This was a retrospective analysis of publicly available data about health conditions posted on Twitter in Pennsylvania. This study was approved by the University of Pennsylvania Institutional Review Board.

We collected tweets originating from Pennsylvania related to 5 of the top causes of death in the United States. The causes of death were then further divided into subcategories: heart disease (heart disease and hypertension), diabetes, stroke, cancer (breast, skin, lung, lymphoma, leukemia, prostate, pancreatic, and ovarian), and chronic lung disease (asthma, chronic obstructive pulmonary disease, COPD).

For each disease, we identified all statistically significant (P<.001) correlations between topics and a binary label indicating whether or not a message contained a reference to the disease. Topics most correlated with asthma were related to first-person accounts of managing the disease (attack and inhaler), discomfort associated with the disease (can’t and breathe), or conditions that pose additional risk (pollution, mold, and dust) such as allergens. The majority of topics associated with cancer referenced some variety of charity campaign (pink, ribbon, and bracelet) or awareness effort (support, awareness, October, and pink). Topics related to stroke were rarely related to cerebrovascular accident, but more often related to other definitions of stroke (eg, golf stroke, paint stroke, and heat stroke). Diabetes, heart disease, and hypertension messages were correlated with topics that focused on disease management (weight loss, insulin, and reduce stress) and lifestyle choices (diet and exercise). Complete topic word clouds for each disease can be found in Multimedia Appendix 2.

There is increasing focus on the potential for big data from digital sources in health care. There are challenges associated with using these sources, as they are not always collected for the purposes of health tracking.

We explored the potential for using Twitter to better understand the Web-based conversation about common health conditions. We identified that in some cases, traditional health metrics are associated with the volume of tweets for a given disease. Although traditional methods of determining disease prevalence are robust, they are often delayed in availability because the process for data acquisition and tracking to determine reliable and valid estimates is considerable. Twitter data are available in real-time, much faster than traditional methods, and with significant volume providing a measure of public discourse about health. While tweets would not replace traditional surveillance in the way initially posed by Google flu trends [14], they do provide something unique that prevalence statistics do not; a narrative about patient and public thoughts, knowledge, and experiences with health. Twitter provides context to the conversation surrounding disease and allows for characterization of public discussion of high prevalence conditions. We identified that individuals are using Twitter to talk about several diseases, although variation exists in the frequency of disease mention and the content.

We observed that people are using Twitter for talking about the most common health conditions in Pennsylvania. Prior work has demonstrated the use of Twitter to monitor influenza [15], postpartum depression [16], concussion [17], epilepsy [18], and migraine [19]. The prevalence of disease has been correlated with the frequency of Twitter posting across a variety of diseases [13,20].

We also identified variability in disease mentions and the specificity of terms. This finding provides us with several insights. First, heart disease and stroke cannot be analyzed without preprocessing owing to the ambiguity of many of the keywords associated with the diseases. To resolve these varying issues, other methods will need to be developed to filter out much of the noise associated with these diseases. However, this finding also assures us that the majority of the language we find associated with other diseases can be analyzed using the open vocabulary methods previously described with minimal preprocessing.

Although disease prevalence often coincides with disease mention on Twitter, we found significant variability. The frequency of mentions of breast cancer on Twitter was several orders of magnitude higher than lung cancer, although lung cancer has a higher rate of death and relatively similar prevalence. Breast cancer has a large social media presence owing to awareness and charity campaigns in conjunction with a large community base from those affected by the disease. Lung cancer is tweeted about less often and is often the result of a pop culture reference from television or a celebrity death.

Traditional metrics provide detailed information about prevalence but not insights about people’s understanding, concerns, and questions about health and disease. Our analysis identified several underlying themes that are specific to some diseases. Asthma tweets included references to personal experiences for both the person with asthma as well as parents expressing concern for their children’s asthma issues. Although the largest portion of tweets for the different types of cancer analyzed often referenced charity and awareness, we observed that across diseases in our sample, cancer conditions had the largest portion of tweets about diagnosis.

Our findings also give insight into potential opportunities for using Twitter to inform public health and health communications practices. Future work could examine temporal relationships between Twitter volume and semantic data and traditional health data over larger timeframes and at varying timescales. Meaningful temporal relationships may indicate that Twitter data have value as an additional signal to augment existing surveillance systems, allowing for more precise health tracking and timely interventions.

Twitter data could enhance community building and engagement. Prior work by Neiger et al [21,22] found that more two-way communication on Twitter between public health entities and individual citizens led to an increase in action and awareness that, in turn, resulted in an improvement in community health. Providing local and state public health entities with more accurate information on the public discourse surrounding health could enhance communication and contribute to the more effective dissemination of pertinent and timely health information to the public.

Finally, understanding the interaction between social media use and individual health can identify opportunities for targeted interventions. Prior work by Park et al [23] showed that interventions targeting the perception of social media interaction have the potential to positively impact individual health. We have shown that it is possible to capture a measure of public perception of individual diseases at the community level via analysis of topics and themes. These methods can be translated to individual subjects, where disease perceptions could be tracked over time and compared with actual measures of health, potentially identifying opportunities for intervention.

We compared data from Twitter for 2012-2015 with disease prevalence from 2014, so there may be some variability by year in these estimates. We evaluated unadjusted data from 1 state, so this may not be representative of the conversation about health conditions across other states or geographic regions. Twitter data primarily originate from urban areas; hence, data may not be the most representative sample across the state of Pennsylvania. Future work could explore variations in language on Twitter relative to the size of geographic regions, socioeconomic factors (eg, race, income, urban or rural), and variations in news events or other triggers. Although our correction method eliminates nondisease references, it does not account for metaphorical and joking tweets. This impacts diseases such as heart disease, diabetes, and hypertension.

The precision of the disease keyword filtering, which is the number of selected tweets that were relevant, is reasonably estimated by the corrected message count. However, the recall of the disease keyword filtering, which is the number of relevant tweets that were selected, is difficult to determine owing to the nature of the data and the subjectivity of relevance in the context of health-related tweets. Hopkins et al [24] provides 3 different models for estimating recall: a hand-coding approach similar to the corrected message count presented here, a supervised learning approach for individual document classification, and a supervised learning approach to estimate document category proportions. Evaluating these methods in terms of cost and accuracy is beyond the scope of this study but should be considered for future work to provide more robust measures of keyword-filtered data quality.

Location identification accuracy is difficult to measure for user-defined locations owing to the relative ambiguity of the data provided. The procedures used to estimate user-defined location provide a “soft” measure of accuracy, but more work is needed to ensure appropriate representation. Additionally, a very small proportion of tweets contains location information, thus, the sample may not be representative of the general Twitter landscape in Pennsylvania. Methods such as those detailed in Liang et al [25] should be considered in future studies to correct for sampling bias.

We identified that the volume of tweets is often related to rates of health conditions across a state. The semantic content provided from Twitter provides insight into public perception and awareness of disease beyond what is available through traditional measures of disease prevalence.