Social media can provide a vast source of insight into a wide variety of applications in public health monitoring and surveillance [1]. The bulk of social media-based health monitoring has relied on Twitter, a microblogging platform with over 300 million active users worldwide [2]. A wide variety of health topics are openly discussed on Twitter [3], providing researchers with a rich source of data for monitoring the spread of disease [4,5], dietary patterns [6,7], drug abuse [8,9], foodborne illness [10,11], and depression [12,13], among many other applications.

While Twitter has strengths as a data source, its dominance in research relative to other public platforms has been explained as a matter of convenience: Twitter provides free APIs to obtain large volumes of random or targeted samples of data [1]. However, microblogs are only one type of social media. Other social media platforms contain different types of data and are used in different ways, like sharing visual media. Image-sharing platforms, such as Flickr, Tumblr, Pinterest, and Instagram, are very popular; for example, Instagram, the most popular image-sharing platform, is more than twice the size of Twitter, with over 700 million active users [14]. Despite their popularity, relatively few public health studies have used these types of platforms as a data source [1].

Most prior health research using image-sharing platforms has focused on lifestyle issues, such as diet and substance use. Mejova et al [15] analyzed posts of food on Instagram, focusing on the relationship between food consumption and obesity. De Choudhury et al [16] also examined food consumption on Instagram, focusing specifically on dietary patterns in locations classified as “food deserts.” Yom-Tov et al [17] and Pless et al [18] examined imagery associated with eating disorders on Flickr and Tumblr, respectively. A few studies have looked at substance use on Instagram, including electronic cigarettes [19], marijuana [20], and opioids [21], as well as the marketing of substances on Instagram [22,23]. Garimella et al [24] looked more broadly at lifestyle choices in Instagram, including diet, physical activity, and drinking.

Many of these studies focused on the text features (eg, hashtags and captions) of the image posts, and some conducted a manual content analysis of the images themselves. Two of the studies cited above-used computer vision—a type of artificial intelligence that can automatically analyze the content of images—to perform automatic identification of certain types of images. Pless et al [18] built image classification models to identify images promoting anorexia, since such content may not be tagged with informative text captions. Garimella et al [24] attempted to estimate county-level health statistics from social media content and experimented with both text features and automatically-extracted image tags, finding that both types of information could be correlated with external health metrics.

This study seeks to characterize the health content shared on Instagram, the most popular image-sharing site, toward the goal of identifying potential areas of research that may benefit from this type of data source. In particular, we consider the following research questions: (1) what health topics are prominently shared in Instagram, and (2) what are the characteristics of those topics, specifically the types of images associated with the topics? This study is related to exploratory topic analyses of other platforms for health research [3,25], with an additional contribution of characterizing the features of images in addition to text. The dataset is made available as a resource to the public health informatics community.

This study was reviewed by the University of Colorado Boulder Institutional Review Board, which determined that it does not constitute human subjects research. However, given that publicly posted images may still be considered sensitive material by the users, we took steps to preserve privacy. We did not download any images as part of our study. Instead, we collected the URL pointers to images hosted on Instagram, which were processed by an external computer vision application programming interface (API). Our data collection only contains the abstract features extracted from the images.

While Instagram provides APIs for specific applications, Instagram does not provide an API to collect public data [26], in contrast to Twitter which provides widely-used streaming APIs [27]. We instead built a “crawler” that queries Instagram’s search engine, which returns the nine most recent posts matching a specified hashtag. The crawler accesses the webpage of the hashtag search engine, analogous to how a person would access the search engine in their browser. The page is downloaded, and the HTML is parsed to extract information such as the set of tags and the caption of the image.

We iteratively queried the search engine for 269 general health-related keywords that were used in previous work to obtain a general collection of health-related tweets [3]. The keywords were obtained from dictionaries of terms related to diseases, symptoms, and treatments, in addition to general words like “sick” and “health” that were added manually. The original keyword set contained over 20,000 terms, which were reduced to 269 words that were most common in Twitter, to conform to API limits on how many keywords can be searched. While developed for Twitter, we use the same list for Instagram here, as the list contains a broad set of terms that have previously been shown to be useful for collecting health-related social media posts.

Instagram only allows searching for hashtags rather than free text, so we treated each keyword as a hashtag (eg, “#flu” instead of “flu”). We repeated these 269 queries continuously from September 29, 2016, through October 25, 2016, attempting to simulate a “streaming” collection as with Twitter, and obtained 174,517 posts. We did not download the same post more than once as measured by a unique post identifier. However, if the same content was shared in multiple posts (eg, if multiple users shared the same image), these posts would be considered separate in the dataset.

Each post includes an image, a set of hashtags, and an optional free text caption. We used langid.py to identify and remove posts containing non-English captions [28]. We also removed hyperlinks, and nonalphanumeric characters. Stop words were removed using the natural language toolkit (NLTK) [29]. After filtering and processing, our dataset contained 96,426 documents posted by 77,327 users with an average of 1.25 posts per user.

