This paper is in the following e-collection/theme issue:

Early Results in Infodemiology and Infoveillance (57) Infodemiology and Infoveillance (1269) Outbreak and Pandemic Preparedness and Management (1673) Public (e)Health, Digital Epidemiology and Public Health Informatics (636)

Published on in Vol 8 (2024)

Preprints (earlier versions) of this paper are available at https://preprints.jmir.org/preprint/46087, first published .
A Novel Approach for the Early Detection of Medical Resource Demand Surges During Health Care Emergencies: Infodemiology Study of Tweets

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

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

Authors of this article:

Mahakprit Kaur1, 2 Author Orcid Image ;   Taylor Cargill1, 2 Author Orcid Image ;   Kevin Hui2, 3 Author Orcid Image ;   Minh Vu2, 3 Author Orcid Image ;   Nicola Luigi Bragazzi4, 5 Author Orcid Image ;   Jude Dzevela Kong2, 4 Author Orcid Image
Article Authors Cited by (1) Tweetations (3) Metrics

    Original Paper

    1Department of Biology, Faculty of Science, York University, Toronto, ON, Canada

    2Africa-Canada Artificial Intelligence and Data Innovation Consortium, Toronto, ON, Canada

    3Department of Computer Science, Lassonde School of Engineering, York University, Toronto, ON, Canada

    4Dalla Lana School of Public Health, University of Toronto, Toronto, ON, Canada

    5Laboratory for Industrial and Applied Mathematics, York University, Toronto, ON, Canada

    Corresponding Author:

    Jude Dzevela Kong, PhD

    Dalla Lana School of Public Health

    University of Toronto

    155 College St, Room 500

    Toronto, ON, M5T 3M7

    Canada

    Phone: 1 416 978 0901

    Email: jdkong@yorku.ca


    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.

    JMIR Form Res 2024;8:e46087

    doi:10.2196/46087

    Keywords

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


    Background

    The World Health Organization declared COVID-19 a pandemic on March 11, 2020 [1]. The emergence of this pandemic, caused by SARS-CoV-2, led to an unprecedented disruption in the global health care system that exposed and exacerbated existing vulnerabilities in health infrastructure around the world. In particular, the COVID-19 pandemic has had a profound impact on the global medical supply chain, leading to people struggling desperately to access crucial medical resources in the face of case surges and high resource demand [2].

    The unprecedented nature of the pandemic and the limited availability of resources, no matter the country, will inevitably lead to the need for prioritizing scarce medical resources to different extents [3,4]. Wealthy countries, such as the United States, experienced shortages of personal protective equipment (PPE) and ventilators [5,6]. This led to the Centers for Disease Control and Prevention developing guidelines for the optimal sourcing of COVID-19 mitigation equipment such as face masks [7]. The pandemic also resulted in an increased strain on hospital capacity around the world. This was especially true for low- and middle-income countries, where health care systems are likely to already be underresourced and stretched thin, making them particularly vulnerable to becoming overwhelmed [8,9]. Considering the potential for future pandemic scenarios and for the recurrence of existing disease outbreaks as new virus variants emerge, the development of accurate real-time methodologies for detecting and forecasting disease impacts is critical for an effective global health response [10,11]. For hospitals that experience volatile demand surges in the face of a finite medical resource supply, timely solutions are required that can allow for rapid and precise decisions to be made regarding resource allocation.

    Given that social media are an emerging source for real-time and easily accessible information, there is potential to leverage data from social media for forecasting real-world outcomes [12-15]. Social media platforms and web search data host a wealth of real-time data that broadly reflect the current state of affairs in a particular region [16]. Although the standards of validation for these new data streams are still being validated because they do not have a track record of use, these unconventional data sources have the potential to aid in short- and long-term surveillance, although the surveillance goals must be clearly defined. Studies have found that leveraging social media to identify shortages has the potential to be a cost-effective solution that can be used in real time [17,18]; for instance, Get Us PPE is a grassroots organization that leveraged Twitter (subsequently rebranded as X) to address medical supply shortages in US health care facilities during the first year of the COVID-19 pandemic [19]. Its success in garnering both public and governmental attention to the PPE shortage crisis has demonstrated that Twitter can be a useful tool for mobilizing efforts to address gaps, identifying regional PPE shortages, and informing decision-making in the health care supply chain. In other countries too, such as India, people used Twitter during the pandemic to amplify demands for medical oxygen and intensive care unit (ICU) beds during periods when health care facilities were overwhelmed by case surges [20].

    Studies examining the role of social media to glean information about the characteristics of the pandemic note that data derived from social media and search engine data were used to predict new cases in countries such as South Korea [21], the United States [22-24], China [25-27], and Iran [28]. Twitter data have been analyzed to understand the population-level spread of disease [29-31]. Furthermore, forecasting models have been created to track demand for ICU capacity planning in countries such as Chile [32,33], Brazil [34], Colombia [35], the United States [36], India [37], and China [38]. Previous studies have applied convergent cross-mapping (CCM) analysis to explore possible relationships involving antiepidemic measure–related tweets [39], the dynamics of misleading news on Twitter [40], and the identification of the global drivers of influenza [41]. However, to our knowledge, there are limited studies examining the potential of social media, particularly Twitter, to better understand hospital bed demand.

    Objectives

    This study aimed to investigate the potential for social media, a relatively novel data stream, to be leveraged as an early warning and detection system for forecasting medical resource shortages. Specifically, this study sought to determine whether the COVID-19 discourse on Twitter could be linked to real-world ICU bed demand. We applied the Granger causality test and CCM analysis to explore whether a causal relationship exists between the volume of negative sentiment COVID-19 tweets and the proportion of ICU bed occupancy in real time in the United States, Brazil, and India. If social media can be successfully leveraged to develop an effective early warning system for forecasting medical resource demand, health care workers and governments may receive real-time insights into pandemic scenarios to inform urgent resource allocation decisions and gain a head start in preparing for demand surges.


    Overview

    For our analyses, the volume of negative sentiment COVID-19 tweets was compared with ICU bed demand data for each subregion in the United States, Brazil, and India. These 3 countries were selected for this study because they have high cumulative COVID-19 death tolls [42], and they are among the top 4 nations in terms of Twitter users [43]; in addition, publicly accessible validation data on ICU bed demand are available for each country. Three main restrictions in terms of how many patients can be treated at a hospital during the pandemic are available PPE, available ICU beds, and available health care professionals per shift [44]. The number of available ICU beds was selected as the validation parameter for our model.

    Data Sets

    For tweets, we used the publicly available Twitter data set Two Billion Multilingual COVID-19 Tweets with Sentiment, Entity, Geo, and Gender Labels (TBCOV), which contains >2 billion COVID-19 multilingual tweets, including geographic location and positive, negative, and neutral sentiment labels [45]. From this data set, negative sentiment tweets were selected to capture the volume of negative Twitter discourse surrounding COVID-19 for the United States, Brazil, and India. From this TBCOV data set, for the period from February 1, 2020, to March 31, 2021, a total of 59,832,393 tweets were extracted for the United States, of which 30,742,934 (51.38%) contained negative sentiment. For Brazil, there were 5,343,723 tweets, of which 3,004,039 (56.22%) contained negative sentiment. For India, there were 9,509,766 tweets, of which 4,199,151 (44.16%) contained negative sentiment.

