold of 5, as suggested by Shrestha [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e"},{"header":"4. Results","content":"\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e4.1. Descriptives\u003c/h2\u003e \u003cp\u003eThe mean value of \u003cem\u003eYouTuber presence\u003c/em\u003e is 70.76% (\u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;21.68). Thus, on average, around 70% of each video showed the YouTuber, with the range spanning from about 18% to 100%. Regarding \u003cem\u003ecamera gaze\u003c/em\u003e, the mean value was 0.76 (\u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.19), and the range spanned between 0.25 to 0.98. This indicates that the model recognized that the YouTubers were looking at the camera for an average of 76% of the total time they appeared on screen. When it comes to the dependent variables, the mean of \u003cem\u003enumber of likes\u003c/em\u003e was 29191.38 (\u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;46405.57), and the mean of \u003cem\u003enumber of comments\u003c/em\u003e was 3672.51 (\u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;5134.51).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e4.2. Bodily address (RQ1)\u003c/h2\u003e \u003cp\u003eThe regression model for RQ1-1 that addresses how \u003cem\u003ebodily address\u003c/em\u003e is associated with \u003cem\u003enumber of likes\u003c/em\u003e showed that \u003cem\u003ebodily address\u003c/em\u003e (\u003cem\u003eB\u003c/em\u003e\u0026thinsp;=\u0026thinsp;384.53, \u003cem\u003eSE\u003c/em\u003e\u0026thinsp;=\u0026thinsp;114.69, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.001) was positively associated with \u003cem\u003enumber of likes\u003c/em\u003e when controlling for \u003cem\u003enumber of views\u003c/em\u003e, \u003cem\u003evideo length\u003c/em\u003e, and \u003cem\u003edays since the upload\u003c/em\u003e (see Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e for details). These results indicate that for every 1% increase in the appearance of medical YouTubers in their videos, those videos received an average of 385 additional likes from the viewers. The \u003cem\u003eR\u003c/em\u003e\u003csup\u003e2\u003c/sup\u003e value of the model indicates that about 68% of the variance in \u003cem\u003enumber of likes\u003c/em\u003e can be explained by this regression model.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eRegression analysis between bodily address and number of likes. \u003csup\u003e\u0026dagger;\u003c/sup\u003e \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.1; \u003csup\u003e*\u003c/sup\u003e \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05; \u003csup\u003e**\u003c/sup\u003e \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.01; \u003csup\u003e***\u003c/sup\u003e \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCoefficient\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cem\u003eSE\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cem\u003eP\u003c/em\u003e value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003ePredictor\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBodily address\u003csup\u003e**\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e384.531\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e114.686\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003eControl variables\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNumber of views\u003csup\u003e***\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.022\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eVideo length (seconds)\u003csup\u003e\u0026dagger;\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-12.956\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e7.130\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.071\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDays since the upload\u003csup\u003e\u0026dagger;\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e19.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e10.480\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.060\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConstant\u003csup\u003e*\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-26585.796\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e10466.498\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.012\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003eAdjusted \u003cem\u003eR\u003c/em\u003e\u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.677\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e\u0026mdash;Insert Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e about here\u0026mdash;\u003c/p\u003e \u003cp\u003eThe regression model for RQ1-2 that addresses how \u003cem\u003ebodily address\u003c/em\u003e is associated with \u003cem\u003enumber of comments\u003c/em\u003e demonstrated that \u003cem\u003ebodily address\u003c/em\u003e (\u003cem\u003eB\u003c/em\u003e\u0026thinsp;=\u0026thinsp;30.47, \u003cem\u003eSE\u003c/em\u003e\u0026thinsp;=\u0026thinsp;14.53, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.038) was also positively associated with \u003cem\u003enumber of comments\u003c/em\u003e when controlling \u003cem\u003enumber of views\u003c/em\u003e, \u003cem\u003evideo length\u003c/em\u003e, and \u003cem\u003edays since the upload\u003c/em\u003e (see Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e for details). These results indicate that for every 1% increase in the appearance of medical YouTubers in their videos, those videos received an average of 30 additional viewer comments. The \u003cem\u003eR\u003c/em\u003e\u003csup\u003e2\u003c/sup\u003e value of the model indicates that about 56% of the variance in \u003cem\u003enumber of comments\u003c/em\u003e can be explained by this regression model.