By András Bota
Pandemics, epidemic outbreaks, and emerging diseases pose a constant risk to society. Technological responses to the greatest public health emergencies of the 20th and early 21st centuries—including HIV, polio, influenza, and the COVID-19 pandemic—demonstrated that innovation can help to mitigate the damage of these outbreaks. Nevertheless, it cannot prevent the reoccurrence of deadly diseases.
The past decades have seen unprecedented technological breakthroughs. For one, advances in long-distance communication now facilitate the rapid exchange of scientific data across the globe. This data exchange allows scientists to model a variety of global processes, such as epidemic outbreaks. Concurrent advances in data science and artificial intelligence (AI) have enabled the processing and analysis of massive amounts of data — including data about epidemics and their effects on society and the economy. Here, we examine some interesting results that have emerged in the overlap between AI and epidemiology.
Outbreak Forecasting
Effective and timely responses to emerging and ongoing disease outbreaks can reduce the number of infected individuals. Response strategies are highly dependent upon accurate forecasts of epidemic spreading patterns in both spatial and temporal dimensions. However, it is often difficult to forecast highly stochastic processes—such as epidemic spread—with traditional statistical methods.
Artificial neural networks, which are inspired by a simplified representation of neuron organization in the brain, are the driving force behind recent developments in AI and machine learning (ML). In particular, researchers can use long short-term memory (LSTM) networks to learn long-term dependencies — especially in time series forecasting. As with other neural networks, scientists train LSTM models on one part of the input dataset and then apply the trained model to a test dataset to ultimately make predictions. LSTM models and other neural networks also share the same drawback in that they are data intensive, especially for complex tasks like the forecasting of epidemic spreading.
Scientists sometimes utilize LSTM networks to forecast “epicurves” — the confirmed or estimated number of infected people, hospitalizations, or fatalities in a region. In 2022, a group of researchers trained an LSTM neural network to provide one-week-ahead predictions for cumulative COVID-19 case counts in Brazil, India, and Russia [4]. They used data from Johns Hopkins University’s COVID-19 Dashboard to train their model, which provided accurate predictions during the given time period. However, the authors acknowledged that forecasting accuracy depends on the quality of available data, which differs between countries.
Figure 1. Most likely transmission routes of COVID-19 cases originating in Beijing, China, as predicted in February 2020. Figure courtesy of Dirk Brockmann and [1].
Public Attitudes Towards COVID-19 and Vaccines
During a serious health emergency like the COVID-19 pandemic, public authorities implement intervention strategies to limit the spread of disease. These efforts can be nonpharmaceutical in nature—such as mandatory masking, the closure of public venues, or travel restrictions—or pharmaceutical, such as vaccinations. The effectiveness of any intervention depends significantly on the population’s response, but attitudes towards these policies are not universally positive. Many people cooperate and follow mitigation measures, but some may opportunistically bypass them (e.g., by not wearing masks properly) and others may actively resist (e.g., by protesting lockdowns). To further complicate matters, intervention strategies and overall communication about the disease in question may vary greatly between cities, states, regions, and countries. Accurate measurements of public viewpoints in response to these strategies can help public authorities encourage adherence and combat misinformation.
A key advantage of ML is its ability to process large amounts of complex data. Researchers can use natural language processing methods, such as topic modeling and sentiment analysis, to digest textual information (in the form of newspaper articles, blog or social media posts, comments, and so forth) and gauge public attitudes towards interventions.
Topic models are statistical probability models that extract implicit or hidden topics—sets of words that are grouped together according to a common theme—from text. Scientists commonly utilize this type of modeling to explore the underlying thematic structures of an input text. A 2022 study used topic modeling to analyze public discussion about COVID-19 in 10 newspapers from six countries from January 2020 to May 2021 [3]. The results suggest that a newspaper’s political alignment and the country’s economic condition both influence the spread of information. Newspaper coverage in less-developed economies focused more on the economic costs of interventions, while coverage in most developed countries favored vaccinations, lockdowns, protests, and the pandemic’s overall effect on society.
