JMIR Publications

ABSTRACT

Background:

Given the ongoing coronavirus disease 2019 (COVID-19) pandemic situation, accurate predictions could greatly help in the health resource management for future waves. However, as a new entity, COVID-19’s disease dynamics seemed difficult to predict. External factors, such as internet search data, need to be included in the models to increase the accuracy of these models. However, it remains unclear whether incorporating online search volumes into models leads to better predictive performances for a long-term prediction.

Objective:

This study aimed to analyze whether search engine query data are important variables that should be included in the models predicting short- and long-term periods of new daily COVID-19 cases and deaths.

Methods:

We used country-level case-related data, NAVER search volumes, and mobility data obtained from Google and Apple for the period of January 20, 2020 to July 31, 2021 in South Korea. Data were aggregated into four subsets (3, 6, 12, and 18 months). The first 80% of the data in all subsets were used as the training set and remaining data served as the test set. Generalized linear models (GLMs) with normal, Poisson, and negative binomial distribution were developed along with linear regression (LR) models with lasso, adaptive lasso, and elastic net regularization. Value of the root mean squared error (RMSE) were defined as a loss function and were used to assess the performance of the models. All analyses and visualizations were conducted in SAS Studio, which is part of the SAS OnDemand for Academics.

Results:

GLMs with different types of distribution functions may have been beneficial in predicting new daily COVID-19 cases and deaths in the early stages of the outbreak. Non-normal distributions of cases and deaths were better predicted using the Poisson or negative binomial function. Over longer periods, as the distribution of cases and deaths became more normally distributed, LR models with regularization may have outperform the GLMs. This study also found that better performances of the models were achieved in predicting new daily deaths compared to new daily cases. In addition, an evaluation of effect of features in the models showed that NAVER search volumes were useful variables in predicting new daily COVID-19 cases, particularly in the first six months of the outbreak. Searches related to logistical needs, particularly for “thermometer” and “mask strap” showed higher feature effects in that period. For longer prediction periods, NAVER search volumes were still found to be an important variable, although with a lower feature effect. This finding suggests that term utilization should be considered to maintain the predictive performance.

Conclusions:

NAVER search volumes were important variables in the short- and long-term prediction with higher feature effects for predicting new daily COVID-19 cases in the first six months of the outbreak. Similar results were also found for death predictions.