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eRegression analysis between bodily address and number of comments. \u003csup\u003e*\u003c/sup\u003e \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05; \u003csup\u003e***\u003c/sup\u003e \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCoefficient\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cem\u003eSE\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cem\u003eP\u003c/em\u003e value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003ePredictor\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBodily address\u003csup\u003e*\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e30.473\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e14.526\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.038\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003eControl variables\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNumber of views\u003csup\u003e***\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.002\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.0002\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eVideo length (seconds)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-1.274\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.905\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNS\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDays since the upload\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.256\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.329\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNS\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConstant\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-748.673\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1330.120\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNS\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003eAdjusted \u003cem\u003eR\u003c/em\u003e\u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.558\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e\u0026mdash;Insert Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e about here\u0026mdash;\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e4.3. Camera gaze (RQ2)\u003c/h2\u003e \u003cp\u003eThe regression model for RQ2-1 that addresses how \u003cem\u003ecamera gaze\u003c/em\u003e is associated with \u003cem\u003enumber of likes\u003c/em\u003e showed that \u003cem\u003ecamera gaze\u003c/em\u003e (\u003cem\u003eB\u003c/em\u003e = -2.29, \u003cem\u003eSE\u003c/em\u003e\u0026thinsp;=\u0026thinsp;1.33, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.087) was marginally associated with \u003cem\u003enumber of likes\u003c/em\u003e when controlling \u003cem\u003enumber of views\u003c/em\u003e, \u003cem\u003evideo length\u003c/em\u003e, and \u003cem\u003edays since the upload\u003c/em\u003e (see Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e for details). For the RQ2-2 model that addresses how \u003cem\u003ecamera gaze\u003c/em\u003e is associated with \u003cem\u003enumber of comments\u003c/em\u003e, there was no significant association between the two variables. To summarize, how much a medical YouTuber looks at the camera does not have a significant association with either the number of likes or comments. These findings contrast with those of the previous models regarding bodily address and user engagement, both of which displayed significant relationships.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eRegression analysis between camera gaze and number of likes. The \u003cem\u003ecamera gaze\u003c/em\u003e scale was adjusted for presentation purposes by multiplying it by 10,000. \u003csup\u003e\u0026dagger;\u003c/sup\u003e \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.1; \u003csup\u003e*\u003c/sup\u003e \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05; \u003csup\u003e***\u003c/sup\u003e \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCoefficient\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cem\u003eSE\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cem\u003eP\u003c/em\u003e value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003ePredictor\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCamera Gaze\u003csup\u003e\u0026dagger;\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-2.292\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.330\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.087\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003eControl variables\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNumber of views\u003csup\u003e***\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.022\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eVideo length (seconds)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-10.946\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e7.289\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNS\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDays since the upload\u003csup\u003e*\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e25.248\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e10.599\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.018\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConstant\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12573.063\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e14114.982\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNS\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003eAdjusted \u003cem\u003eR\u003c/em\u003e\u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.659\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e\u0026mdash;Insert Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e about here\u0026mdash;\u003c/p\u003e \u003c/div\u003e"},{"header":"5. Discussion","content":"\u003cp\u003eThis study explores the relationship between parasocial visual cues and user engagement with medical YouTubers\u0026rsquo; videos, with a particular focus on COVID-related health communication. Interestingly, the results for the two parasocial cues showed clear contrasts: bodily address was positively associated with both the number of likes and comments, whereas camera gaze was not significantly associated with either. In other words, medical YouTubers\u0026rsquo; presence in their videos could significantly help increase both the number of likes and comments, while their camera gaze may have nothing to do with viewer engagement.\u003c/p\u003e \u003cp\u003eAs mentioned, our results showed that bodily address has a positive association with users\u0026rsquo; support for the video, which was measured by the number of likes. Considering that our sample is composed of medical YouTubers\u0026rsquo; COVID-related videos, these results may correspond to typical patterns in our \u003cem\u003eoffline\u003c/em\u003e health communication in everyday life; if the situation is serious, we might prefer to see medical professionals in person rather than relying on voice communication over the phone. The unique characteristics of the pandemic are also noteworthy. During this period, we experienced a lockdown situation where physical contact was restricted for safety reasons. In the wake of the lockdown, offline social interactions decreased significantly, which resulted in people using social media as an alternative venue for social interaction [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]. Studies on computer-mediated communication during the pandemic suggest that at the time, social media users were drawn to more active and dynamic forms of communication, as evidenced by the soaring popularity of TikTok [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e]. Medical YouTubers\u0026rsquo; videos might not be an exception in this regard, with viewers preferring more dynamic visual cues from the YouTuber\u0026rsquo;s presence on screen as opposed to \u0026ldquo;B-roll\u0026rdquo; footage.\u003c/p\u003e \u003cp\u003eOur findings about bodily address and the number of likes may also relate to the findings of prior literature on the affective aspect of parasocial interactions. Horton and Wohl\u0026rsquo;s initial work [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] suggested that onscreen performers talking to the audience can increase an \u0026ldquo;illusion of intimacy\u0026rdquo; (p. 217). Extending the literature, Cummins and Cui [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] empirically confirmed that bodily address can significantly enhance viewers\u0026rsquo; affective empathy towards a performer, particularly in the context of TV viewership. Although these studies did not explore the connection between parasocial cues and viewer engagement with the content, they demonstrated that parasocial cues could influence viewers\u0026rsquo; affective aspect, as opposed to their cognitive aspect. The present study adds a novel finding to this body of research: the influence of parasocial cues on viewers\u0026rsquo; affective aspect remains valid in video-based social media, particularly in the context of health communication.\u003c/p\u003e \u003cp\u003eOur findings also show that medical YouTubers\u0026rsquo; bodily address encouraged their viewers to engage more in the comment threads. Writing a comment is, no doubt, a more active form of user engagement and commitment than simply clicking the \u0026ldquo;like\u0026rdquo; button. In fact, prior literature on health communication has extensively examined social media comment threads as venues for user interactivity. Heldman et al.\u0026rsquo;s work [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e], one of the earliest studies investigating health communication and user engagement in social media, underscores interactivity as one of the key features of social media, which distinguishes these platforms as avenues for new types of communication, rather than simple sources of information or news. This importance was further emphasized during the COVID-19 pandemic [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e]; participation in comment threads allowed users to position themselves as active discussants of the COVID-19 situation, as opposed to passive observers. For example, the comment thread under one of our sample videos, titled \u0026ldquo;Why we need LOCKDOWN // UK DOCTOR // Covid-19 Vlog #6\u0026rdquo; from \u003cem\u003eDr Hope\u0026rsquo;s Sick Notes\u003c/em\u003e, is filled with viewer comments about the lockdown, where some clearly expressed their opinions. Thus, the commenters were not merely seeking COVID-19 information but were also voicing their perspectives.