Detecting the period of a disease is of great importance to building information management capacity in disease control and prevention. This paper aims to optimize the disease surveillance process by further identifying the infectious or recovered period of flu cases through social media. Specifically, this paper explores the potential of using public sentiment to detect flu periods at word level. At text level, we constructed a deep learning method to classify the flu period and improve the classification result with sentiment polarity. Three important findings are revealed. Firstly, bloggers in different periods express significantly different sentiments. Blogger sentiments in the recovered period are more positive than in the infectious period when measured by the interclass distance. Secondly, the optimized disease detection process can substantially improve the classification accuracy of flu periods from 0.876 to 0.926. Thirdly, our experimental results confirm that sentiment classification plays a crucial role in accuracy improvement. Precise identification of disease periods enhances the channels for the disease surveillance processes. Therefore, a disease outbreak can be predicted credibly when a larger population is monitored. The research method proposed in our work also provides decision making reference for proactive and effective epidemic control and prevention in real time.
Keywords: disease detection; flu; sentiment analysis; social media; text classification.