Estimating influenza incidence using search query deceptiveness and generalized ridge regression

Area of Science:

• Epidemiology
• Computational Public Health
• Data Science

Background:

• Seasonal influenza causes significant mortality and morbidity annually.
• Current surveillance relies on time-consuming and costly in-person patient contact.
• Internet activity data offers a potential alternative for timely outbreak detection.

Purpose of the Study:

• To quantitatively assess the risk of deceptive internet activity traces in influenza incidence estimation.
• To evaluate methods for mitigating estimation errors caused by non-informative internet data.
• To improve the accuracy and timeliness of influenza surveillance using digital data.

Main Methods:

• Utilized linear regression to analyze simulated and real internet activity traces.
• Varied the 'deceptiveness' of data to quantify its impact on influenza incidence estimates.
• Employed a semantic distance measure from Wikipedia categories as a proxy for data deceptiveness.

Main Results:

• Knowledge of data deceptiveness significantly reduced estimation errors.
• Automatically selected features performed comparably to or better than human-curated features.
• Wikipedia category semantic distance effectively indicated data deceptiveness.

Conclusions:

• Influenza incidence models should incorporate measures of data deceptiveness alongside feature-incidence mappings.
• This approach enhances the reliability of internet data for public health surveillance.
• Improved digital surveillance can decrease costs and increase the speed of outbreak detection.