\u003c/p\u003e \u003cp\u003eCorresponding to this importance of user participation in comment threads during the pandemic, the role of bodily address may be significant in encouraging COVID-19 discussions. While some previous studies frame bodily address as a production technique that may influence cognitive processes [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e], our findings suggest that effectively deploying this parasocial cue could yield more affective results than typically expected from such a production technique. We recommend that future research further explore how different production techniques can possibly function as parasocial cues, ultimately fostering viewer engagement not only with the video itself but also with the broader topic it addresses.\u003c/p\u003e \u003cp\u003eThe present study did not find a significant association between camera gaze and either the number of likes or comments. Interestingly, this finding is not aligned with previous studies on interpersonal communication, which found that mutual gaze is a key component for engagement [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e]. Since YouTubers\u0026rsquo; gaze towards the camera can create a mediated mutual gaze [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], it would be reasonable to expect a positive correlation between the extent of camera gaze and the number of likes or comments. However, it is also true that our findings are somewhat in line with Ferchaud et al.\u0026rsquo;s study regarding parasocial cues on YouTube [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. They found that while the presence of YouTubers on screen is significantly associated with perceived authenticity, whether they look at the camera or somewhere else is not. Both their study and the current study, which examine parasocial interactions on the same platform, found camera gaze to be a non-significant factor, while YouTuber presence was significant. Since these findings differ from those from previous studies on communication in other settings [\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e], this may indicate that the reason behind the results is related to the affordances of the platform rather than the variable itself [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAlthough the exact reason behind this phenomenon is beyond the scope of the current study, we present one possible explanation based on prior literature. Several studies in communication science have highlighted the importance of screen size in video consumption as a potential factor for viewer engagement and emotional response [\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e]. In traditional TV viewing, content is presented on a large screen, typically located in the living room. In contrast, YouTube videos are often consumed on mobile devices, tablets, or laptops, which have smaller screens in comparison to typical TV sets. As more details in the mise-en-sc\u0026egrave;ne of a video are visible on larger screens than on smaller ones, it is possible that smaller screen sizes render the YouTuber\u0026rsquo;s gaze a less significant factor. We hope future research will probe into this possibility or other potential explanations for this phenomenon.\u003c/p\u003e \u003cp\u003eThe current study employs a novel method, automated visual analysis via computer vision, to explore a long-researched topic, parasocial interactions. While this study has non-negligible limitations as an exploratory attempt, which we will outline in the Limitations and Future Directions section, it also has significant methodological contributions.\u003c/p\u003e \u003cp\u003eFirst, this study proves that computer vision can be a valid research tool for communication and media research by confirming the findings of prior literature based on manual analysis. It is worth noting that our results align well with those of our two key references [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], from which we derived our chosen parasocial visual cues. In fact, few social media studies using computer vision have focused on confirming previous findings [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e]; most have been directed towards exploring new topics rather than validating results from manual analyses [\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e]. While new explorations are undoubtedly worthwhile, we believe that connecting traditional, manual research methods and computer vision is essential to help create a continuum in our research, rather than separating it into distinct avenues based on methodology.\u003c/p\u003e \u003cp\u003eThe second methodological contribution of this study is that by employing computer vision, the current study significantly enhances the granularity of observation in parasocial interaction research. Previous studies on parasocial visual cues typically chose a video clip as the unit of their observation [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. In contrast, the current study measured every other frame, which brought remarkably higher granularity. For instance, one of the videos in our sample from \u003cem\u003eDoctor Mike Hansen\u003c/em\u003e\u0026rsquo;s channel is 10 minutes and 26 seconds long. If this video were explored using the traditional approach, only one data point would represent this video. Meanwhile, our computer vision approach analyzed the video as 7,512 data points, demonstrating a significant improvement in granularity.