A Social Media Time-Series Data Analytics Approach for Digital Epidemiology

Abstract

Various events and their perspectives around the world are discussed or posted at every moment on social media platforms like Twitter in near real-time, forming an enriched repository of information as historical records or time series. These include people's sentiments, emotions, opinions, and other information such as situational aspects of the spreading of a particular disease, ailment, or a population explosion of some vectors or pathogens. Exploring and harnessing such information about a disease for surveillance to prevent and control its spreading or becoming epidemic or pandemic is worthwhile for a country or the world. In this paper, we correlate tweeting activity with the reported disease cases, and take advantage of the predictive power of neural networks and auto-regressive models to estimate disease incidences for the current week (aka nowcasting) considering the social media data and the disease case counts reported by the Government agencies. We propose Long Short-Term Memory (LSTM) network models and autoregressive moving average models with two channels of inputs to incorporate social media and historic disease case count data for predicting current disease case counts. We employ various LSTM network models and autoregressive moving average models to estimate the current week's disease case count and compared their performance considering tweets as exogenous input to these models. The experimental results establish the efficacy of the LSTM network models with dynamically merged inputs for predicting disease case count with least prediction error.