We extracted “tag” features from each image using Microsoft’s Computer Vision API [30]. This service returns key phrase descriptors of images, such as “person” or “running”. The API was able to extract at least 1 tag for 79.24% (76,407/96,426) of images in our dataset. We refer to these tags as “image tags” to distinguish them from hashtags.

Once extracted, we treated the image tags as an additional type of text, along with captions and hashtags. In the final collection, there are 96,426 posts with a nonempty list of hashtags for an average of 15.2 tags per post. There were 95,208 posts (95,208/96,426; 98.74%) with nonempty captions for an average of 21.6-word tokens per caption. There were also 76,407 posts (76,407/96,426; 79.24%) with at least 1 image tag for an average of 3.7 tags per image. The dataset is shown in Multimedia Appendix 1.

We use probabilistic topic models [31] to characterize the major themes of health-related discussion in Instagram. Topic models are tools for clustering related words into themes or concepts called “topics” and for identifying the topic composition of documents. Topic models have been used in health research as a method of performing content analyses of large datasets [25,32-34].

A topic model is a statistical model with many latent variables and parameters that can be inferred by fitting the model to data. In this model, each “topic” has a probability distribution over words, estimated from data, and topics are usually represented by presenting the 10-20 most probable words in the topic. Additionally, each document has a probability distribution over topics, which can be used to characterize the topic composition of a document and to identify documents that describe particular topics.

Topic models take documents, represented as vectors of word counts, as input. The model parameters (ie, the distribution over topics θ_d in each document d, and the distribution over words ϕ_k for each topic k) are estimated to fit the observed data (ie, the word counts in each document). The estimated parameters are often interpretable to people, and the words associated with each topic cluster can be used to assign a concept to the topic [35].

The topic model gives the probability of each topic k in each document d, θ_dk. To get the overall prevalence of each topic, we calculate the average proportion, where D is the number of documents: To estimate the prevalence of the coarser categories, we simply define each category’s prevalence of the sum of its topic proportions, where C_j is the set of topics in category j: To summarize the degree to which 2 topics have a tendency to occur together in documents, we calculate the Pearson correlation of the θ_dk values for all pairs of topics. Most of the topic pairs with high correlations were similar topics in the same category. To discover less obvious topic co-occurrences, we focus only on topic pairs from different categories.

The topic model results show a large and diverse set of health topics are discussed in Instagram. Qualitatively, we find that the top hashtags tend to be the best descriptors of topics, while caption words give some indication of what kind of messages are associated with the topics, such as whether they are more informational or conversational. The extracted image tags are generally much less coherent, though they do help characterize the types of images that are associated with each topic. For example, many of the topics related to pain contain images of animals, perhaps because users post cheerful images in response to pain. The tag “posing” appears in some topics, suggesting these posts may be informational rather than personal. The Cancer topic contains the image tags “group” and “crowd;” it appears these many posts in this topic are about cancer awareness events. In some cases, image tags were the defining characteristics that distinguished clusters that were otherwise very similar, which suggests that images are informative beyond the hashtags and captions to conduct content analyses of Instagram posts.

Qualitatively, it appears that in most cases the image tags are not specific enough to be useful for directly identifying posts relevant to a specific health application. However, tags of food and beverages appear to be fairly specific and accurate, suggesting that computer vision may help in identifying posts for studies of diet and food consumption. The only previous work we are aware of that used automatically extracted image tags for this purpose is [24], which found that image tags were predictive of lifestyle factors; for example, “glass”, “liquid” and “beverage” were associated with alcohol consumption. The authors suggested that image tags may be useful for identifying stigmatizing behaviors, where social media users may post images of an activity but not explicitly tag the activity. Even nonstigmatized activities, like general food consumption, may not be tagged by a user in a way that is specific enough to identify by text search, while computer vision may help. We observed that image tags extracted from the computer vision API did not usually identify a specific dish, but could at least identify broad categories, like “meat” and “vegetable,” and in some cases were more specific, like “potato” and “doughnut.” We, therefore, argue that this type of computer vision tool can expand the amount of data available for studying patterns in food consumption.

We gained additional insights by considering the co-occurrences of health topics in the data. For example, the Vitamins & supplements topic is less likely to appear in a post in isolation but instead co-occurs with other topics, likely because supplements are discussed in the context in which they are used. Using this data to study nutrition in a population may, therefore, be able to show how nutrition is discussed and applied to specific aspects of health. Some pairs of topics with high correlations may indicate comorbidities, such as Chronic illness and Mental health [39].

Co-occurrences with nonhealth topics may give insights into other contexts in which health is discussed. We observed that many health topics frequently co-occur with inspirational topics, such as topics containing poetry and quotes, or topics about nature, as well as topics related to spirituality and religion. These types of posts may give insight into how individuals cope with and support others with, illness and disease.

An additional observation in some topics is the use of certain hashtags to identify a specific community of users [40], such as the #wlscommunity in the weight loss topics. Online health communities have been studied to understand social support and behavior change in managing health conditions [41]. Instagram-based communities may be a unique source for studying similar issues. Communities for specific demographic groups (eg, #girlswholift) are also present. In some cases, demographic associations could be gleaned from the image tags, even if the text tags were not explicitly gendered.

