Currently, we are working with the New York City Department of Health and Mental Hygiene to generate projections of COVID-19 in the city to support decision making, logistics, and planning. Our model-inference/projection system serves three main objectives: 1) assess effectiveness of control measures (e.g. social distancing); 2) predict epidemic outcomes (new infections, hospitalizations, ICU admissions, intubations) under different control scenarios; and 3) predict demands to the healthcare systems (numbers of hospital beds, ICU beds, and ventilators needed) under different control scenarios.
Our initial model-inference/projection system used a simple Susceptible-Exposed-Infectious-Removed (SEIR) model in conjunction with the Ensemble Adjustment Kalman Filter (EAKF), accounting for delay in reporting and under-reporting. The projection system was first trained/calibrated to weekly incidence (i.e. confirmed case counts) and then projected to the next 8 weeks under different control scenarios. This system was used for projections generated from 3/16 to 4/3/2020.
Our 2nd model-inference/projection system used a similar SEIR-EAKF system, however, further account for differences by age (e.g. reporting rate and disease severity) and was trained using age-specific incidence data. This system is used for projections generated from 4/3/2020 onwards.
Our most recent model-inference/projection system further accounts for spatial heterogeneity using a network model capturing connections among 42 United Hospital Fund (UHF) neighborhoods in NYC. This new system also incorporates mobility data provided by Safegraph.com to gauge changes in transmission within each UHF neighborhood and between pairs of neighborhoods. These neighborhood-level projections are available since 4/29/2020.
All projections are publicly available on Github: