Background

Natural language processing methods: A brief overview

Natural language processing combines linguistics, artificial intelligence, and machine learning to enable machines to process human language, such as that used in social media posts, and to extract meaning from it. ^{Reference Ding19} Social media analysis requires that text is preprocessed prior to analysis. For example, text is “cleaned” by removing hyperlinks, user tags, or words with little analytical value (eg, “is” or “and”). Words are often then lemmatized, or brought back to their root form (ie, “distancing” to “distance” or “viruses” to “virus“), and text is segmented into 1- and 2-word phrases called named unigrams and bigrams, respectively. Words and phrases with extremely low or high frequency can be removed for simplification and dimensionality reduction. Machine-learning algorithms are used to identify clusters of texts grouped by similar words in a process known as topic modeling, which allows researchers to visualize the distribution and frequency of topics over time or to explore more granular analyses for a particular topic cluster. ^{Reference Medford, Saleh, Sumarsono, Perl and Lehmann8} This methodology requires a level of technical skill; however, companies such as Symplur (Pasadena, CA), an online analytics platform, offer researchers automated natural language processing services to identify and analyze trending language on social media. ²⁰

Early use of web-based data

Examples of surveillance using social media

Social media posts reflect what occupies the minds of people worldwide in near real time creating an increase in interconnectivity termed the “global village.” ²³ Social media may be useful for understanding disease outbreaks through a method known as syndromic surveillance. ^{Reference Wang, Wang, Zhang, Jia, Li and Dang24} The CDC defines syndromic surveillance as a method for using health-related data (like symptom groups) to predict the probability of an impending outbreak. ²⁵ Social media can complement traditional surveillance methods by offering additional data types. For example, the frequency of pertinent hashtags may correlate with disease incidence; geotagged social media entries may indicate crowding and risk of spread; and social media posts can report on school or store closings or supply shortages. ^{Reference Bour, Ahne, Schmitz, Perchoux, Dessenne and Fagherazzi26} Social media reports on public perceptions of nonpharmacological interventions like “work from home” or “wear a mask” orders may be compared with new infection rates to determine effectiveness and to provide early insight into transmission modalities. During the 2009 H1N1 pandemic, Signorini et al ^{Reference Signorini, Segre and Polgreen27} collected tweets with relevant terms to track disease prevalence and public sentiment regarding the infection and prevention efforts. Social media has also been effective in forecasting (ie, the prediction of an event based on past or present data) and monitoring public perceptions and behaviors during the Zika, Dengue, MERS, and Ebola outbreaks. ^{Reference McGough, Brownstein, Hawkins and Santillana3–Reference Odlum and Yoon7}

Applications for COVID-19 surveillance

Our review of social media in surveillance of emerging infections shows, less newsworthy illnesses seem better suited for prediction using social media because there is less influence through media reports. A feared disease, such as Ebola, is likely to receive more press reports; thus, its actual effects on the population may be distorted. Despite this media distortion, social media-derived data have provided valuable insights. Tsao et al ^{Reference Tsao, Chen, Tisseverasinghe, Yang, Li and Butt10} reviewed 91 studies regarding social media in the COVID-19 era and modified a novel framework called “Social Media and Public Health Epidemic and Response (SPHERE).” ^{Reference Tsao, Chen, Tisseverasinghe, Yang, Li and Butt10,Reference Schillinger, Chittamuru and Ramírez29}

Three main categories describe the use of social media as it relates to COVID-19: (1) “Social media as a contagion and a vector” refers to the “infodemic” or rapid dissemination of pandemic information via social media. (2) “Social media as disease control” refers to efforts by public health organizations to create and disseminate materials educating the public with accurate information about the pandemic via social media. (3) “Social media as surveillance” refers to the monitoring of public attitudes, mental health, and case detection during the pandemic. ^{Reference Tsao, Chen, Tisseverasinghe, Yang, Li and Butt10}

Of 81 studies in the third category, only 6 investigated COVID-19 surveillance using social media platforms. Qin et al ^{Reference Qin, Sun and Wang11} collected social media search frequency from a popular Chinese social network (Baidu) for infectious disease key words (eg, dry cough, fever, chest distress, pneumonia, and coronavirus) and used a lagged series of search frequencies to predict suspected case numbers. After evaluating several prediction methods, these researchers used the “subset selection method,” in which 10 of the 50 available predictors that most accurately predicted case numbers were selected and used for analysis. With this method, the number of suspected COVID-19 cases correlated best with search frequencies recorded 6–9 days earlier. Social media search frequencies 10 days earlier correlated best with the number of confirmed COVID-19 cases. ^{Reference Qin, Sun and Wang11} The correlation between social media search indices and suspected COVID-19 cases was stronger than its correlation with PCR confirmed COVID-19 cases.