\u003c/p\u003e \u003cp\u003eLastly, this study, which validates computer vision as a research tool for YouTube, offers valuable insights for future research on parasocial interactions on other video-based social media platforms, where manual analysis is less feasible. For example, the platform Twitch, frequently studied for parasocial interactions [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e], presents challenges for manual analysis due to the length of its content; a recent study shows that the average length of popular Twitch streamers\u0026rsquo; videos exceeds six hours [\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e]. This lengthy content may not be amenable to manual research methods, but can be analyzed using computer vision, which likely maintains validity regardless of video duration.\u003c/p\u003e"},{"header":"6. Limitations and future directions","content":"\u003cp\u003eOnce again, we clarify that current computer vision techniques are far from perfect. Thus, in any study, they should be regarded as research tools that can provide one of multiple ways to look at a phenomenon, rather than an absolute pathway to the truth. The current study is no exception, and we additionally share specific methodological limitations in our case.\u003c/p\u003e \u003cp\u003eFirst, in employing computer vision, our sample had to include only videos suitable for our automated approach, which resulted in the exclusion of some valid videos. After our initial sampling, we ruled out 39 videos that were not amenable to computer vision analysis. Those excluded videos that displayed too many variations of the YouTuber\u0026rsquo;s face, either alternating between showing their face with and without a mask or between indoor and outdoor locations, which made it impossible to identify the YouTuber using our current computational approach. Second, the method we used to decide on the YouTuber\u0026rsquo;s facial cluster has a limited level of validity. Our model selected the cluster containing the most frequently shown face throughout the video as the YouTuber. However, it is possible that some faces, closely relevant to the specific video topic, may appear more often in the video than the YouTuber. Hypothetically, a video about the U.S. lockdown during the early pandemic might include images of President Trump more frequently than the YouTuber. Although our manual check revealed no such cases in our sample, future research with large-scale samples should consider this as a possible hurdle.\u003c/p\u003e \u003cp\u003eAdditionally, we suggest possible directions for future research investigating parasocial interactions via computer vision. Ferchaud et al. [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e] suggested \u003cem\u003eposture of the YouTuber\u003c/em\u003e and \u003cem\u003ecamera angle\u003c/em\u003e as potential parasocial visual cues. The authors suggested specific categories for \u003cem\u003eposture of the YouTuber\u003c/em\u003e (e.g., \u0026ldquo;primarily seated,\u0026rdquo; \u0026ldquo;primarily standing,\u0026rdquo; \u0026ldquo;primarily in action\u0026rdquo;) and for \u003cem\u003ecamera angle\u003c/em\u003e (e.g., \u0026ldquo;straight on,\u0026rdquo; \u0026ldquo;looking up at personality/ies,\u0026rdquo; \u0026ldquo;looking down at personality/ies\u0026rdquo;), which could be useful clues for future research. While the current study, as an initial attempt in this research avenue, did not include these variables, we hope that future research will consider these visual cues with computer vision models specifically designed for this purpose.\u003c/p\u003e \u003cp\u003eThis study also has non-methodological limitations. First, the scope of this study is limited, which constrains the generalizability of our findings. We focused on a very specific topic, medical YouTubers\u0026rsquo; health communication during the pandemic; although this is a topic worthy of academic attention, our findings may not apply to other cases involving parasocial interactions with different video topics or on other social media platforms. Our limited sample size may not reflect broader patterns on social media, which future research could consider. This study also did not explore the mechanisms behind the significant correlations between bodily address and the number of likes and comments. Examining the mechanisms through experimental research could substantially enhance our understanding of social media engagement. In such experimental research, key external variables, such as the viewing device or the YouTuber\u0026rsquo;s verbal style, could be controlled, and necessary control variables, such as demographic factors, could be incorporated into the analysis model to yield more accurate results. Specifically, investigating how parasocial cues might encourage viewers to comment on social media videos could be a meaningful topic, not only for health communication but also for understanding broader participation culture in online spaces.