    Real-world hospital bed demand was defined as inpatient_beds_used_covid_coverage from the US Health Data COVID-19 Reported Patient Impact and Hospital Capacity by State Timeseries (RAW) data set [46] for the United States and as ICU beds needed from the Institute for Health Metrics and Evaluation COVID-19 Projections data set for Brazil and India [47]. The COVID-19 Projections data set contains daily information about each region’s need and capacity for hospital beds overall, including ICU beds. In India, similar to Brazil, the greatest medical supply demand during the pandemic was for oxygen cylinders and ICU beds [48]. Amid the pandemic, the insufficient oxygen-manufacturing capacity and the fragmented nature of the Indian health care system made it extremely difficult for people to obtain the supplies they needed in time [49]. Hashtags and sample tweets posted by the Indian public during the pandemic to secure oxygen cylinders and express the urgent need for ICU beds in specific regions have been documented [50,51].

    Data collection and analysis were conducted using Python (Python Software Foundation).

    Granger Causality Test Analysis

    A time series of each region was generated using the number of negative sentiment tweets per week. This time series was standardized with mean and SD calculated from historical tweet data. Another time series, the ground truth frequency of ICU bed demand, was generated from our preprocessed medical data. In general, all time-series data were binned in intervals of 1 week.

    The Granger causality test was used to determine whether past negative tweet frequency contains information that can help forecast ICU bed demand, in addition to the information contained in the past values of ICU bed demand alone [52]. In theory, this test can be applied to a stationary time series. For a nonstationary time series, first or higher difference can be used instead [53,54]. To see whether the time series could satisfy the requirement for the Granger causality test, the augmented Dickey-Fuller (ADF) test, which determines whether a time series is stationary or nonstationary, was used. In our implementation, the functions grangercausalitytests and adfuller from the statsmodels package for Python were used.

    CCM Analysis

    The CCM analysis workflow consisted of embedding, cross-mapping, and convergence analysis as well as validation and performance testing. In embedding, the negative sentiment tweets and ICU bed demand for each region were embedded into higher dimensional spaces to capture their underlying dynamics. In cross-mapping, the embedded time series were compared to identify their relationship. In convergence analysis, the results were assessed using statistical measures to determine whether there is a robust relationship between negative sentiment tweets and ICU bed demand.

    Put another way, given 2 time series X and Y, their data point entries can be considered to exist in a vector space with x and y axes, and the points over time form a trajectory in the space. Likewise, one can include the time-delayed values of X as new axes, where the vectors can be <X(t), X(t−3), X(t−6),...>, <X(t−1), X(t−4), X(t−7),...>, etc.

    If the values of X over time do indeed influence or are linked to the values of Y, then a distance-weighted k-nearest neighbor model in the X, X with delay 1, X with delay 2, etc vector space applied to the same Y (and delay axes) space can have its output converge to the actual observed values of Y, that is, predict the value of Y. If the convergence between modeled Y from X with delays and the actual observed values of Y is close, we can say that the model constructed represents the causality relation between X and Y.

    In our implementation, the causal_ccm package for Python was used.

    Ethical Considerations

    Ethics approval was not required for our study because all data and information are publicly available. In addition, all user-identifiable information was excluded from the study results.


    Overview

    From the TBCOV data set, for the period from February 1, 2020, to March 31, 2021, a total of 30,742,934 tweets containing negative sentiment were extracted for the United States; 3,004,039 tweets containing negative sentiment were extracted for Brazil; and 4,199,151 tweets containing negative sentiment were extracted for India. Our results can be categorized into (1) Granger causality test analysis and (2) CCM analysis.

    Granger Causality Test Analysis

    United States

    Figure 1 shows time series graphs comparing negative sentiment COVID-19 tweets with real-world ICU bed demand data for each of the 50 US states.

    Before performing the Granger causality test, the 2 time series were checked to determine whether they were stationary or nonstationary using ADF tests. After taking the second difference of the 2 time series, the P values of the ADF tests for negative sentiment COVID-19 tweets and ICU bed demand were found to be <.05 for all US states, meaning that we were able to reject the null hypothesis (H0) that a unit root was present in the time series samples; in other words, the 2 time series were stationary. The results are summarized in Multimedia Appendix 1.

    The results for the Granger causality test with H0, that is, negative sentiment COVID-19 tweets do not Granger-cause ICU bed demand in US states, are presented in Table 1. At lag 2, H0 was rejected for 14 (28%) of the 50 US states (P<.05).

    ‎
    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.
    Table 1. Granger causality test for all 50 US states.
    StateP valueNumber of lagsReject null hypothesis
    Alaska.3710No
    Alabama.041Yes
    Arkansas.3410No
    Arizona<.0011Yes
    California.9110No
    Colorado.6110No
    Connecticut.9910No
    Delaware.4010No
    Florida.042Yes
    Georgia.9910No
    Hawaii.9310No
    Iowa.9710No
    Idaho.4510No
    Illinois.042Yes
    Indiana.9910No
    Kansas.2510No
    Kentucky.9710No
    Louisiana.1010No
    Massachusetts.0012Yes
    Maryland.9910No
    Maine.9510No
    Michigan.023Yes
    Minnesota.5910No
    Missouri.4310No
    Mississippi.9910No
    Montana.3210No
    North Carolina.024Yes
    North Dakota.0043Yes
    Nebraska.9310No
    New Hampshire.011Yes
    New Jersey.9310No
    New Mexico.8010No
    Nevada<.0014Yes
    New York.0011Yes
    Ohio.2810No
    Oklahoma.0472Yes
    Oregon.5410No
    Pennsylvania.5310No
    Rhode Island.0021Yes
    South Carolina.7810No
    South Dakota.1910No
    Tennessee.9410No
    Texas.3210No
    Utah.9710No
    Virginia.027Yes
    Vermont.7010No
    Washington.6910No
    Wisconsin.9510No
    West Virginia.9910No
    Wyoming.2410No
    Brazil

    Figure 2 shows time series graphs comparing negative sentiment COVID-19 tweets with real-world ICU bed demand data for each of the 27 Brazilian federative units.

    Before performing the Granger causality test, the 2 time series were checked to determine whether they were stationary or nonstationary using ADF tests, and H0 was rejected for all Brazilian federative units (P<.05), qualifying all of them for the Granger causality test. The results are summarized in Multimedia Appendix 2.

    The results for the Granger causality test with H0, that is, negative sentiment COVID-19 tweets do not Granger-cause ICU bed demand in Brazilian federative units, are presented in Table 2. At lag 2, H0 was rejected for 6 (22%) of the 27 Brazilian federative units (P<.05).

