Twitter-Based Influenza Detection After Flu Peak via Tweets With Indirect Information: Text Mining Study

Area of Science:

• Computational epidemiology
• Public health surveillance
• Social media analytics

Background:

• Social networking services (SNSs) are increasingly used for health surveillance, with users acting as social sensors.
• A key challenge in SNS-based surveillance is the inconsistency of results from small or inactive populations.
• Accurate disease outbreak prediction requires reliable data extraction from SNS platforms.

Purpose of the Study:

• To propose a novel approach for estimating patient number trends using indirect social media information.
• To enhance disease surveillance by incorporating data from both urban and rural areas.
• To improve the reliability of SNS-based health surveillance systems.

Main Methods:

• Developed a TRAP (Trend Analysis and Prediction) model integrating direct and indirect information from social media posts.
• Utilized a dataset of 7 million Japanese influenza-related tweets over three years.
• Treated indirect information as a factor that inhibits direct information to manage noise and age of data.

Main Results:

• The proposed TRAP model combined with Natural Language Processing (NLP) significantly improved estimation accuracy.
• The TRAP+NLP model achieved a correlation coefficient of .70, a substantial increase from the baseline model's .36.
• Performance improvements were observed in both urban and rural settings.

Conclusions:

• The proposed TRAP model effectively utilizes indirect social media information to enhance disease trend estimation.
• The method demonstrates improved accuracy in both urban and rural areas, validating its effectiveness for public health surveillance.
• Integrating NLP classification with the TRAP model offers a robust solution for SNS-based epidemic prediction.