\u003c/p\u003e"},{"header":"7. Conclusions","content":"\u003cp\u003eThis study investigates an underexplored topic, the relationship between parasocial visual cues and user engagement with medical YouTubers\u0026rsquo; COVID-related videos, using an emerging research method, computer vision. Based on prior literature [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], bodily address and camera gaze were measured as parasocial visual cues, while the number of likes and comments, which reflect viewers\u0026rsquo; support and engagement with the discussion raised by the YouTuber, respectively, were selected as user engagement metrics. Our linear regression analyses revealed contrasting findings: bodily address was positively associated with both the number of likes and comments, whereas camera gaze was not significantly associated with either. This indicates that medical YouTubers\u0026rsquo; presence in their videos could significantly increase both the number of likes and comments, whereas their camera gaze may have little to do with viewer engagement. While the study has non-negligible limitations, especially in its methods, we believe that these novel findings contribute to the literature on parasocial interactions on social media and that this study serves as a meaningful first step in parasocial interaction research using computer vision, offering suggestions for future research directions.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe dataset analyzed in this study is available at https://osf.io/efmgn/overview?view_only=0de589ba3d6145ce84d070ad1dc4ac28.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interest.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research was funded by the U.S. National Science Foundation (award number 2152423).\u0026nbsp;\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eGieryn, T. F. Boundary-work and the demarcation of science from non-science: strains and interests in professional ideologies of scientists. \u003cem\u003eAm. Sociol. Rev.\u003c/em\u003e\u003cstrong\u003e48\u003c/strong\u003e, 781\u0026ndash;795 (1983).\u003c/li\u003e\n \u003cli\u003eHara, N. \u0026amp; Chae, S. W. Social media affordances for mediated science communication during the COVID-19 pandemic. In \u003cem\u003eSocial Informatics\u003c/em\u003e (eds Hara, N. \u0026amp; Fichman, P.) (Taylor \u0026amp; Francis, 2025).\u003c/li\u003e\n \u003cli\u003eShukla, A. COVID-19 pandemic: an analysis of popular YouTube videos as an alternative health information platform. \u003cem\u003eHealth Inform. J.\u003c/em\u003e\u003cstrong\u003e27\u003c/strong\u003e, 1460458221994878 (2021).\u003c/li\u003e\n \u003cli\u003eChae, S. W., Hara, N., Shiroiya, H. R., Chen, J. \u0026amp; Ogihara, E. Being vulnerable with viewers: exploring how medical YouTubers communicated about COVID-19 with the public. \u003cem\u003ePLoS ONE\u003c/em\u003e\u003cstrong\u003e19\u003c/strong\u003e, e0313857 (2024).\u003c/li\u003e\n \u003cli\u003eMiller, K. \u003cem\u003ePlaying along: Digital games, YouTube, and virtual performance\u003c/em\u003e (Oxford Univ. Press, 2012).\u003c/li\u003e\n \u003cli\u003eFerchaud, A., Grzeslo, J., Orme, S. \u0026amp; LaGroue, J. Parasocial attributes and YouTube personalities: exploring content trends across the most subscribed YouTube channels. \u003cem\u003eComput. Hum. Behav.\u003c/em\u003e\u003cstrong\u003e80\u003c/strong\u003e, 88\u0026ndash;96 (2018).\u003c/li\u003e\n \u003cli\u003eSherrick, B., Smith, C., Jia, Y., Thomas, B. \u0026amp; Franklin, S. B. How parasocial phenomena contribute to sense of community on Twitch. \u003cem\u003eJ. Broadcast. Electron. Media\u003c/em\u003e\u003cstrong\u003e67\u003c/strong\u003e, 47\u0026ndash;67 (2022).\u003c/li\u003e\n \u003cli\u003eKim, M. \u0026amp; Bas, O. Seeing the Black Lives Matter movement through computer vision? an automated visual analysis of news media images on Facebook. \u003cem\u003eSoc. Media Soc.\u003c/em\u003e\u003cstrong\u003e9\u003c/strong\u003e, 20563051231195582 (2023).\u003c/li\u003e\n \u003cli\u003ePeng, Y., Lock, I. \u0026amp; Salah, A. A. Automated visual analysis for the study of social media effects: opportunities, approaches, and challenges. \u003cem\u003eCommun. Methods Meas.\u003c/em\u003e\u003cstrong\u003e18\u003c/strong\u003e, 163\u0026ndash;185 (2024).\u003c/li\u003e\n \u003cli\u003eCummins, R. G. \u0026amp; Cui, B. Reconceptualizing address in television programming: the effect of address and affective empathy on viewer experience of parasocial interaction. \u003cem\u003eJ. Commun.\u003c/em\u003e\u003cstrong\u003e64\u003c/strong\u003e, 723\u0026ndash;742 (2014).\u003c/li\u003e\n \u003cli\u003eMadathil, K. C., Rivera-Rodriguez, A. J., Greenstein, J. S. \u0026amp; Gramopadhye, A. K. Healthcare information on YouTube: a systematic review. \u003cem\u003eHealth Inform. J.\u003c/em\u003e\u003cstrong\u003e21\u003c/strong\u003e, 173\u0026ndash;194 (2015).\u003c/li\u003e\n \u003cli\u003eSchellewald, A. Understanding the popularity and affordances of TikTok through user experiences. \u003cem\u003eMedia Cult. Soc.\u003c/em\u003e\u003cstrong\u003e45\u003c/strong\u003e, 1568\u0026ndash;1582 (2023).\u003c/li\u003e\n \u003cli\u003eHerrmann, L. K. \u0026amp; Elutilo, B. B. Exercise with YouTube: an evaluation of YouTube\u0026rsquo;s workout #withme playlist. \u003cem\u003eInt. J. Health Wellness Soc.\u003c/em\u003e\u003cstrong\u003e14\u003c/strong\u003e, 21\u0026ndash;36 (2023).\u003c/li\u003e\n \u003cli\u003eKohler, S. \u0026amp; Dietrich, T. C. Potentials and limitations of educational videos on YouTube for science communication. \u003cem\u003eFront. Commun.\u003c/em\u003e\u003cstrong\u003e6\u003c/strong\u003e, 581302 (2021).\u003c/li\u003e\n \u003cli\u003eO\u0026rsquo;Mara, B. Social media, digital video and health promotion in a culturally and linguistically diverse Australia. \u003cem\u003eHealth Promot. Int.\u003c/em\u003e\u003cstrong\u003e28\u003c/strong\u003e, 466\u0026ndash;476 (2013).\u003c/li\u003e\n \u003cli\u003eOsman, W., Mohamed, F., Elhassan, M. \u0026amp; Shoufan, A. Is YouTube a reliable source of health-related information? a systematic review. \u003cem\u003eBMC Med. Educ.