stevegoldstein / covid19-severity-prediction

COVID19 data + modeling at the county-level + hospital-level.

Overview

Note: This repo is actively maintained - for any questions, please file an issue.

Our group at UC Berkeley is working to help forecast the severity of the epidemic both for individual counties and individual hospitals. As a byproduct, we have and will continue to produce models, visualizations, and curated datasets (including confirmed cases/deaths, demographics, risk factors, social distancing data) that can be used by other teams in the fight against COVID-19. We are collaborating with Response4Life, a non-profit organization, whose goal is to blunt the effect of COVID-19 through the production and appropriate distribution of PPE, medical equipment, and medical personnel to healthcare facilities across the United States. Paper link: "Curating a COVID-19 data repository and forecasting county-level death counts in the United States".

• Visualizations (updated daily): see interactive county-level map + interactive hospital-level map
• Data (updated daily): We have compiled and cleaned a large corpus of hospital- and county-level data from a variety of public sources to aid data science efforts to combat COVID-19.
  • At the hospital level, the data include the location of the hospital, the number of ICU beds, the total number of employees, the hospital type, and contact information
  • At the county level, our data include socioeconomic factors, social distancing scores, and COVID-19 cases/deaths from USA Facts and NYT
• Modeling: Using this data, we have developed a short-term (3-5 days) forecasting model for mortality at the county level. This model combines a county-specific exponential growth model and a shared exponential growth model through a weighted average, where the weights depend on past prediction accuracy.
• Severity index: The Covid pandemic severity index (CPSI) is designed to help aid the distribution of medical resources to hospitals. It takes on three values (3: High, 2: Medium, 1: Low), indicating the severity of the covid-19 outbreak for a hospital on a certain day. It is calculated in three steps.
  1. county-level predictions for number of deaths are modeled
  2. county-level predictions are allocated to hospitals within counties proportional the their total number of employees
  3. final value is decided by thresholding the number of cumulative predicted deaths for a hospital (=current recorded deaths + predicted future deaths)

Quickstart with the data + models

Data

Can download, load, and merge the data via:

import load_data
# first time it runs, downloads and caches the data
df = load_data.load_county_level(data_dir='/path/to/data') 

• for more data details, see ./data/readme.md
• see also the county_quickstart notebook
• we are constantly monitoring and adding new data sources (+ relevant data news here)
• output from running the daily updates is stored here

Prediction

• To get deaths predictions for our current best-performing model, the simplest way is to call (for more details, see ./modeling/readme.md)

from modeling.fit_and_predict import add_preds
df = add_preds(df, NUM_DAYS_LIST=[1, 3, 5]) # adds keys like "Predicted Deaths 1-day", "Predicted Deaths 3-day"
# NUM_DAYS_LIST is list of number of days in the future to predict

Related county-level projects

• County-level data summaries from JHU
• UChicago GeoData visualization team

Acknowledgements

To reference, please cite the paper: Curating a COVID-19 data repository and forecasting county-level death counts in the United States

The UC Berkeley Departments of Statistics, EECS led by Professor Bin Yu (group members are all alphabetical by last name)

• Yu group team (Data/modeling): Nick Altieri, Rebecca Barter, James Duncan, Raaz Dwivedi, Karl Kumbier, Xiao Li, Robbie Netzorg, Briton Park, Chandan Singh (student lead), Yan Shuo Tan, Tiffany Tang, Yu Wang
• response4Life team and volunteers (Organization/distribution)
• Kolak group team (Geospatial visualization): Qinyun Lin
• Medical team (Advice from a medical perspective): Roger Chaufournier, Aaron Kornblith, David Jaffe
• Shen Group team (IEOR): Junyu Cao, Shunan Jiang, Pelagie Elimbi Moudio
• Helpful input from many including: SriSatish Ambati, Rob Crockett, Tina Elassia-Rad, Marty Elisco, Nick Jewell, Valerie Isham, Valerie Karplus, Andreas Lange, Ying Lu, Samuel Scarpino, Jas Sekhon, Phillip Stark, Jacob Steinhardt, Suzanne Tamang, Brian Yandell, Tarek Zohdi