JMIR Publications

ABSTRACT

Background:

Numerous studies have demonstrated the remarkable impact of disasters on mental health, with a substantial proportion of affected individuals developing psychiatric disorders. The timely identification of disaster-related mental health problems is essential to prevent long-term negative consequences and improve individual and community resilience. The inherent constraints observed in previous research that exclusively examined the effects of isolated incidents were removed by analyzing the impact of a disaster arising from annual events.

Objective:

We conducted a sentiment analysis to investigate the usefulness of big data obtained during a recurrent Halloween event in Itaewon, South Korea, which tragically ended in a crowd crush incident in 2022. It was hypothesized that this disaster would considerably influence the distribution of sentiments in Korean Twitter data.

Methods:

We collected tweets two weeks before and after the annual festival from 2020 to 2022 to consider the variability over the years and on the days before the disaster. The collected tweets were then subjected to sentiment analysis using a pre-trained RoBERTa neural network model that had been fine-tuned using public sentiment datasets. We analyzed tweet numbers and proportional distribution across seven pre-defined emotional categories: anger, neutrality, sadness, happiness, disgust, fear, and surprise. The resulting sentiment data were transformed into daily time-series data for an in-depth analysis. We considered each sentiment category’s numerical counts and proportions to eliminate bias because of fluctuating tweet counts. An interrupted time series analysis was conducted using Seasonal Autoregressive Integrated Moving Average with Exogenous Regressor (SARIMAX) models.

Results:

The number of tweets across all seven sentiment categories increased in volume, with an increase in tweets indicating “Sadness” in the year of the disaster (2022) compared with those in previous years. Furthermore, the proportion of tweets indicating “Sadness” and “Fear” extensively increased during the post-disaster period. The SARIMAX model exhibited the same trend. Specifically, all sentiments showed a notable increase in tweets, including a surprising increase in those categorized as happy. Regarding proportions, significant changes were observed only in tweets categorized as “Sadness” [0.046 (95% CI: 0.024–0.068, P<.0001)] and “Fear” [0.033 (95% CI: 0.014–0.051, P<.0001)].

Conclusions:

Our approach offers evidence that using sentiment data from social media combined with sentiment classification is feasible for assessing public mental health features distinctive to each disaster.