A Novel Approach for the Early Detection of Medical Resource Demand Surges During Health Care Emergencies: Infodemiology Study of Tweets

Abstract

Background: The COVID-19 pandemic has highlighted gaps in the current handling of medical resource demand surges and the need for prioritizing scarce medical resources to mitigate the risk of health care facilities becoming overwhelmed.

Objective: During a health care emergency, such as the COVID-19 pandemic, the public often uses social media to express negative sentiment (eg, urgency, fear, and frustration) as a real-time response to the evolving crisis. The sentiment expressed in COVID-19 posts may provide valuable real-time information about the relative severity of medical resource demand in different regions of a country. In this study, Twitter (subsequently rebranded as X) sentiment analysis was used to investigate whether an increase in negative sentiment COVID-19 tweets corresponded to a greater demand for hospital intensive care unit (ICU) beds in specific regions of the United States, Brazil, and India.

Methods: Tweets were collected from a publicly available data set containing COVID-19 tweets with sentiment labels and geolocation information posted between February 1, 2020, and March 31, 2021. Regional medical resource shortage data were gathered from publicly available data sets reporting a time series of ICU bed demand across each country. Negative sentiment tweets were analyzed using the Granger causality test and convergent cross-mapping (CCM) analysis to assess the utility of the time series of negative sentiment tweets in forecasting ICU bed shortages.

Results: For the United States (30,742,934 negative sentiment tweets), the results of the Granger causality test (for whether negative sentiment COVID-19 tweets forecast ICU bed shortage, assuming a stochastic system) were significant (P<.05) for 14 (28%) of the 50 states that passed the augmented Dickey-Fuller test at lag 2, and the results of the CCM analysis (for whether negative sentiment COVID-19 tweets forecast ICU bed shortage, assuming a dynamic system) were significant (P<.05) for 46 (92%) of the 50 states. For Brazil (3,004,039 negative sentiment tweets), the results of the Granger causality test were significant (P<.05) for 6 (22%) of the 27 federative units, and the results of the CCM analysis were significant (P<.05) for 26 (96%) of the 27 federative units. For India (4,199,151 negative sentiment tweets), the results of the Granger causality test were significant (P<.05) for 6 (23%) of the 26 included regions (25 states and the national capital region of Delhi), and the results of the CCM analysis were significant (P<.05) for 26 (100%) of the 26 included regions.

Conclusions: This study provides a novel approach for identifying the regions of high hospital bed demand during a health care emergency scenario by analyzing Twitter sentiment data. Leveraging analyses that take advantage of natural language processing-driven tweet extraction systems has the potential to be an effective method for the early detection of medical resource demand surges.

Keywords: COVID-19; Granger; ICU bed; Twitter; causal analysis; convergent cross-mapping; global health; intensive care unit bed; medical supply shortage; pandemic; social media.

Figure 1

Time series with a comparison of trends for intensive care unit (ICU) bed use with trends for the volume of negative sentiment COVID-19 tweets across all 50 US states. TBCOV: Two Billion Multilingual COVID-19 Tweets with Sentiment, Entity, Geo, and Gender Labels.

Figure 2

Time series with a comparison of trends for intensive care unit (ICU) bed use with trends for the volume of negative sentiment COVID-19 tweets across Brazil’s 27 subdivisions (states and administrative divisions). TBCOV: Two Billion Multilingual COVID-19 Tweets with Sentiment, Entity, Geo, and Gender Labels.

Figure 3

Time series with a comparison of trends for intensive care unit (ICU) bed use with trends for the volume of negative sentiment COVID-19 tweets across Indian states. TBCOV: Two Billion Multilingual COVID-19 Tweets with Sentiment, Entity, Geo, and Gender Labels.

Figure 4

Graphs showing the convergence of the correlation coefficient (r) in convergent cross-mapping analysis as the time series length (L) approaches the maximum possible value for each US state. Series X=negative sentiment COVID-19 tweet proportion and series Y=intensive care unit bed demand. The blue graph represents the correlation coefficient when modeling X causing Y (X→Y). The orange graph represents the correlation coefficient when modeling Y causing X (Y→X). The correlation coefficients ranged from 0.0405 (Vermont) to 0.7670 (New York). The P values ranged from .44 (Vermont) to <.001 (New York). Of the 50 US states, 46 (92%) had P values <.05.

Figure 5

Graphs showing the convergence of the correlation coefficient (r) in convergent cross-mapping analysis as the time series length (L) approaches the maximum possible value for each Brazil subregion. Series X=negative sentiment COVID-19 tweet proportion and series Y= intensive care unit bed demand. The blue graph represents the correlation coefficient when modeling X causing Y (X→Y). The orange graph represents the correlation coefficient when modeling Y causing X (Y→X). The correlation coefficients ranged from 0.0751 (Amapá) to 0.8010 (Amazonas). The P values ranged from .14 (Amapá) to <.001 (Amazonas). Of the 27 Brazilian federative units, 26 (96%) had P values <.05.

Figure 6

Graphs showing the convergence of the correlation coefficient (r) in convergent cross-mapping analysis as the time series length (L) approaches the maximum possible value for each of the 26 Indian states included in the analysis. Series X=negative sentiment COVID-19 tweet proportion and series Y=intensive care unit bed demand. The blue graph represents the correlation coefficient when modeling X causing Y (X→Y). The orange graph represents the correlation coefficient when modeling Y causing X (Y→X). The correlation coefficients ranged from 0.1630 (Sikkim) to 0.8060 (Delhi). The P values ranged from .001 (Sikkim) to <.001 (Delhi). All 25 states and the national capital region of Delhi had P values <.05.