    ‎
    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.
    Table 2. Granger causality test for Brazilian federative units.
    Federative unitP valueNumber of lagsReject null hypothesis
    Acre.9910No
    Alagoas.9010No
    Amazonas.2810No
    Amapá.9910No
    Bahia.9910No
    Ceará<.0019Yes
    Federal District.9910No
    Espirito Santo.5510No
    Goias.6010No
    Maranhão.046Yes
    Minas Gerais.8510No
    Mato Grosso do Sul.9910No
    Mato Grosso.9910No
    Pará.0062Yes
    Paraíba.041Yes
    Parana.6010No
    Pernambuco.012Yes
    Piaui.9010No
    Rio de Janeiro.1610No
    Rio Grande do Norte.9910No
    Rondônia.9110No
    Roraima.0024Yes
    Rio Grande do Sul.9910No
    Santa Catarina.9910No
    Sergipe.4810No
    São Paulo.9110No
    Tocantins.5010No
    India

    Figure 3 shows time series graphs comparing negative sentiment COVID-19 tweets with real-world ICU bed demand data for the 25 Indian states included in the analysis (Assam, Meghalaya, and Tamil Nadu were excluded owing to lack of data) and the national capital region of Delhi.

    Before performing the Granger causality test, the 2 time series were checked to determine whether they were stationary or nonstationary using ADF tests, and H0 was rejected for all Indian states and the national capital region of Delhi (P<.05), qualifying all for the Granger causality test. The results are summarized in Multimedia Appendix 3.

    The results for the Granger causality test with H0, that is, negative sentiment tweets do not Granger-cause ICU bed demand in Indian states and the national capital region, are presented in Table 3. At lag 2, H0 was rejected for 6 (23%) of the 26 included regions (25 Indian states and the national capital region; P<.05).

    ‎
    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.
    Table 3. Granger causality test for Indian states and the national capital region of Delhi.
    State or national capital regionP valueNumber of lagsReject null hypothesis
    Haryana.9910No
    Madhya Pradesh.6010No
    Andhra Pradesh.9910No
    Uttarakhand.9910No
    Gujarat.8610No
    Manipur.00310Yes
    Himachal Pradesh.8310No
    Punjab.9710No
    Karnataka.9910No
    Jharkhand.3210No
    Bihar.4510No
    Arunachal Pradesh.9910No
    Sikkim.9910No
    Mizoram.9910No
    Goa.8110No
    Kerala<.0018Yes
    West Bengal.2810No
    Maharashtra.0026Yes
    Tripura.6410No
    Delhi.9110No
    Uttar Pradesh.9910No
    Rajasthan.8810No
    Nagaland.9910No
    Odisha.9810No
    Telangana.9910No
    Chhattisgarh.8910No

    CCM Analysis

    United States

    Figure 4 illustrates the results of the CCM analysis for the United States, where a significant causal relationship was found between ICU bed demand and negative sentiment COVID-19 tweets for 46 (92%) of the 50 states (P<.05). The full list of correlation coefficients (r) and P values for the CCM analysis across US states is presented in Table 4.

    ‎
    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.
    Table 4. Results of the convergent cross-mapping analysis for each US state. The correlation coefficients (r) ranged from 0.0405 (Vermont) to 0.7670 (New York). Series X=negative sentiment COVID-19 tweet proportion and series Y=intensive care unit bed demand. X→Y refers to the correlation coefficient when modeling X causing Y; Y→X refers to the correlation coefficient when modeling Y causing X.
    StateCorrelation coefficient, rP valueX→YY→X
    Alaska0.1080.040.1080−0.1520
    Alabama0.2280<.0010.22800.0211
    Arkansas0.2320<.0010.2340−0.0166
    Arizona0.5740<.0010.57300.5460
    California0.6270<.0010.62600.1100
    Colorado0.5750<.0010.57600.3770
    Connecticut0.7010<.0010.70000.2770
    Delaware0.4810<.0010.48000.5290
    Florida0.7430<.0010.74300.5600
    Georgia0.4050<.0010.40400.0673
    Hawaii0.1650<.0010.16500.3960
    Iowa0.3930<.0010.38700.2240
    Idaho0.2080<.0010.2070−0.0386
    Illinois0.5700<.0010.57000.4190
    Indiana0.2150<.0010.21900.3560
    Kansas0.3710<.0010.37500.3380
    Kentucky0.2790<.0010.27900.1840
    Louisiana0.3880<.0010.38800.5810
    Massachusetts0.3590<.0010.35700.4900
    Maryland0.3980<.0010.39800.5700
    Maine0.0526.310.05180.2660
    Michigan0.5480<.0010.55000.3720
    Minnesota0.2960<.0010.29300.2920
    Missouri0.2920<.0010.28900.2170
    Mississippi0.5950<.0010.59500.2920
    Montana0.2600<.0010.26000.2020
    North Carolina0.5470<.0010.54000.5850
    North Dakota0.6000<.0010.59900.6830
    Nebraska0.1580.0020.15700.2520
    New Hampshire0.4080<.0010.40800.0395
    New Jersey0.6830<.0010.67700.4970
    New Mexico0.4810<.0010.48000.3220
    Nevada0.1800<.0010.18000.2380
    New York0.7670<.0010.76600.4700
    Ohio0.4840<.0010.48300.1210
    Oklahoma0.0565.270.05590.0657
    Oregon0.2590<.0010.25700.1350
    Pennsylvania0.6290<.0010.62800.4440
    Rhode Island0.3990<.0010.39900.3960
    South Carolina0.2550<.0010.25400.2780
    South Dakota0.4140<.0010.41300.1500
    Tennessee0.4920<.0010.49200.0250
    Texas0.4430<.0010.44300.5220
    Utah0.2970<.0010.2950−0.0438
    Virginia0.3760<.0010.37400.1550
    Vermont0.0405.440.03410.0602
    Washington0.5930<.0010.59300.2970
    Wisconsin0.6140<.0010.61400.2920
    West Virginia0.0746.140.0752−0.0510
    Wyoming0.2700<.0010.26900.3740
    Brazil

    Figure 5 illustrates the results of the CCM analysis for Brazil, where a significant causal relationship was found between ICU bed demand and negative sentiment COVID-19 tweets for 26 (96%) of the 27 Brazilian federative units (P<.05). The full list of correlation coefficients (r) and P values for the CCM analysis across Brazil is presented in Table 5.

    ‎
    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.
    Table 5. Results of the convergent cross-mapping analysis for each Brazilian federative unit. The correlation coefficients (r) ranged from 0.0751 (Amapá) to 0.8010 (Amazonas). Series X=negative sentiment COVID-19 tweet proportion and series Y=intensive care unit bed demand. X→Y refers to the correlation coefficient when modeling X causing Y; Y→X refers to the correlation coefficient when modeling Y causing X.
    Federative unitCorrelation coefficient, rP valueX→YY→X
    Acre0.5800<.0010.58000.7230
    Alagoas0.6010<.0010.60100.6250
    Amazonas0.8010<.0010.80100.8500
    Amapá0.0751.140.07580.6270
    Bahia0.5940<.0010.59400.4860
    Ceará0.3140<.0010.31700.7180
    Federal District0.2720<.0010.26000.1720
    Espirito Santo0.1940<.0010.19400.4150
    Goias0.3290<.0010.32900.2490
    Maranhão0.3970<.0010.39600.7280
    Minas Gerais0.5370<.0010.53600.3490
    Mato Grosso do Sul0.1910<.0010.19100.3350
    Mato Grosso0.3400<.0010.34000.3070
    Pará0.4750<.0010.48300.6790
    Paraíba0.3780<.0010.37800.5710
    Parana0.6110<.0010.61500.2810
    Pernambuco0.4350<.0010.43500.7090
    Piaui0.2990<.0010.29900.3390
    Rio de Janeiro0.4760<.0010.47600.5930
    Rio Grande do Norte0.2810<.0010.27800.5440
    Rondônia0.2760<.0010.27700.4130
    Roraima0.5170<.0010.51600.3440
    Rio Grande do Sul0.5040<.0010.50400.6780
    Santa Catarina0.5400<.0010.54900.2040
    Sergipe0.3230<.0010.32200.3400
    São Paulo0.6240<.0010.62500.5690
    Tocantins0.3530<.0010.35300.2430
    India

    Figure 6 illustrates the result of the CCM analysis for India, where a significant causal relationship was found between ICU bed demand and negative sentiment COVID-19 tweets for 26 (100%) of the 26 included regions (25 states and the national capital region; P<.05). The full list of correlation coefficients (r) and P values for the CCM analysis across Indian states and the national capital region is presented in Table 6.

    ‎
    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.
    Table 6. Results of the convergent cross-mapping analysis for each of the 25 Indian states included in the analysis and the national capital region. Correlation coefficients (r) ranged from 0.163 (Sikkim) to 0.806 (Delhi). Series X=negative sentiment COVID-19 tweet proportion and series Y=intensive care unit bed demand. X→Y refers to the correlation coefficient when modeling X causing Y; Y→X refers to the correlation coefficient when modeling Y causing X.
    State or national capital regionCorrelation coefficient, rP valueX→YY→X
    Haryana0.486<.0010.48600.4840
    Madhya Pradesh0.417<.0010.41700.3950
    Andhra Pradesh0.258<.0010.25800.2940
    Uttarakhand0.489<.0010.48900.5920
    Gujarat0.669<.0010.66800.6120
    Manipur0.511<.0010.51100.3720
    Himachal Pradesh0.470<.0010.47000.3000
    Punjab0.244<.0010.24400.2260
    Karnataka0.646<.0010.64800.3830
    Jharkhand0.517<.0010.51700.4900
    Bihar0.685<.0010.68500.3870
    Arunachal Pradesh0.232<.0010.23200.2700
    Sikkim0.163<.0010.16300.4870
    Mizoram0.321<.0010.31800.3080
    Goa0.493<.0010.49200.1650
    Kerala0.612<.0010.61200.3420
    West Bengal0.631<.0010.63600.4420
    Maharashtra0.739<.0010.74200.5630
    Tripura0.470<.0010.46800.0105
    Delhi0.806<.0010.80900.4340
    Uttar Pradesh0.707<.0010.70700.6370
    Rajasthan0.582<.0010.58200.5200
    Nagaland0.315<.0010.29200.1790
    Odisha0.614<.0010.61400.6770
    Telangana0.573<.0010.57300.2660
    Chhattisgarh0.390<.0010.39000.1740

    Principal Findings

    Given the need to prioritize the use of limited medical resources during a health care emergency scenario such as the COVID-19 pandemic, social media hold promise in identifying shortages and can be a cost-effective tool for the proper allocation of medical resources, particularly ICU beds, because when governments and organizations are well informed with real-time shortage data, they have the capacity to adequately address the immediate funding and supply needs of health care facilities. This strategy may help mitigate immediate risk to the public until a more systematic solution is possible.

    This study sought to determine which patterns existed between negative sentiment COVID-19 tweets and real-world ICU bed shortages during the pandemic, with the aim of leveraging social media to pinpoint regional surges in ICU bed demand in the United States, Brazil, and India. Two statistical tests were conducted to investigate this: the Granger causality test and CCM analysis.

    The Granger causality test aims to identify causalities where, in a stochastic system, 1 separable variable is useful for forecasting another. The results of the Granger causality test for this analysis (Figures 1-3 and Tables 1-3) indicate that negative sentiment COVID-19 tweets Granger-caused ICU bed shortage (P<.05) for 14 (28%) of the 50 US states, 6 (22%) of the 27 Brazilian federative unit, and 6 (23%) of the 26 Indian regions included in the analysis (25 states and the national capital region). By contrast, the CCM analysis aims to identify causalities for nonseparable variables that are linked in a dynamic system and can identify and quantify weak to moderate causalities that may be missed by the Granger causality test. For the 3 countries, nearly all subregions—46 (92%) of the 50 US states, 26 (96%) of the 27 Brazilian federative units, and 26 (100%) of the 26 Indian regions included in the analysis—had a significant (P<.05) result in the CCM analysis (Figures 4-6 and Tables 4-6), indicating a relationship between negative sentiment COVID-19 tweets and ICU bed demand.

    For the United States (Tables 1 and 4), of the 50 states, 13 (26%) had a significant result for both the Granger causality test and the CCM analysis (Alabama, Arizona, Florida, Illinois, Massachusetts, Michigan, North Carolina, North Dakota, New Hampshire, Nevada, New York, Rhode Island, and Virginia), 33 (66%) passed the CCM test but not the Granger causality test (Alaska, Arkansas, California, Colorado, Connecticut, Delaware, Georgia, Hawaii, Iowa, Idaho, Indiana, Kansas, Kentucky, Louisiana, Maryland, Minnesota, Missouri, Mississippi, Montana, Nebraska, New Jersey, New Mexico, Ohio, Oregon, Pennsylvania, South Carolina, South Dakota, Tennessee, Texas, Utah, Washington, Wisconsin, Wisconsin, and Wyoming), 1 (2%) passed the Granger causality test only (Oklahoma), and 3 (6%) passed neither the Granger causality test nor the CCM test (Maine, Vermont, and West Virginia). Considering that the majority of the US states (33/50, 66%) passed the CCM test but not the Granger causality test, it can be inferred that the causal relationship between negative sentiment COVID-19 tweets and ICU bed shortage is weak to moderate for US states because CCM analysis is better at detecting weak to moderate causalities than the Granger test. This also implies that the relationship between the 2 variables (social media sentiment and ICU bed shortage) is dynamic and influenced by a number of complex interacting factors such that CCM analysis may be the more appropriate method for detecting and modeling this relationship.

    For Brazil (Tables 2 and 5), a similar pattern occurred such that nearly all federative units (26/27, 96%) passed the CCM test, but only a few (6/27, 22%) passed the Granger causality test. Of the 27 federative units, 6 (22%) passed both the Granger causality test and the CCM test (Ceará, Maranhão, Pará, Paraíba, Pernambuco, and Roraima), 20 (74%) passed the CCM test but not the Granger causality test (Acre, Alagoas, Amazonas, Bahia, Federal District, Espirito Santo, Goias, Minas Gerais, Mato Grosso do Sul, Mato Grosso, Parana, Piaui, Rio de Janeiro, Rio Grande do Norte, Rondônia, Rio Grande do Sul, Santa Catarina, Sergipe, São Paulo, and Tocantins), none passed the Granger causality test only, and 1 (4%) state passed neither test (Amapá). This result again supports the idea of a dynamic and complex causal relationship being detected between negative sentiment COVID-19 tweets and ICU bed shortage.