Li et al ^{Reference Li, Chen, Chen, Zhang, Pang and Chen12} evaluated the prediction strength of search data from Baidu, Google, and Sina Weibo (Chinese social media platform) search engines. They correlated the frequency of 2 search terms, “coronavirus” and “pneumonia,” with daily numbers of suspected and PCR-confirmed COVID-19 cases in China. A correlation was detected with Internet search volume 5–7 days prior to suspected cases and 8–10 days prior to PCR-confirmed cases. ^{Reference Li, Chen, Chen, Zhang, Pang and Chen12} Peng et al ^{Reference Peng, Wang, Liu and Wu13} used geotagged data from Sina Weibo to analyze the spatiotemporal distribution of COVID-19 cases in Wuhan, China. Sina Weibo developed a COVID-19 informational channel early in the pandemic to help patients without access to timely treatment. Help seekers provided information such as name, age, address, contact number, and illness details to an Internet form. Spatial distribution of help seekers reflected the severity of infections clustered in families as well as concentrated regional patterns across the city. Regions with higher population density, especially regions with higher density of elderly individuals, had the highest concentration of COVID-19 help seekers. ^{Reference Peng, Wang, Liu and Wu13}

Machine learning and natural language processing can be used to understand the general public’s perceptions during the pandemic (Public Attitudes – Modified SPHERE framework), and these data can complement surveillance and forecasting efforts. Medford et al ^{Reference Medford, Saleh, Sumarsono, Perl and Lehmann8} studied Twitter data to understand the pandemic’s effect on the emotions, beliefs, and thoughts of the public. ^{Reference Medford, Saleh, Sumarsono, Perl and Lehmann8} They extracted relevant Tweets from the weeks prior and after the activation of the CDC Emergency Operations Center using pertinent hashtags such as #2019nCoV, #coronavirus, #nCoV2019, #wuhancaronavirus, and #wuhanvirus. Variables of interest included tweet text, time, included images or links, the type of tweet (retweet or reply), and the number of likes, retweets, and replies. They applied natural language processing methods to prepare tweets for sentiment analysis and topic modeling. Analyses showed tweets most commonly expressed sentiments of fear and surprise. The economic impact of the pandemic was the most discussed topic. ^{Reference Medford, Saleh, Sumarsono, Perl and Lehmann8} Saleh et al ^{Reference Saleh, Lehmann, McDonald, Basit and Medford14} performed a similar analysis with the goal of understanding public perception of social distancing and found most tweets expressed positive polarity. ^{Reference Saleh, Lehmann, McDonald, Basit and Medford14} Joy was the predominant emotion and was expressed in >50% of tweets. The most common tweet topics were “public opinion and values,” “media and entertainment,” and “quarantine measures and effects.” ^{Reference Saleh, Lehmann, McDonald, Basit and Medford14} Sentiment analysis complements transmission models, as these analyses give researchers insight on public perception on prevention and control measures. Medford et al ^{Reference Medford, Saleh, Sumarsono, Perl and Lehmann8} also highlighted the utility of this technology for combating misinformation on social media and the associated fear and mistrust. Individuals or autonomous “bots” on social media platforms spread malicious and dangerous misinformation. ^{Reference Abeshouse30,Reference Hern31} Natural language processing methods give platforms the ability to recognize potentially misleading posts and counter them with educational information, making them likely a worthwhile investment for organizations such as the World Health Organization (WHO) or the CDC. ^{Reference Medford, Saleh, Sumarsono, Perl and Lehmann8}

