-
Notifications
You must be signed in to change notification settings - Fork 46
/
metadata-USC-SIkJalpha.txt
55 lines (55 loc) · 5.48 KB
/
metadata-USC-SIkJalpha.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
team_name: University of Southern California
model_name: SI_kJalpha
model_abbr: USC-SI_kJalpha
model_version: 2022-03-13
model_contributors: Ajitesh Srivastava <[email protected]>, Majd Al Aawar <[email protected]>
website_url: https://scc-usc.github.io/ReCOVER-COVID-19
license: mit
team_model_designation: primary
methods: Uses SIKJalpha which models temporally varying infection, death, and hospitalization rates. Learning is performed by reducing the problem to multiple simple linear regression problems. Effects of variants are modeled separately as competing contagions.
modeling_NPI: [Round 13 update] No NPI included.
compliance_NPI: Not modeled.
contract_tracing: Not applicable.
testing: Not applicable.
vaccine_efficacy_transmission: Not explicitly modeled
vaccine_efficacy_delay: 14 days
vaccine_hesitancy: A model (much like "suceptible-infected") fitted on each age group. Maximum coverage of each group taken as the upper bound on the coverage on the lower age-group. Boosters are modeled such that the "susceptible" population is the eligible population who are yet to get a booster. To achieve the desired saturation level to create sub-scenarios, we increase the adoption rate over time.
vaccine_immunity_duration: Only waning immunity is considered! The probability of transferring to partially immune state by at time t is modeled as a gamma distribution such that: (i) the median is as per the scenario; (ii) the efficacy (given the partial immune protection against infection) after the first 60 days is the "initial" vaccine efficacy (e.g. 80% for 65+, 90% for <65).
natural_immunity_duration: Same as above
case_fatality_rate: Learned from recent data.
infection_fatality_rate: Learned from recent data.
asymptomatics: Using a combination of wastewater data from Biobot and sero-prevelance data from the CDC. Note that this covers the currect cumulative fraction of unreported, untested, and asymptomatic.
age_groups: [0-4, 5-11, 12-17, 18-65, 65+] Age-wise vaccine, cases, hospitalizations, deaths for training as well as projections. Population with waned immunity tracked for each age-group
importations: Not applicable
confidence_interval_method: We generate multiple trajectories by (i) Uniform sampling of the recently seen infection rates, under-reporting/asymptomatic population, vaccine coverage, waning parameters uncertainty, and variant prevelance. (ii) For each of the above case trajectories, we use multiple possibilities of death and hospitalization rates that have been seen in the recent past. The above results in multiple points for each day. Quantiles are generated based on sampling among these points for each day.
calibration: All calibrations using regression. No manual-tuning other than the listed assumption and trivial settings: non-zero lag between infection and death.
spatial_structure: Not applicable.
team_funding: National Science Foundation Award # 2135784
data_inputs: Cases and deaths - JHU, hospitalization - HHS time-series
citation:
[1] Ajitesh Srivastava, Tianjian Xu and Viktor K. Prasanna, "Fast and Accurate Forecasting of COVID-19 Deaths using the SIkJalpha Model". https://arxiv.org/abs/2007.05180.
[2] Ajitesh Srivastava and Viktor K. Prasanna, "Data-driven Identification of Number of Unreported Cases for COVID-19: Bounds and Limitations". SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2020) Health Day Track.
[3] Ajitesh Srivastava and Viktor K. Prasanna, "Learning to Forecast and Forecasting to Learn from the COVID-19 Pandemic". https://arxiv.org/abs/2004.11372
methods_long: We use our own epidemic model called SI-kJalpha, preliminary version
of which we have successfully used during DARPA Grand Challenge 2014. Our model
can consider the effect of many complexities of the epidemic process and yet be
simplified to a few parameters that are learned using fast linear regressions. Therefore,
our approach can learn and generate forecasts extremely quickly.Despite being fast, the accuracy of our forecasts is
on par with the state-of-the-art as demonstrated on the leaderboard page. at https://scc-usc.github.io/ReCOVER-COVID-19.
Our model is able to quickly adapt to changing trends, and the variations in parameters
during different times/policies allow us to forecast different scenarios such as
what would happen if we were to disregard social distancing suggestions.
For each state, we model hospitalizations as a separate compartment, as a linear
function of recent cases with heterogeneous rates. This means that for some
hyper-parameter J, those who were infected at time t-1 to t-J will have a separate
rate from those infected at t-(J+1) to t-2J, and so on. This is similar to how we
perform death forecasts. For long-term forecasts (more than a few days in the future),
the cases are forecasted first based on our SIkJalpha model, which forms the input to
hospitalization prediction. Age-specific compartments are tracked for individuals with waned immunity.
While we account for changing trends by putting more emphasis on recently seen data,
we assume that the trends will remain the same in the future for our point forecasts.
Quantiles are generated by considering variations in the infection/fatality/hopsitalization
over recent window. Variants are modeled separately as competing contagions
after estimating prevalence using GISAID/GenBank data. Provided immune escape parameters are applied to prior
infection and 2-dose vaccinations. We track all possible permutations of vaccinations and (re)infections
to compute immunity in the population.