    For India (Tables 3 and 6), a similar pattern emerged. Of the 26 regions included in the analysis (25 states and the national capital region), 3 (12%) passed both the Granger causality test and the CCM test (Manipur, Kerala, and Maharashtra), whereas the remaining 23 (88%) passed the CCM test only (Haryana, Madhya Pradesh, Andhra Pradesh, Uttarakhand, Gujarat, Himachal Pradesh, Punjab, Karnataka, Jharkhand, Bihar, Arunachal Pradesh, Sikkim, Mizoram, Goa, West Bengal, Tripura, Delhi, Uttar Pradesh, Rajasthan, Nagaland, Odisha, Telangana, and Chhattisgarh), meaning that all 25 states and the national capital region of Delhi passed at least the CCM test. This further demonstrates that CCM analysis is capable of successfully detecting a causal pattern between negative sentiment pandemic tweets and real-world medical resource shortage and that this can potentially be used to pinpoint specific regions that are expected to face surges in medical resource demand at a given time.

    Overall, these results suggest that a significant relationship exists between negative sentiment COVID-19 tweets and real-world ICU bed demand at subnational scales and that this relationship can be effectively detected and modeled using CCM analysis. These findings also indicate that the data contained within social media discourse regarding the COVID-19 pandemic can indeed be leveraged to identify and forecast the real-world impacts of the pandemic in the form of ICU bed demand surges. Further optimization of methods for identifying patterns between X sentiment and real-world medical emergencies can support the development of an early warning system for the robust real-time prediction of health care demand surges. Such a system may give health care workers and government decision makers a critical head start when deciding how to most effectively allocate medical resources in a crisis.

    Future Directions

    These results open up the possibility to develop tools that can forecast hospital bed demand in certain regions, although further research is required. This modeling approach can have a significant impact in the context of the health care supply chain. Forecasting ability is a potential factor affecting supply chain performance [27-29]. One study found that tweets related to food insecurity were strongly correlated with real food insufficiencies [16]; the authors noted that there is potential for tweet sentiment analysis to be used as a cost-effective early warning system that can help direct food-related interventions. Similarly, our results suggest that there is potential for negative sentiment COVID-19 posts to relate to actual medical resource shortage in regions where people use public discourse platforms such as Twitter (since rebranded as X). Although the current iteration of the methods described in this paper is only relevant to the time before a peak infection in the region, such methods have the potential to advance preparedness measures for future pandemics as they become more robust [31,32].

    Digital data sources can aid in the identification of changes in disease activity, and it is worth exploring whether they show better performance in this regard than traditional COVID-19 metrics such as confirmed cases [33]. In low- and middle-income countries, there is potential for social media to act as a cost-effective early warning system to identify priority regions for medical resource allocation in real time. User behavior data can be extracted, given the unique social network parameters of a region, including the language spoken and the preference for 1 social media platform over another [34]. Similar to how a multilingual data set was used in the analyses in this study, Lopreite et al [55] analyzed a data set of tweets posted in various European languages and found that there were early warning signals of COVID-19 outbreaks before public announcements about an outbreak were made. Social networks that are particular to a region can provide user behavior data that can inform early warning detection systems specific to that region; for example, using Baidu search data, Qin et al [56] were able to predict the number of new COVID-19 cases such as fever, coronavirus, and pneumonia. Their study, along with similar social media–based early warning detection efforts [37-40], shows potential for the creation of a more effective yet affordable model to forecast new cases.

    As noted in the previous subsection, there are complex interacting factors that may explain the results observed. Thus, further investigation is needed to unveil the dynamics underpinning the relationship between X negative sentiment pandemic tweets and medical resource demand; for instance, sociopolitical and economic challenges may have had varying influences on the X discourse by affecting public perceptions of COVID-19 management measures across countries. For Brazil (Figure 2), it can be seen that an increased proportion of negative sentiment preceded the ICU bed demand surge. Concern about preventing the proliferation of COVID-19 was among the major emergent X topics in Brazil, and politics also influenced the X landscape [57]. Brazil has a decentralized health system where the federal government finances the states and municipalities that provide health care services. Different levels of government must be able to coordinate effectively, and a failure to do so can lead to a disjunction in the care provided. A study by Neves et al [58] found that key government stakeholders underestimated the magnitude of the pandemic in the early weeks; for example, the Minas Gerais State Health Department posted messages on social media about not suspending the carnival, which is Brazil’s largest festival. When questionable COVID-19 management decisions are made that increase public health risk, such as decisions not to suspend major public events (eg, Brazil’s annual carnival), it is possible for the X discourse to reflect negative public perceptions of these government decisions before their negative impacts on case count and ICU bed demand. Further research is required to better understand the reliability of social media discourse in reflecting current pandemic management landscapes so that web-based public sentiment can accurately forecast pandemic impacts.

    Incongruencies among different levels of government may have also contributed to the results observed. The then President of Brazil, Jair Messias Bolsonaro, issued messages that conflicted with those issued by states and municipalities, such as defending hydroxychloroquine as a COVID-19 treatment and countering mask use and physical distancing [59]. At a state-specific level, the governor of Rio de Janeiro state faced charges regarding irregularities in contracts for building emergency field hospitals, which prevented efforts to relieve strain on hospitals [60]. In addition, the state government of São Paulo was being investigated for ventilator fraud, which led to a shortage of ventilators that prevented citizens from accessing life-saving equipment [61]. Overall, the lack of coordinated effort and strategy for dealing with the pandemic contributed to the inconsistent implementation of preventive measures across the states and resulted in confusion for the public. State-specific challenges may have also contributed to the rise in negative sentiment and increased medical resource demand, which can help explain the results observed.

    A similar exploration of potential factors for India can shed light on how they contributed to the nuanced and complex relationship observed between negative sentiment COVID-19 tweets and medical resource demand. India emerged as an interesting case study during the pandemic because of the novel use of X by citizens as a way of sourcing ICU beds and ventilators for their loved ones. As citizens found out that there was a shortage of equipment in the hospitals, they took to X in an effort to source medical supplies that were desperately needed. Adherence to government policy is another factor to consider in the Indian X discourse; for example, both COVID-19 waves were associated with nationwide shutdowns, but as the months wore on, people started to adhere less to measures such as masking and physical distancing. This may have contributed to an increase in cases, overwhelming hospitals [62].

    Further exploration into region-specific factors as well as social and political contexts will be important for refining our forecasting models and gaining a better understanding of the complex relationship between X negative sentiment pandemic tweets and medical resource demand.

    Limitations

    One process in our methodology involved using tweets that already had sentiment labels and analyzing them. However, we are aware that there may not be data sets containing pandemic-related tweets in future contexts. Therefore, 1 recommendation would be to first extract relevant keyword–related tweets and increase accuracy with the help of supervised natural language processing models. A similar study with the aim of predicting medical resource shortages based on tweets was conducted for the state of New York [63]. The method consisted of using supervised learning to find related tweets, which is a more robust method.

    One other limitation is that the ICU bed data sets for Brazil and India that were used to validate the model have decreased reliability because they only provide an estimate of the actual data [18], and they may not have accounted for fluctuations in ICU bed supply. However, because there were no other publicly available data sets, it was reasonable to use these data sets.

    The use of X data has some important ethical considerations that lie at the intersection of privacy and data collection; for example, social media data collection can be considered a double-edged sword. On the one hand, it provides very valuable data that can increase the possibility of creating important solutions such as a more cost-effective and faster method of gauging PPE shortages and medical resource demand. On the other hand, current artificial intelligence and data collection practices have raised concerns about privacy and the selling of personal data.

    Conclusions

    The COVID-19 pandemic has made it clear that adequate demand-based allocation of medical resources and adequate preparedness for surges in hospital admissions are paramount to reduce cases and deaths at the onset of a pandemic. Between the time that a pandemic hits and the time that a vaccine is developed and distributed, it is vital that the medical supply is carefully managed to ensure that all health care facilities have adequate capacity and proper plans for meeting unprecedented medical resource demand surges. This study analyzed negative sentiment COVID-19 tweets that were compared with real-world ICU bed use data. Our results show promise that leveraging social media, particularly X, has the potential to provide a cost-effective relatively rapid method that can inform resource allocation to facilities that need it most.

    Further investigation into the potential of X data in the modeling of medical supply shortages may lead to the development of powerful tools that can inform health care decision-making in pandemic scenarios. X causal analysis in shortage forecasting has the potential to be applied broadly in a global context for identifying medical resource demand and informing health care supply chain decisions.

    Acknowledgments

    The authors would like to thank the STEM Fellowship for the opportunity to participate in the Inter-University Big Data Challenge 2022 during which this project was developed. JDK acknowledges support from Canada’s International Development Research Centre (grant 109981) and the International Development Research Centre and the Swedish International Development Cooperation Agency (grant 109559-001). JDK equally acknowledges support from the Natural Sciences and Engineering Research Council of Canada (NSERC) Discovery Grant (grant RGPIN-2022-04559), the NSERC Discovery Launch Supplements (grant DGECR-2022-00454), and the New Frontiers in Research Fund–Exploration Stream (grant NFRFE-2021-00879).

    Data Availability

    The data sets generated and analyzed during this study are available in the Two Billion Multilingual COVID-19 Tweets with Sentiment, Entity, Geo, and Gender Labels repository [46]; the Institute for Health Metrics and EvaluationCOVID-19 Projections repository [48]; and the COVID-19 Reported Patient Impact and Hospital Capacity by State Timeseries (RAW) [47].

    Conflicts of Interest

    None declared.

    Multimedia Appendix 1

    Augmented Dickey-Fuller Test for US states.

    DOCX File , 22 KB
    Multimedia Appendix 2

    Augmented Dickey-Fuller Test for Brazil subregions.

    DOCX File , 18 KB
    Multimedia Appendix 3

    Augmented Dickey-Fuller Test for India states.