The works of Medford et al and Saleh et al are important in the context of Kermack and McKendrick’s “Susceptible, Infected, Removed or Recovered (SIR)” model, which is a well-tested disease transmission model. ^{Reference Kermack and McKendrick32} The model partitions the population into compartments based on who in the population is “susceptible” to a disease, who is “infected,” and who is “removed” from the population due to past infection resulting in either death or immunity. The model allows researchers to simulate epidemics by using differential equations to model the flow of individuals among the 3 compartments. The model’s behavior is dependent on the inherent infectivity of the disease and the population density of the susceptible population. ^{Reference Pickering, Venkatesan, Lehmann, Saleh and Medford33} This model has been well tested for various scenarios, but it assumes stochastic interaction that is not realistic in populations with a more complex social structure. For example, an infected individual with a large social network may infect more individuals than someone with a small number of contacts. In the context of COVID-19 modeling, assumptions about the proportion of people following social distancing guidelines or participating in vaccination efforts are important for determining the number of individuals susceptible. Natural language processing for analyzing social media posts can improve these assumptions based on the proportion of users posting in favor of practices such as social distancing and vaccination.

Another potential source of disease surveillance derives from crowdfunding sites. ^{Reference Saleh, Lehmann and Medford15} As the number of COVID-19 cases and its economic impact increased in the United States, reports noted an increase in web-based crowdfunding related to its costs. Saleh et al ^{Reference Saleh, Lehmann and Medford15} examined the web-based crowdfunding response in the early stages of the COVID-19 pandemic in the United States using campaigns with narratives on GoFundMe. ^{Reference Saleh, Lehmann and Medford15} A substantial increase in overall crowdfunding campaigns in March was largely attributable to COVID-19–related campaigns. However, as the COVID-19 pandemic progressed, the number of campaigns per COVID-19 case declined more than 10-fold, and there was a lack of a case-dependent response. Saleh et al ^{Reference Saleh, Lehmann and Medford15} concluded web-based crowdfunding appears to be a stopgap for only a minority of campaigners. However, crowdfunding activity may serve an early signal for emerging needs and could aid the provision governmental disaster relief.

Limitations of social media

Clinical surveillance programs for emerging infectious diseases are labor intensive, costly, and time-consuming, and they require a substantial workforce. ^{Reference McGough, Brownstein, Hawkins and Santillana3–Reference Odlum and Yoon7} Leveraging high volume, publicly available data from social media posts provide a time and cost-efficient alternative to traditional measures. Thus, utilization of social media posts for pandemic surveillance is of increasing relevance as social media use becomes increasingly ubiquitous; however, several limitations must be considered.

Underdeveloped areas, both locally and abroad, with less Internet and technology access, are also areas at high risk of being disproportionately affected by pandemics due to crowding, multigenerational households, and decreased access to information. ^{Reference Beaunoyer, Dupéré and Guitton34} Lack of access likely causes underrepresentation of these populations in disease models, further aggravating health disparities already faced. Additionally, natural language processing methods are language dependent and, although researchers can complete their analyses in multiple languages, it would be challenging and expensive to adequately include posts in less common languages or dialects, resulting in further selection bias.

Social media surveillance measures are dependent on natural language processing models, which also present some inherent limitations. Not every social media post with certain language or key words is reflective of an infected individual but may be reflective of interest in the disease or prevalence of the disease in the media. Although the natural language processing model can be optimized to extract the most relevant data, the model may not be able to consistently distinguish between infected and interested users or between serious and sarcastic posts. Post volume sufficient to generate enough signal is important to compensate for noise. Thus, this type of surveillance more difficult to execute in populations with less access to social media.

In conclusion, social media posts have the potential to provide real-time information about disease incidence in the absence of timely testing or disease reporting. Public health entities should consider leveraging these types of data for disease surveillance going forward. Currently, documented studies describe the use of social media for disease surveillance are retrospective in nature and have been carried out in response to an emerging pandemic or epidemic. ^{Reference McGough, Brownstein, Hawkins and Santillana3–Reference Saleh, Lehmann and Medford15} A major benefit of using social media posts for disease surveillance is the potential for early warning systems. Future work in the field should focus on proactive approaches for monitoring known and emerging infectious diseases. Although studies of surveillance claim that it is cost-effective in comparison to traditional surveillance methods, further research is necessary to determine the comparative cost of employing individuals properly trained in data science and natural language processing. ^{Reference Marques-Toledo, Degener and Vinhal4} Additionally, there is untapped potential in the analysis of posts on video-based social media platforms, and these methods should be explored further. Social media data are valuable to surveillance of infectious diseases and will remain a valuable resource for healthcare knowledge generation in the future.