\u003c/em\u003e\u003cstrong\u003e22\u003c/strong\u003e, 382 (2022).\u003c/li\u003e\n \u003cli\u003eHorton, D. \u0026amp; Wohl, R. R. Mass communication and para-social interaction: observations on intimacy at a distance. \u003cem\u003ePsychiatry\u003c/em\u003e\u003cstrong\u003e19\u003c/strong\u003e, 215\u0026ndash;229 (1956).\u003c/li\u003e\n \u003cli\u003ePerse, E. M. \u0026amp; Rubin, R. B. Attribution in social and parasocial relationships. \u003cem\u003eCommun. Res.\u003c/em\u003e\u003cstrong\u003e16\u003c/strong\u003e, 59\u0026ndash;77 (1989).\u003c/li\u003e\n \u003cli\u003eDibble, J. L., Hartmann, T. \u0026amp; Rosaen, S. F. Parasocial interaction and parasocial relationship: conceptual clarification and a critical assessment of measures. \u003cem\u003eHum. Commun. Res.\u003c/em\u003e\u003cstrong\u003e42\u003c/strong\u003e, 21\u0026ndash;44 (2016).\u003c/li\u003e\n \u003cli\u003eLiebers, N. \u0026amp; Schramm, H. Parasocial interactions and relationships with media characters: an inventory of 60 years of research. \u003cem\u003eCommun. Res. Trends\u003c/em\u003e\u003cstrong\u003e38\u003c/strong\u003e, 4\u0026ndash;31 (2019).\u003c/li\u003e\n \u003cli\u003eDibble, J. L. \u0026amp; Rosaen, S. F. Parasocial interaction as more than friendship. \u003cem\u003eJ. Media Psychol.\u003c/em\u003e (2011).\u003c/li\u003e\n \u003cli\u003eHartmann, T. \u0026amp; Goldhoorn, C. Horton and Wohl revisited: exploring viewers\u0026rsquo; experience of parasocial interaction. \u003cem\u003eJ. Commun.\u003c/em\u003e\u003cstrong\u003e61\u003c/strong\u003e, 1104\u0026ndash;1121 (2011).\u003c/li\u003e\n \u003cli\u003eSchramm, H. \u0026amp; Hartmann, T. The PSI-process scales: a new measure to capture the intensity and breadth of parasocial processes. \u003cem\u003eCommunications\u003c/em\u003e\u003cstrong\u003e33\u003c/strong\u003e, 385\u0026ndash;401 (2008).\u003c/li\u003e\n \u003cli\u003eKowert, R. \u0026amp; Daniel, E. Jr. The one-and-a-half sided parasocial relationship: the curious case of live streaming. \u003cem\u003eComput. Hum. Behav. Rep.\u003c/em\u003e\u003cstrong\u003e4\u003c/strong\u003e, 100150 (2021).\u003c/li\u003e\n \u003cli\u003eLabrecque, L. I. Fostering consumer\u0026ndash;brand relationships in social media environments: the role of parasocial interaction. \u003cem\u003eJ. Interact. Mark.\u003c/em\u003e\u003cstrong\u003e28\u003c/strong\u003e, 134\u0026ndash;148 (2014).\u003c/li\u003e\n \u003cli\u003eGiles, D. C. Parasocial interaction: a review of the literature and a model for future research. \u003cem\u003eMedia Psychol.\u003c/em\u003e\u003cstrong\u003e4\u003c/strong\u003e, 279\u0026ndash;305 (2002).\u003c/li\u003e\n \u003cli\u003eKabadayi, S. \u0026amp; Price, K. Consumer\u0026ndash;brand engagement on Facebook: liking and commenting behaviors. \u003cem\u003eJ. Res. Interact. Mark.\u003c/em\u003e\u003cstrong\u003e8\u003c/strong\u003e, 203\u0026ndash;223 (2014).\u003c/li\u003e\n \u003cli\u003eKim, J. The meaning of numbers: effect of social media engagement metrics in risk communication. \u003cem\u003eCommun. Stud.\u003c/em\u003e\u003cstrong\u003e72\u003c/strong\u003e, 195\u0026ndash;213 (2021).\u003c/li\u003e\n \u003cli\u003eMunaro, A. C., Barcelos, R. H., Maffezzolli, E. C. F., Rodrigues, J. P. S. \u0026amp; Paraiso, E. C. To engage or not engage? the features of video content on YouTube affecting digital consumer engagement. \u003cem\u003eJ. Consum. Behav.\u003c/em\u003e\u003cstrong\u003e20\u003c/strong\u003e, 1336\u0026ndash;1352 (2021).\u003c/li\u003e\n \u003cli\u003eAraujo, T., Lock, I. \u0026amp; van de Velde, B. Automated visual content analysis (AVCA) in communication research: a protocol for large scale image classification with pre-trained computer vision models. \u003cem\u003eCommun. Methods Meas.\u003c/em\u003e\u003cstrong\u003e14\u003c/strong\u003e, 239\u0026ndash;265 (2020).\u003c/li\u003e\n \u003cli\u003eDeng, J. et al. Retinaface: single-stage dense face localisation in the wild. \u003cem\u003earXiv\u003c/em\u003e (2019).\u003c/li\u003e\n \u003cli\u003eHahsler, M., Piekenbrock, M. \u0026amp; Doran, D. dbscan: fast density-based clustering with R. \u003cem\u003eJ. Stat. Softw.\u003c/em\u003e\u003cstrong\u003e91\u003c/strong\u003e, 1\u0026ndash;30 (2019).\u003c/li\u003e\n \u003cli\u003eSchroff, F., Kalenichenko, D. \u0026amp; Philbin, J. Facenet: a unified embedding for face recognition and clustering. In \u003cem\u003eProc. IEEE Conf. Comput. Vis. Pattern Recognit.\u003c/em\u003e 815\u0026ndash;823 (2015).\u003c/li\u003e\n \u003cli\u003eChong, E. et al. Detection of eye contact with deep neural networks is as accurate as human experts. \u003cem\u003eNat. Commun.\u003c/em\u003e\u003cstrong\u003e11\u003c/strong\u003e, 6386 (2020).\u003c/li\u003e\n \u003cli\u003eShrestha, N. Detecting multicollinearity in regression analysis. \u003cem\u003eAm. J. Appl. Math. Stat.\u003c/em\u003e\u003cstrong\u003e8\u003c/strong\u003e, 39\u0026ndash;42 (2020).\u003c/li\u003e\n \u003cli\u003eOsler, L. \u0026amp; Zahavi, D. Sociality and embodiment: online communication during and after COVID-19. \u003cem\u003eFound. Sci.\u003c/em\u003e\u003cstrong\u003e28\u003c/strong\u003e, 1125\u0026ndash;1142 (2023).\u003c/li\u003e\n \u003cli\u003eGottfried, J. Americans\u0026rsquo; social media use: YouTube and Facebook are by far the most used online platforms among U.S. adults; TikTok\u0026rsquo;s user base has grown since 2021. \u003cem\u003ePew Res. Cent.\u003c/em\u003e (31 Jan 2024).