Universal Risk Phenotype of US Counties for Flu-like Transmission Dynamics

The spread of a transmissible virus is a complex spatio-temporal process shaped by the specific transmission mechanism of the virus, the survivability of the pathogen outside the host under varying enviromental factors such as temperature and humidity, and the ease of access to new viable hosts broadly determined by the local population density and possibly the level of enforcement of social distancing policies. Other factors such as demographic and medical factors might make specific host more susceptible, which might also impact spread. While the key factors that shape transmission for influenza and COVID-19 are broadly understood, making precise actionable forecasts is still a diffcult modeling problem. Faced with the challenge of forecasting the case count in the context of the COVID-19 pandemic, a diversity of modeling approaches have been suggested. However a single "best" model is yet to emerge. In this study we introduce the concept of a universal geo-spatial risk measure, to quantify the risk phenotype of US counties for spread mediated via flu-like transmission mechanisms. We demonstrate that the new risk phenotype significantly improves incidence forecasts, and emerges as the most important factor "explaining" county-specific incidence trends.

The Hypothesis

We hypothesize that the seasonal flu epidemic encodes complex geo-spatial stochastic incidence patterns that are common to, and drive, COVID epidemiology. Using a novel approach to quantifying similarity of stochastic sample paths, we learn from nearly a decade of historical flu seasonal epidemics, and effectively incorporate these patterns to model and forecast COVID case counts. Our approach demonstrates that broad commonalities in transmission dynamics may be effectively leveraged for forcasting the number of new cases, even if the specific pathogen of interest is biologically distinct from the one whose historical incidence patterns are actually available. In the context of influenza and SARS-CoV2 viral transmission modeling, the geome corresponding to influenza is useful for predicting COVID-19 cases, since the tranbsmission dynamics is believed to be largely similar.

Risk Calculation Approach

  • We collect countywise infection count data over a period of 9 years from 2003 to 2013 flu season in US
  • We use a novel approach to quantifying similarity between categorical time series samples to identify natural clustering of the counties, according to the sequential count variations
  • Using insight from our previous work (https://elifesciences.org/articles/30756), and this similarity metric, we identify the emergent cluster of counties which includes the set corresponding to estimated epidemic initiation over the flu seasons. These are coastal counties with high population and the right conjunction of weather, demographic, and socio-economic factors that trigger the seasonal flu epidemic. We hypothesize that the risk of initiation of the yearly flu season also track the risk of COVID-19 infections.

  • Using the infection count variations in these "high initial risk" counties (lets denote this set as \(G_0\) and our prior work on inferring finite state probabilistic automata (PFSA) models of finite valued stochastic processes (https://royalsocietypublishing.org/doi/full/10.1098/rsta.2011.0543), we infer a generative model \(\mathcal{G}_0 \) from the set of quantized count variatoions \( G_0\).

  • Finally we use sequence likelihood divergence to quantify the deviation of the count variation in each county from this model, to formulate a time-independent spatially varying measure of universal initiation risk (\(\mathbf{u}_0\)), which we call the geome.

  • We construcnt a General Linear Model (Poisson log-linear regression) to connect this covariate along with other suspected covariates to bi-weekly COVID-19 case counts.

Results & Implications

The GLM solution has the following implications, results and observations:

  1. The \(\mathbf{u}_0 \times r\) is the most imprtant covaraite where \( r \) is the fraction of the population living in an urban environment (non-rural as defined by US census) in each county
  2. The model we find correlates very well with confirmed COVID cases across US counties
  3. The model is robust, in the sense that it is stable to diverse perturbations we experimented with
  4. We can use this model to make county-specific forecast of COVID counts in the near term

Impact of \(\mathbf{u}_0 \times r\) Dominance

This shows that learning patterns from the seasonal flu epidemic may be meaningfully transferred to the COVID-19 epidemiology, or perhaps any virus with a similar transmission mechanism. And the information encoded in these patterns is hard to replicate by manually including new covariates in the model; there are always some factors that are unmodeled, or observed patterns not effectively explained by covriates we can a priori imagine to have an impact in the transmission dynamics. Instead of attempting to identify all of such factors along their potentially complex dependencies, our approach treats the past sequential observations as providing the set of similarities or patterns that we expect to drive future infections. Our key insight here is that these potentially unmodelled spatially varying degrees of freedom (the geome) in the underlying driving dynamics are similar for COVID-19 and flu, vindicated by our result that a measure of this risk scaled by the fraction of county-specific urban population, indeed emerges as the most important covariate in standard regression models, over socio-economic and demograhic factors typically considered in the literature. Also, model output correlates very well with actual count data collected within the same modeling period, as well as in the immediate future (2 weeks). Other measures of goodness of fit and explainability are also demonstrated, clearly establishing that the geome does indeed significantly improve county wise count prediction for the COVID-19 pandemic in the USA.

Forecasts: What we Bring To the Table

State of the Art

The Center for Disease COntrol has gathered a suite of modeling approaches fof COVID-19 case count, with bi-weekly prjections at the nation and state level. These models use a vast diversity of approaches. https://www.cdc.gov/coronavirus/2019-ncov/cases-updates/forecasts-cases.html Our aooroach here is to use the geome data along with suspected covariates to infer a predictor for data within a 2-week period, and then project the counts over the next 2 week period. In that sense, a sketch for our algorithm is as follows:

 Input: u_0, v_1, \cdots v_m
 Infer: GLM(u_0, v_1, \cdots v_m)
 Predict: Y from GLM model
 Infer Feedforward Neural Net: X_t=NN(X_{t-1},Y) where X_{t-1} is the case count vector in the previous period t-1
 Predict: NN(X_t,Y) as the case count in period t+1

GenESeSS Command used

./genESeSS -f

Innovation

Impact

Epidemiologically Informed Precise Incidence Forecasts for COVID-19 and Influenza-like Illnesses

Geome of Flu-like Transmission

Table of Contents