    DOCX File , 17 KB
    1. Cucinotta D, Vanelli M. WHO declares COVID-19 a pandemic. Acta Biomed. Mar 19, 2020;91(1):157-160. [FREE Full text] [CrossRef] [Medline]
    2. Zhu G, Chou MC, Tsai CW. Lessons learned from the COVID-19 pandemic exposing the shortcomings of current supply chain operations: a long-term prescriptive offering. Sustainability. Jul 21, 2020;12(14):5858. [CrossRef]
    3. Srinivas G, Maanasa R, Meenakshi M, Adaikalam JM, Seshayyan S, Muthuvel T. Ethical rationing of healthcare resources during COVID-19 outbreak: review. Ethics Med Public Health. Mar 2021;16:100633. [FREE Full text] [CrossRef] [Medline]
    4. Emanuel EJ, Persad G, Upshur R, Thome B, Parker M, Glickman A, et al. Fair allocation of scarce medical resources in the time of Covid-19. N Engl J Med. May 21, 2020;382(21):2049-2055. [CrossRef] [Medline]
    5. Ranney ML, Griffeth V, Jha AK. Critical supply shortages - the need for ventilators and personal protective equipment during the Covid-19 pandemic. N Engl J Med. Apr 30, 2020;382(18):e41. [CrossRef] [Medline]
    6. McMahon DE, Peters GA, Ivers LC, Freeman EE. Global resource shortages during COVID-19: bad news for low-income countries. PLoS Negl Trop Dis. Jul 2020;14(7):e0008412. [FREE Full text] [CrossRef] [Medline]
    7. Livingston E, Desai A, Berkwits M. Sourcing personal protective equipment during the COVID-19 pandemic. JAMA. May 19, 2020;323(19):1912-1914. [CrossRef] [Medline]
    8. Ivers LC, Walton DA. COVID-19: global health equity in pandemic response. Am J Trop Med Hyg. Jun 2020;102(6):1149-1150. [FREE Full text] [CrossRef] [Medline]
    9. Egger D, Miguel E, Warren SS, Shenoy A, Collins E, Karlan D, et al. Falling living standards during the COVID-19 crisis: quantitative evidence from nine developing countries. Sci Adv. Feb 2021;7(6):eabe0997. [FREE Full text] [CrossRef] [Medline]
    10. Dodds WK. Disease now and potential future pandemics. In: The World's Worst Problems. Cham, Switzerland. Springer International Publishing; 2019.
    11. Castillo-Chavez C, Curtiss R, Daszak P, Levin SA, Patterson-Lomba O, Perrings C, et al. Beyond Ebola: lessons to mitigate future pandemics. Lancet Global Health. Jul 2015;3(7):e354-e355. [CrossRef]
    12. Borges D, Nascimento MC. COVID-19 ICU demand forecasting: a two-stage Prophet-LSTM approach. Appl Soft Comput. Aug 2022;125:109181. [FREE Full text] [CrossRef] [Medline]
    13. Gitto S, Di Mauro C, Ancarani A, Mancuso P. Forecasting national and regional level intensive care unit bed demand during COVID-19: the case of Italy. PLoS One. Feb 25, 2021;16(2):e0247726. [FREE Full text] [CrossRef] [Medline]
    14. Nikolopoulos K, Punia S, Schäfers A, Tsinopoulos C, Vasilakis C. Forecasting and planning during a pandemic: COVID-19 growth rates, supply chain disruptions, and governmental decisions. Eur J Oper Res. Apr 01, 2021;290(1):99-115. [FREE Full text] [CrossRef] [Medline]
    15. Shinde GR, Kalamkar AB, Mahalle PN, Dey N, Chaki J, Hassanien AE. Forecasting models for coronavirus disease (COVID-19): a survey of the state-of-the-art. SN Comput Sci. Jun 11, 2020;1(4):197. [FREE Full text] [CrossRef] [Medline]
    16. Goetz SJ, Heaton C, Imran M, Pan Y, Tian Z, Schmidt C, et al. Food insufficiency and Twitter emotions during a pandemic. Appl Econ Perspect Policy (forthcoming). Apr 03, 2022:10.1002/aepp.13258. [FREE Full text] [CrossRef] [Medline]
    17. Klein AZ, Magge A, O'Connor K, Flores Amaro JI, Weissenbacher D, Gonzalez Hernandez G. Toward using Twitter for tracking COVID-19: a natural language processing pipeline and exploratory data set. J Med Internet Res. Jan 22, 2021;23(1):e25314. [FREE Full text] [CrossRef] [Medline]
    18. Althouse BM, Scarpino SV, Meyers LA, Ayers JW, Bargsten M, Baumbach J, et al. Enhancing disease surveillance with novel data streams: challenges and opportunities. EPJ Data Sci. 2015;4(1):17. [FREE Full text] [CrossRef] [Medline]
    19. Dąbrowska J, Keränen, J, Mention AL. The emergence of community-driven platforms in response to COVID-19. Res Technol Manage. Aug 19, 2021;64(5):31-38. [CrossRef]
    20. Scarr S, Sharma M, Chowdhury J, Dutta PK. #SOS: India's tweets for help. Reuters. May 7, 2021. URL: https://www.reuters.com/graphics/HEALTH-CORONAVIRUS/INDIA-TWITTER/oakpekqlrpr/ [accessed 2022-06-09]
    21. Husnayain A, Shim E, Fuad A, Su EC. Predicting new daily COVID-19 cases and deaths using search engine query data in South Korea from 2020 to 2021: infodemiology study. J Med Internet Res. Dec 22, 2021;23(12):e34178. [FREE Full text] [CrossRef] [Medline]
    22. Cousins HC, Cousins CC, Harris A, Pasquale LR. Regional infoveillance of COVID-19 case rates: analysis of search-engine query patterns. J Med Internet Res. Jul 30, 2020;22(7):e19483. [FREE Full text] [CrossRef] [Medline]
    23. Yousefinaghani S, Dara R, Mubareka S, Sharif S. Prediction of COVID-19 waves using social media and google search: a case study of the US and Canada. Front Public Health. 2021;9:656635. [FREE Full text] [CrossRef] [Medline]
    24. Mackey T, Purushothaman V, Li J, Shah N, Nali M, Bardier C, et al. Machine learning to detect self-reporting of symptoms, testing access, and recovery associated with COVID-19 on Twitter: retrospective big data infoveillance study. JMIR Public Health Surveill. Jun 08, 2020;6(2):e19509. [FREE Full text] [CrossRef] [Medline]
    25. Fang J, Zhang X, Tong Y, Xia Y, Liu H, Wu K. Baidu index and COVID-19 epidemic forecast: evidence from China. Front Public Health. 2021;9:685141. [FREE Full text] [CrossRef] [Medline]
    26. Li C, Chen LJ, Chen X, Zhang M, Pang CP, Chen H. Retrospective analysis of the possibility of predicting the COVID-19 outbreak from internet searches and social media data, China, 2020. Euro Surveill. Mar 2020;25(10):2000199. [FREE Full text] [CrossRef] [Medline]
    27. Yan Q, Shan S, Sun M, Zhao F, Yang Y, Li Y. A social media infodemic-based prediction model for the number of severe and critical COVID-19 patients in the lockdown area. Int J Environ Res Public Health. Jul 01, 2022;19(13):8109. [FREE Full text] [CrossRef] [Medline]
    28. Ayyoubzadeh SM, Ayyoubzadeh SM, Zahedi H, Ahmadi M, R Niakan Kalhori S. Predicting COVID-19 incidence through analysis of google trends data in Iran: data mining and deep learning pilot study. JMIR Public Health Surveill. Apr 14, 2020;6(2):e18828. [FREE Full text] [CrossRef] [Medline]
    29. Kogan NE, Clemente L, Liautaud P, Kaashoek J, Link NB, Nguyen AT, et al. An early warning approach to monitor COVID-19 activity with multiple digital traces in near real time. Sci Adv. Mar 2021;7(10):eabd6989. [FREE Full text] [CrossRef] [Medline]
    30. Paul MJ, Dredze M, Broniatowski D. Twitter improves influenza forecasting. PLoS Curr. Oct 28, 2014;6 [FREE Full text] [CrossRef] [Medline]
    31. Nagar R, Yuan Q, Freifeld CC, Santillana M, Nojima A, Chunara R, et al. A case study of the New York City 2012-2013 influenza season with daily geocoded Twitter data from temporal and spatiotemporal perspectives. J Med Internet Res. Oct 20, 2014;16(10):e236. [FREE Full text] [CrossRef] [Medline]
    32. Goic M, Bozanic-Leal MS, Badal M, Basso LJ. COVID-19: short-term forecast of ICU beds in times of crisis. PLoS One. Jan 13, 2021;16(1):e0245272. [FREE Full text] [CrossRef] [Medline]
    33. Quiroga BF, Vásquez C, Vicuña MI. Nonlinear time-series forecasts for decision support: short-term demand for ICU beds in Santiago, Chile, during the 2021 COVID-19 pandemic. Int Trans Oper Res. Oct 20, 2022;30(6):3400-3428. [FREE Full text] [CrossRef] [Medline]