\u003c/li\u003e\n \u003cli\u003eHeldman, A. B., Schindelar, J. \u0026amp; Weaver, J. B. Social media engagement and public health communication: implications for public health organizations being truly \u0026ldquo;social\u0026rdquo;. \u003cem\u003ePublic Health Rev.\u003c/em\u003e\u003cstrong\u003e35\u003c/strong\u003e, 1\u0026ndash;18 (2013).\u003c/li\u003e\n \u003cli\u003eChae, S. W. \u0026amp; Hara, N. Exploring how a YouTube channel\u0026rsquo;s political stance is associated with early COVID-19 communication on YouTube. \u003cem\u003eInf. Commun. Soc.\u003c/em\u003e\u003cstrong\u003e27\u003c/strong\u003e, 618\u0026ndash;644 (2024).\u003c/li\u003e\n \u003cli\u003eHutson, J. P., Smith, T. J., Magliano, J. P. \u0026amp; Loschky, L. C. What is the role of the film viewer? the effects of narrative comprehension and viewing task on gaze control in film. \u003cem\u003eCogn. Res. Princ. Implic.\u003c/em\u003e\u003cstrong\u003e2\u003c/strong\u003e, 1\u0026ndash;30 (2017).\u003c/li\u003e\n \u003cli\u003eGoffman, E. \u003cem\u003eBehavior in public places\u003c/em\u003e (Free Press, 1963).\u003c/li\u003e\n \u003cli\u003eKleinke, C. L. Gaze and eye contact: a research review. \u003cem\u003ePsychol. Bull.\u003c/em\u003e\u003cstrong\u003e100\u003c/strong\u003e, 78\u0026ndash;100 (1986).\u003c/li\u003e\n \u003cli\u003eBurgoon, J. K., Coker, D. A. \u0026amp; Coker, R. A. Communicative effects of gaze behavior. \u003cem\u003eHum. Commun. Res.\u003c/em\u003e\u003cstrong\u003e12\u003c/strong\u003e, 495\u0026ndash;524 (1986).\u003c/li\u003e\n \u003cli\u003eDunaway, J. \u0026amp; Soroka, S. Smartphone-size screens constrain cognitive access to video news stories. \u003cem\u003eInf. Commun. Soc.\u003c/em\u003e\u003cstrong\u003e24\u003c/strong\u003e, 69\u0026ndash;84 (2021).\u003c/li\u003e\n \u003cli\u003eKim, K. J. \u0026amp; Sundar, S. S. Mobile persuasion: can screen size and presentation mode make a difference to trust? \u003cem\u003eHum. Commun. Res.\u003c/em\u003e\u003cstrong\u003e42\u003c/strong\u003e, 45\u0026ndash;70 (2016).\u003c/li\u003e\n \u003cli\u003ePeng, Y. What makes politicians\u0026rsquo; Instagram posts popular? analyzing social media strategies of candidates and office holders with computer vision. \u003cem\u003eInt. J. Press/Polit.\u003c/em\u003e\u003cstrong\u003e26\u003c/strong\u003e, 143\u0026ndash;166 (2021).\u003c/li\u003e\n \u003cli\u003ed\u0026rsquo;Andrea, C. \u0026amp; Mintz, A. Studying the live cross-platform circulation of images with computer vision API: an experiment based on a sports media event. \u003cem\u003eInt. J. Commun.\u003c/em\u003e\u003cstrong\u003e13\u003c/strong\u003e, 21 (2019).\u003c/li\u003e\n \u003cli\u003eChae, S. W. Twitch aggression profile: exploring aggression on a live mixed-media platform. \u003cem\u003eInf. Commun. Soc.\u003c/em\u003e\u003cstrong\u003e1\u0026ndash;19\u003c/strong\u003e (2025).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"parasocial interaction, computer vision, automated visual analysis, YouTube, COVID-19, health communication","lastPublishedDoi":"10.21203/rs.3.rs-8101465/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8101465/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThis study explores the relationship between parasocial visual cues and user engagement with medical YouTubers\u0026rsquo; COVID-19-related videos, using a novel approach with computer vision. Based on prior literature, we measured parasocial visual cues through bodily address\u0026mdash;where the YouTuber is seen speaking to the audience\u0026mdash;and camera gaze\u0026mdash;where the YouTuber is looking at the camera. For user engagement, we recorded the numbers of likes and comments, which respectively reflect viewers\u0026rsquo; support and engagement with the discussions raised by the YouTubers. Our linear regression analyses revealed contrasting findings: bodily address was positively associated with both the number of likes and comments, whereas camera gaze was not significantly associated with either. In other words, medical YouTubers\u0026rsquo; presence in their videos could significantly increase both the number of likes and comments, whereas their camera gaze may have little to do with viewer engagement. Our findings on bodily address align with previous parasocial interaction studies conducted through manual analysis. However, our findings on camera gaze correspond only with literature specific to parasocial interactions on YouTube and diverge from broader arguments in the general parasocial interaction literature. The possible rationales behind these contrasting results are discussed from both theoretical and practical perspectives.\u003c/p\u003e","manuscriptTitle":"Investigating medical YouTubers’ parasocial visual cues in their COVID-related videos: a computer vision approach","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-19 08:40:42","doi":"10.21203/rs.3.rs-8101465/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"83d6d75a-5b96-4050-8fb3-e997443cb73b","owner":[],"postedDate":"December 19th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":59834485,"name":"Health sciences/Health care"},{"id":59834486,"name":"Biological sciences/Psychology"},{"id":59834487,"name":"Social science/Psychology"}],"tags":[],"updatedAt":"2026-02-06T09:41:25+00:00","versionOfRecord":[],"versionCreatedAt":"2025-12-19 08:40:42","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8101465","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8101465","identity":"rs-8101465","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}