JMIR Publications

ABSTRACT

Background:

The COVID-19 pandemic has exposed major flaws in the PPE supply chain and healthcare facilities are at high risk of becoming overwhelmed beyond their capacity. Therefore, it is paramount to develop new technologies that can be leveraged as an early warning and detection system.

Objective:

In this study, Twitter sentiment analysis was utilized for two objectives. One was to determine if an increase in tweets pertaining to PPE shortages correspond to an actual increase in PPE shortages in long-term care facilities in California. The second objective was to apply the same methodology toward medical supply shortage prediction in the context of developing countries; namely, to determine if negative sentiment tweets pertaining to COVID-19 in Brazil and India correspond to a greater use of hospital Intensive Care Unit (ICU) beds in these countries.

Methods:

Two time series were created including frequency of negative sentiment tweets (extracted using VADER) and ground truth frequency of medical resource shortages or demands. Then, the Granger causality test was used to determine if the time series for tweets can be useful for forecasting medical resource shortages.

Results:

The sample size for the negative sentiment tweet analysis was 6,970 tweets pertaining to California, 24,1105 tweets for the analysis of ICU bed demand in Brazil, and 8,613,049 tweets for India. For California, the results of the Granger test were significant at lag 2 (P = 0.035) and lag 5 (P = 0.005). For Brazil, the Granger test was significant for six of the 25 regions which passed the Augmented-Dickey Fuller test: Amazonas (P = 0.039, lag 4), Bahia (P = 0.019, lag 1), Federal District (P = 0.010, lag 1), Espírito Santo (P = 0.006, lag 3), Roraima (P = 0.030, lag 1), and São Paulo (P = 0.013, lag 1). For India, the results of the Granger test were significant for ten of the 27 regions which passed the Augmented-Dickey Fuller test: Tripura (P = 0.020, lag 1), Gujarat (P = 0.023 , lag 3), West Bengal (P = 0.008, lag 2), Haryana (P = 0.045, lag 1), Bihar (P = 0.126, lag 1), Karnataka (P = 0.037, lag 3), Odisha (P = 0.048, lag 4), Andhra Pradesh (P = 0.037, lag 4), Jharkhand (P = 0.0004, lag 1), and Himachal Pradesh (P = 0.020, lag 3).

Conclusions:

This study provides a novel approach for identifying regions of PPE shortage and high hospital bed demand by analyzing Twitter sentiment data given that Twitter can be used as a useful tool for rapidly organizing relief efforts. Natural language processing-driven Tweet extraction systems have the potential to be an effective method that allows for early detection of medical resource demand surges.