Social Media as an Emotional Barometer: Bidirectional Encoder Representations From Transformers-Long Short-Term Memory Sentiment Analysis on the Evolution of Public Sentiments During Influenza A on Sina Weibo

Background: Starting in October 2023, China experienced successive outbreaks and the spread of influenza A. During this period, Sina Weibo users sought emotional stability and psychological resilience by sharing information and expressing personal opinions. The content generated by users, including text posts, can be analyzed to reveal fluctuations in their emotions and psychological dynamics, thereby providing a valuable reference for assessing their mental health status. Objective: This study aimed to understand the evolution of emotions expressed on social media during the various phases of the influenza A outbreak. Methods: We used the bidirectional encoder representations from transformers-long short-term memory model to classify emotions in relevant posts from September 2023 to April 2024 and to correlate these emotions with objective influenza infection rates. Results: The positivity rate of influenza A first showed an upward trend and reached its peak between November and February of the following year. During this period, the predominant emotional response of the public was sadness. Even after the influenza positivity rate declined, sadness persisted for a while, highlighting the long-term emotional consequences of influenza on individual psychological well-being. In contrast, the emotion of surprise fluctuated very little throughout the observation period, indicating that the public’s unexpected emotions toward the influenza outbreak were not significant. As the influenza season progressed, the public’s emotional responses changed. In the early stages of the influenza outbreak, neutral emotions were weakened due to the dominance of negative emotions, such as sadness, fear, and anger. However, neutral emotions later rebounded and stabilized at a higher level, indicating that the public regained rationality and emotional balance during the peak of the influenza outbreak. In addition, happiness decreased in the early stages of the influenza outbreak due to the overshadowing of negative emotions but gradually increased as the holiday season approached. Overall, the emotional landscape shifted from being dominated by negative emotions in the early stages to a coexistence of positive, negative, and neutral emotions. This evolution of emotional dynamics is closely related to the adaptability of the public’s psychology, the effectiveness of government control measures and information dissemination, and external factors such as holidays and large-scale population movements. The model ultimately achieved a weighted average F₁-score of 0.8918 and a weighted average accuracy of 0.8980. Specifically, the accuracy was 0.8384 for sadness, 0.9537 for neutral sentiment, 0.9559 for happiness, and 0.7500 for surprise. Conclusions: The phenomenon of social sharing of emotions provides valuable theoretical insights into the collective expression of emotions on social media and the reciprocal influences among individuals. The findings not only offer a novel perspective on the mechanisms of emotional transmission during public health events but also supply empirical evidence to inform public opinion and emotional management in the context of influenza A.