Using Twitter to Measure Public Discussion of Diseases: A Case Study

Abstract

BackgroundTwitter is increasingly used to estimate disease prevalence, but such measurements can be biased, due to both biased sampling and inherent ambiguity of natural language.ObjectiveWe characterized the extent of these biases and how they vary with disease.MethodsWe correlated self-reported prevalence rates for 22 diseases from Experian’s Simmons National Consumer Study (n=12,305) with the number of times these diseases were mentioned on Twitter during the same period (2012). We also identified and corrected for two types of bias present in Twitter data: (1) demographic variance between US Twitter users and the general US population; and (2) natural language ambiguity, which creates the possibility that mention of a disease name may not actually refer to the disease (eg, “heart attack” on Twitter often does not refer to myocardial infarction). We measured the correlation between disease prevalence and Twitter disease mentions both with and without bias correction. This allowed us to quantify each disease’s overrepresentation or underrepresentation on Twitter, relative to its prevalence.ResultsOur sample included 80,680,449 tweets. Adjusting disease prevalence to correct for Twitter demographics more than doubles the correlation between Twitter disease mentions and disease prevalence in the general population (from .113 to .258, P <.001). In addition, diseases varied widely in how often mentions of their names on Twitter actually referred to the diseases, from 14.89% (3827/25,704) of instances (for stroke) to 99.92% (5044/5048) of instances (for arthritis). Applying ambiguity correction to our Twitter corpus achieves a correlation between disease mentions and prevalence of .208 ( P <.001). Simultaneously applying correction for both demographics and ambiguity more than triples the baseline correlation to .366 ( P <.001). Compared with prevalence rates, cancer appeared most overrepresented in Twitter, whereas high cholesterol appeared most underrepresented.ConclusionsTwitter is a potentially useful tool to measure public interest in and concerns about different diseases, but when comparing diseases, improvements can be made by adjusting for population demographics and word ambiguity.

Statistical physics of vaccination

“…A recent study on Twitter shows that the time series of the word zombie is a very good predictor of the seasonal flu [724]. Furthermore, matching searches or keywords to context (in this case disease) without taking in account that the same word can be used in discussions completely unrelated to illness might introduce spurious correlations or decrease the signal to noise ratio [753]. Notwithstanding, it is fundamental to acknowledge the potential of these tools in the fight against infectious diseases.…”

Characterizing Tweet Volume and Content About Common Health Conditions Across Pennsylvania: Retrospective Analysis (Preprint)

“…Prevalence of disease has been correlated with frequency of Twitter posting across a variety of diseases. [13][20]…”

“…The scaling is accomplished via a correction factor based on the manual review of tweets by two researchers using the methods outlined in Weeg, et al [13] To calculate the correction factor for a disease, a sample of 30 tweets for each keyword were sampled. Those tweets were then classified as being a reference to disease or not a reference to a disease.…”

Characterizing Physical Activity in a Health Social Network

“…Self-reported diagnoses have been used to examine mental health conditions, such as depression, PTSD, bipolar disorder, and seasonal affective disorder [13,12], and physical health conditions, such as asthma and diabetes [30], and flu [27]. Kumar et al [19] studies the Werther effect, which describes the increased rate of completed or attempted suicides following the coverage of a celebrity's suicide in the media.…”