One of our methodological contributions was to repurpose an existing tool, the polylingual topic model, for a new task of combining different modalities of data in a topic model. We showed that automatically extracted image tags from a computer vision API can be treated as text tokens in an existing topic model. Beyond topic models, our observations of the results suggest that these extracted image tags are in some cases useful descriptors of images. We suggest that this type of tool can be applied to images for health research more broadly.

We observed the same broad set of topics in Instagram that have previously been seen in Twitter [3], suggesting that Instagram could serve as a potential data source for many of the same applications for which Twitter is used. Moreover, the presence of the first-person language (eg, “i’ve”) in some topics indicates that health posts on Instagram include personal health mentions, which is an essential characteristic for some types of surveillance [42]. This has implications for social media-based health surveillance because this suggests that Instagram could be used as a data source for similar areas of research, while having the potential benefit of covering a larger population than Twitter. We do not suggest that one platform is universally better than another, but instead, using data from multiple platforms can result in better surveillance than reliance on one platform [43,44].

Instagram may complement Twitter as a data source because it has a different demographic distribution. The user base of Instagram is younger, lower income, and more urban compared to Twitter, [45]. These demographics cover populations that are traditionally harder to reach in health research [46], and so Instagram may be well-suited for studying such populations. This argument has been made for using Twitter [47], yet Instagram has an even heavier bias toward these populations. Additionally, Instagram has a gender bias that Twitter does not have, being nearly 50% more popular among women than men [45].

Not all health topics are discussed widely on Instagram, which may be a limitation of using Instagram. By far the most common topics in Instagram are related to diet and exercise, while topics on acute illness, which would be needed for a task like influenza surveillance, are the least common. This may explain why all prior work we identified using image-sharing platforms for health research was related to lifestyle factors, such as diet (most common) and physical activity. Nonetheless, topics about infectious disease do exist on Instagram, and so it may be worth investigating the utility of contributing this data to an ensemble surveillance system [43]. To the best of our knowledge, no prior work has studied Instagram for infectious disease surveillance, which would be a good candidate for future research. However, this study did not collect Instagram posts from a large enough span of time to validate the data for such a task.

Another limitation of using Instagram is the limited availability of metadata. When crawling Instagram, it is difficult to sample data uniformly across time, as Instagram does not provide a streaming API analogous to Twitter’s widely-used APIs, which would make it difficult to extract the long-term pattern, for example, to validate influenza tracking. Location data also appears to be difficult to obtain when crawling from the web. In our dataset, 46% of posts contained a user-specified location string, but these were not in a standard format, and many of them were names of businesses or other specific locations, without reference to a geographic area. Geolocation from Instagram is less well understood in social media research, as well as inference of other demographic attributes that may be important in public health research. Richer data may be available from certain resellers; for example, Gnip, who is the official seller of Twitter data, also sells Instagram data, which can be searched by either tag or geolocation.

In addition to limitations of this data, there are limitations with the topic model methodology. Topic model evaluation is notoriously difficult [35], though research has found that this methodology can provide overlapping insights with more traditional, manual text analysis [48]. There is subjectiveness in choosing the number of topics and labeling and categorizing the topics, which we mitigated by having two researchers involved in each step. An advantage of the topic modeling approach for this study is that it can be applied to the entire dataset of nearly 100,000 posts, and the word distributions highlight the features associated with each topic across the three modalities. Furthermore, with such a large number of topics in the data (ie, 47 health topics identified by the approach used here), a typical sample size for manual content analysis, on the order of 1,000 posts [19-21], would be insufficient for accurately learning the prevalence of each topic.

Another limitation of topic modeling is that the topics characterize what is being discussed, but it is difficult to describe how the content is presented. For example, the topic model can identify posts that are related to marijuana, but it does not distinguish between personal marijuana use, information about marijuana, or advertisements for cannabis products—distinctions that have been made in prior work using more qualitative methods [20]. However, the topic model is still an essential first step of filtering and retrieval, after which topic-specific posts could be analyzed in more depth.

We note that there exist other methods for identifying thematic patterns in text beyond probabilistic topic models that have been used in health research, such as network-based clustering on term co-occurrence graphs [49]. Most such methods, including topic models, rely on co-occurrence statistics of words and have similar properties and limitations. We used the polylingual topic model due to its ability to integrate different “languages” or modalities of data.

Finally, the grouping of topics into ten overlapping categories is also limited. Some topics were difficult to categorize, and the boundaries of some categories were difficult to define. However, the goal of the categorization is to present the raw results (available from our dataset in Multimedia Appendix 1) more concisely. The mapping of topics to categories is transparent (viewable in Multimedia Appendix 2) so that the results can be interpreted correctly.

This study shows that health is discussed on Instagram in a variety of ways, and there is potential for computer vision techniques to automatically characterize health-related images, which could extend public health surveillance of social media beyond text-based analysis. Our dataset of nearly 100,000 posts is available to allow for the study of specific topics and image tags in more depth. There are pragmatic reasons why this popular platform has been used in research relatively little compared to platforms like Twitter and Facebook, but our results and discussion point to ideas that image-sharing platforms like Instagram may complement other social media data sources in health research.