Another relevant ML method is sentiment analysis, which—as the name suggests—identifies the sentiments of texts. These feelings can be as simple as “positive,” “negative,” or “neutral,” but sentiment analysis also involves subfields like sarcasm, subjectivity, or the detection of hate speech. In 2021, a group of researchers analyzed social media posts on Facebook, Twitter, and Reddit to investigate public perceptions of three nonpharmaceutical interventions in 12 U.S. states [2]. They applied sentiment analysis to these posts and concluded that users were more positive about mask policies than shelter-in-place mandates and delayed school reopenings.
Societal Effects of the Pandemic
The COVID-19 pandemic’s societal ramifications extended well beyond medical aspects. In the early stages of the outbreak, vaccines were not yet available and authorities in many countries instead introduced nonpharmaceutical interventions, which in turn negatively impacted the economy. Travel restrictions had devastating effects on both domestic and international travel; employees worked from home whenever possible; and schools, gyms, restaurants, and other gathering places shuttered their doors. When vaccines finally became available, vaccine hesitancy affected their uptake. Quantitative assessments of people’s attitudes towards these repercussions can provide valuable feedback for decision-makers as they encourage adherence and combat misinformation.
As we have seen, ML tools from natural language processing—such as topic modeling and sentiment analysis—can analyze vast amounts of textual information. A 2021 study applied these methods to more than one million tweets from March 30 to July 5, 2020, in order to understand attitudes towards enforced remote work [5]. The authors reported mildly positive sentiments that acknowledged the positive and negative aspects of working from home. Perceived benefits included an increase in productivity, flexibility, and the creative use of technology. However, many people considered online conferences to be tiring and missed the social connections of their jobs. Impressions of beneficial changes in work-life balance were mixed with complaints about simultaneously managing family and work duties, especially for parents with children at home.
In May 2023, the World Health Organization ceased classifying the COVID-19 pandemic as a public health emergency of international concern. However, this does not mean that COVID-19 is no longer dangerous or that future outbreaks will not pose a significant risk to society. The use of AI techniques over the last few years has enabled fundamental breakthroughs in multiple scientific fields. AI clearly has strong applications in epidemiology, but there remain plenty of opportunities for further contributions. Continued research in this direction will make society more resilient to future pandemics and public health crises.
András Bota delivered a minisymposium presentation on this research at the 2023 SIAM Conference on Computational Science and Engineering, which took place in Amsterdam, the Netherlands, last year.
References
[1] Cohen, J. (2020, February 7). Scientists are racing to model the next moves of a coronavirus that's still hard to predict. Science. Retrieved from https://www.science.org/content/article/scientists-are-racing-model-next-moves-coronavirus-thats-still-hard-predict.
[2] Figueira, O., Hatori, Y., Liang, L., Chye, C., & Liu, Y. (2021). Understanding COVID-19 public sentiment towards public health policies using social media data. In 2021 IEEE global humanitarian technology conference (GHTC) (pp. 8-15). Seattle, WA: Institute of Electrical and Electronics Engineers.
[3] Sittar, A., Major, D., Mello, C., Mladenić, D., & Grobelnik, M. (2022). Political and economic patterns in COVID-19 news: From lockdown to vaccination. IEEE Access, 10, 40036-40050.
[4] Xu, L., Magar, R., & Farimani, A.B. (2022). Forecasting COVID-19 new cases using deep learning methods. Comput. Biol. Med., 144, 105342.
[5] Zhang, C., Yu, M.C., & Marin, S. (2021). Exploring public sentiment on enforced remote work during COVID-19. J. Appl. Psychol., 106(6), 797-810.
András Bota is a postdoctoral researcher at the Luleå University of Technology. He began working with networks during his Ph.D. studies, focusing first on community detection and then on modeling the spread of information, economic events, and diseases. After completing his doctorate, Bota spent two years at the University of New South Wales and another year at Umeå University before moving to Luleå.