    34. Angelo SA, Arruda EF, Goldwasser R, Lobo MS, Salles A, Lapa e Silva JR. Demand forecast and optimal planning of intensive care unit (ICU) capacity. Pesqui Oper. 2017;37(2):229-245. [CrossRef]
    35. Rivera-Rodriguez C, Urdinola BP. Predicting hospital demand during the COVID-19 outbreak in Bogotá, Colombia. Front Public Health. Nov 11, 2020;8:582706. [FREE Full text] [CrossRef] [Medline]
    36. Preiss A, Hadley E, Jones K, Stoner MC, Kery C, Baumgartner P, et al. Incorporation of near-real-time hospital occupancy data to improve hospitalization forecast accuracy during the COVID-19 pandemic. Infect Dis Model. Mar 2022;7(1):277-285. [FREE Full text] [CrossRef] [Medline]
    37. Darapaneni N, GM M, Paduri AR, Talupuri SK, Konanki V, Galande S, et al. Predicting hospital beds utilization for COVID-19 In Telangana, India. In: Proceedings of the IEEE International IOT, Electronics and Mechatronics Conference (IEMTRONICS). Presented at: IEEE International IOT, Electronics and Mechatronics Conference (IEMTRONICS); April 21-24, 2021, 2021; Toronto, ON. URL: https://ieeexplore.ieee.org/document/9422598/authors#authors [CrossRef]
    38. Zhuang Z, Cao P, Zhao S, Han L, He D, Yang L. The shortage of hospital beds for COVID-19 and non-COVID-19 patients during the lockdown of Wuhan, China. Ann Transl Med. Feb 2021;9(3):200. [FREE Full text] [CrossRef] [Medline]
    39. Zhang Q, Yi GY, Chen LP, He W. Sentiment analysis and causal learning of COVID-19 tweets prior to the rollout of vaccines. PLoS One. Feb 24, 2023;18(2):e0277878. [FREE Full text] [CrossRef] [Medline]
    40. Castioni P, Andrighetto G, Gallotti R, Polizzi E, De Domenico M. The voice of few, the opinions of many: evidence of social biases in Twitter COVID-19 fake news sharing. R Soc Open Sci. Oct 26, 2022;9(10):220716. [FREE Full text] [CrossRef] [Medline]
    41. Deyle ER, Maher MC, Hernandez RD, Basu S, Sugihara G. Global environmental drivers of influenza. Proc Natl Acad Sci U S A. Nov 15, 2016;113(46):13081-13086. [FREE Full text] [CrossRef] [Medline]
    42. Cumulative confirmed COVID-19 deaths by world region. Our World in Data. URL: https:/​/ourworldindata.​org/​grapher/​cumulative-covid-deaths-region?tab=table&time=earliest.​.2023-01-02&facet=entity [accessed 2024-01-02]
    43. Countries with the most X/Twitter users 2023. Statista. URL: https://www.statista.com/statistics/242606/number-of-active-twitter-users-in-selected-countries/ [accessed 2024-01-02]
    44. Gonzatto Junior OA, Nascimento DC, Russo CM, Henriques MJ, Tomazella CP, Santos MO, et al. Safety-stock: predicting the demand for supplies in Brazilian hospitals during the COVID-19 pandemic. Knowl Based Syst. Jul 08, 2022;247:108753. [FREE Full text] [CrossRef] [Medline]
    45. Imran M, Qazi U, Ofli F. TBCOV: two billion multilingual COVID-19 tweets with sentiment, entity, geo, and gender labels. Data. Jan 10, 2022;7(1):8. [CrossRef]
    46. COVID-19 reported patient impact and hospital capacity by state timeseries (RAW). HealthData.gov. URL: https://healthdata.gov/Hospital/COVID-19-Reported-Patient-Impact-and-Hospital-Capa/g62h-syeh [accessed 2022-07-04]
    47. COVID-19 projections. Institute for Health Metrics and Evaluation. URL: https://covid19.healthdata.org/projections [accessed 2022-07-04]
    48. Shivji S. In India, people are begging for hospital beds and oxygen cylinders on social media. CBC News. Apr 29, 2021. URL: https://www.cbc.ca/news/world/india-covid19-pandemic-social-media-1.6005433 [accessed 2024-01-02]
    49. Natarajan KV, Prasad S. A supply chain perspective on India's COVID-19 crisis: lessons from the second wave and preparing for a potential third wave. J Glob Health. Oct 23, 2021;11:03116. [FREE Full text] [CrossRef] [Medline]
    50. Jena A. COVID-19 and SOS tweets in India. Lancet Infect Dis. Aug 2021;21(8):1072-1073. [CrossRef]
    51. Syed Abdul S, Ramaswamy M, Fernandez-Luque L, John O, Pitti T, Parashar B. The pandemic, infodemic, and people's resilience in India: viewpoint. JMIR Public Health Surveill. Dec 08, 2021;7(12):e31645. [FREE Full text] [CrossRef] [Medline]
    52. Seth A. Granger causality. Scholarpedia. 2007;2(7):1667. [CrossRef]
    53. Weisstein EW. Finite difference. Wolfram MathWorld. URL: https://mathworld.wolfram.com/FiniteDifference.html [accessed 2024-01-02]
    54. Sugihara G, May R, Ye H, Hsieh CH, Deyle E, Fogarty M, et al. Detecting causality in complex ecosystems. Science. Oct 26, 2012;338(6106):496-500. [CrossRef] [Medline]
    55. Lopreite M, Panzarasa P, Puliga M, Riccaboni M. Early warnings of COVID-19 outbreaks across Europe from social media. Sci Rep. Jan 25, 2021;11(1):2147. [FREE Full text] [CrossRef] [Medline]
    56. Qin L, Sun Q, Wang Y, Wu KF, Chen M, Shia BC, et al. Prediction of number of cases of 2019 novel coronavirus (COVID-19) using social media search index. Int J Environ Res Public Health. Mar 31, 2020;17(7):2365. [FREE Full text] [CrossRef] [Medline]
    57. Garcia K, Berton L. Topic detection and sentiment analysis in Twitter content related to COVID-19 from Brazil and the USA. Appl Soft Comput. Mar 2021;101:107057. [FREE Full text] [CrossRef] [Medline]
    58. Neves JC, de França TC, Bastos MP, de Carvalho PV, Gomes JO. Analysis of government agencies and stakeholders’ Twitter communications during the first surge of COVID-19 in Brazil. Work. Oct 19, 2022;73(s1):S81-S93. [CrossRef]
    59. Brum P, Cândido Teixeira M, Vimieiro R, Araújo E, Meira WJ, Lobo Pappa G. Political polarization on Twitter during the COVID-19 pandemic: a case study in Brazil. Soc Netw Anal Min. Sep 23, 2022;12(1):140. [FREE Full text] [CrossRef] [Medline]
    60. Urban M, Saad-Diniz E. Why Brazil’s COVID-19 response is failing. The Regulatory Review. Jun 22, 2020. URL: https://www.theregreview.org/2020/06/22/urban-saad-diniz-brazil-covid-19-response-failing/ [accessed 2024-01-02]
    61. Scalzaretto N. Police investigate ventilator fraud in São Paulo. The Brazilian Report. Jun 12, 2020. URL: https:/​/brazilian.​report/​liveblog/​coronavirus/​2020/​06/​12/​police-investigate-ventilators-deal-fraud-in-sao-paulo/​ [accessed 2024-01-02]
    62. Lande J, Pillay A, Chandra R. Deep learning for COVID-19 topic modelling via Twitter: alpha, delta and omicron. PLoS One. Aug 1, 2023;18(8):e0288681. [FREE Full text] [CrossRef] [Medline]
    63. Adiga PS. Predicting the personal protective equipment (PPE) demand and shortages during the COVID-19 pandemic. State University of New York. 2021. URL: https:/​/www.​proquest.com/​openview/​6fcbca65a527b025b4d78670c3306980/​1?pq-origsite=gscholar&cbl=18750&diss=y [accessed 2024-01-15]

    ‎
    ADF: augmented Dickey-Fuller
    CCM: convergent cross-mapping
    ICU: intensive care unit
    PPE: personal protective equipment
    TBCOV: Two Billion Multilingual COVID-19 Tweets with Sentiment, Entity, Geo, and Gender Labels

    Edited by A Mavragani; submitted 30.01.23; peer-reviewed by W Ceron, H Kim, Anonymous; comments to author 15.04.23; revised version received 07.07.23; accepted 20.11.23; published 29.01.24.

    Copyright

    ©Mahakprit Kaur, Taylor Cargill, Kevin Hui, Minh Vu, Nicola Luigi Bragazzi, Jude Dzevela Kong. Originally published in JMIR Formative Research (https://formative.jmir.org), 29.01.2024.

    This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Formative Research, is properly cited. The complete bibliographic information, a link to the original publication on https://formative.jmir.org, as well as this copyright and license information must be included.