Interpreting County Level COVID-19 Infection and Feature Sensitivity using Deep Learning Time Series Models

Introduction

This work combines sensitivity analysis with heterogeneous time-series deep learning model prediction, which corresponds to the interpretations of Spatio-temporal features from what the model has actually learned. We forecast county-level COVID-19 infection using the Temporal Fusion Transformer (TFT). We then use the sensitivity analysis extending Morris Method to see how sensitive the outputs are with respect to perturbation to our static and dynamic input features. We have collected 2.5 years of socioeconomic and health features over 3142 US counties. Using the proposed framework, we conduct extensive experiments and show our model can learn complex interactions and perform predictions for daily infection at the county level.

Folder Structure

• dataset_raw: Contains the collected raw dataset and the supporting files. To update use the Update dataset notebook.
• TFT-PyTorch: Contains all codes and merged feature files used during the TFT experimentation setup and interpretation. For more details, check the README.md file inside it. The primary results are highlighted in results.md.

Features

Most of the features are results from the COVID-19 Pandemic Vulnerability Index (PVI) Dashboard maintained by National Institute of Environmental Health Sciences. They have two different versions of the dashboard model (11.2 and 12.4). Since model 12.4 only has data since 2021, we have used model 11.2. These are the features currently being used in the current model. Note that, both dynamic and known futures are used as past inputs by TFT.

Feature	Type
Age Distribution	Static
Health Disparities	Static
Disease Spread	Dynamic
Social Distancing	Dynamic
Transmissible Cases	Dynamic
Vaccination Full Dose	Dynamic
Linear Space	Known Future
SinWeekly	Known Future
CosWeekly	Known Future

Details of Features from PVI Model (11.2)

Data Domain Component(s)	Update Freq.		Description/Rationale	Source(s)
Age Distribution
% age 65 and over	Static		Aged 65 or Older from 2014-2018 ACS. Older ages have been associated with more severe outcomes from COVID-19 infection.	2018 CDC Social Vulnerability Index
Health Disparities
Uninsured	Static		Percentage uninsured in the total civilian noninstitutionalized population estimate, 2014- 2018 ACS. Individuals without insurance are more likely to be undercounted in infection statistics, and may have more severe outcomes due to lack of treatment.	2018 CDC Social Vulnerability Index
Transmissible Cases
	Daily		Population size divided by cases from the last 14 days. Because of the 14-day incubation period, the cases identified in that time period are the most likely to be transmissible. This metric is the number of such “contagious” individuals relative to the population, so a greater number indicates more likely continued spread of disease.	USA Facts
Disease Spread
	Daily		Fraction of total cases that are from the last 14 days (one incubation period). Because COVID-19 is thought to have an incubation period of about 14 days, only a sustained decline in new infections over 2 weeks is sufficient to signal reduction in disease spread. This metric is always between 0 and 1, with values near 1 during exponential growth phase, and declining linearly to zero over 14 days if there are no new infections.	USA Facts
Social Distancing
	Daily		Unacast social distancing scoreboard grade is assigned by looking at the change in overall distance travelled and the change in nonessential visits relative to baseline (previous year), based on cell phone mobility data. The grade is converted to a numerical score, with higher values being less social distancing (worse score) is expected to increase the spread of infection because more people are interacting with other.	Unacast
Vaccination Full Dose
Series_Complete_Pop_Pct	Daily		Percent of people who are fully vaccinated (have second dose of a two-dose vaccine or one dose of a single-dose vaccine) based on the jurisdiction and county where recipient lives.	CDC