Open-Source COVID-19 ML Datasets; Models; Tools for Data Scientists

In this blog, I'll discuss essential parameters for Covid19 disease spread, in machine learning models, the types of different models, tools, and datasets available.

• Monitoring and forecasting disease spread
• Improving health equity and minimizing secondary effects of the pandemic
• Supporting healthcare workers
• Slowing transmission by advancing the science of contact tracing, environmental sensing
• Devising effective vaccination plans.

Age-based Mortality Models

This type of ML model is trained using age-stratified Generalized linear models (GLMs) with component-wise gradient boosting. It helps to predict the probability of death based on information available for patients before they contracted the virus. The process of stratification of the overall model by age groups helps to reduce the variability in age and to identify risk factors of different ages.

In the overall model, 18 features were identified in at least 20% of the models (2 of 10) as being associated with increased mortality risk. Data scientists/researchers took Odds Ratios (ORs) with interquartile ranges (IQR) to compare the relative importance of the variables for predicting mortality. Of these features, age had the most prominent association—median OR: 2.82 (iqr: 0.03)—for predicting mortality.

Google Data Studio

SELECT
*
FROM
bigquery-public-data.covid19_public_forecasts.state_28d
WHERE
state_fips_code = “48”
AND prediction_date >= forecast_date
ORDER BY
prediction_date

Covid19 datasets and models

• Google Aggregated Mobility Research Dataset – Aggregates weekly flows of users from region to region, where the region is at a resolution of 5km2
• : contains dataset from aggregated data sources, both US and non-US
• : Making data freely accessible for better public outcomes
• by
• : containing USA’s daily report state-wise with confirmed cases, time-series summary, etc
•  

Regression-based Time-Series Modeling with Covid19

This is a forecasting approach for COVID-19 case prediction relying on Graph Neural Networks and mobility data. This ML modeling approach uses a single large-scale Spatio-temporal graph, with the following assumptions.

• Spatial domain –  Edges represent direct location-to-location movement and are weighted by mobility flows, based on, the amount of flow internal to the location.
• Temporal domain – Edges represent a binary connection to past days.
• Each node contains features for the state, county, day, past cases, and past deaths

The above COVID-19 graph showing spatial and temporal edges (highlighted in red) across three days. Each slice represents spatial connections between counties, while the connections between slices represent temporal relationships. Every node in the graph has direct temporal edges to nodes in d previous days.

The above figure represents the 2-hop Skip-Connection model. Multiple layers of spatial aggregations are used on temporal embedding vectors. At each layer, the embedding of the seed node (represented in blue) is concatenated and propagated up to the next embedding layer. The final embedding is passed through an MLP and used to predict P

Fairness on Covid-19 Datasets

As Google is committed to Responsible AI principles, it has come forward, to study the disproportionate impact, the disease has had in the United States. As a pioneer of Fairness in AI, Google’s AI team could try to follow “Avoid creating or reinforcing unfair bias”, to study the actual impact of the disease.

CDC research has shown that communities of color in the United States have been the hardest hit by COVID-19 with disproportionately high rates of cases and deaths. The causes of it are related to structural racism, various systemic inequities in access to healthcare, inherent systemic bias, and underlying negatively impacting social determinants of health.

The below figure illustrates:

• The absolute error in predicting covid19 deaths is significantly higher for counties with a higher proportion of younger and middle-aged people.
• The demographic sections of these countries comprising of younger and middle-aged groups have a higher proportion of COVID-19 case counts.
• Further, after the absolute errors are normalized by actual death counts, there is less difference between the confidence intervals across the demographic groups.

During the analysis of median income, the analysis was done by bucketing (segregating them to bins) county populations according to their income. The results are represented by the bottom figure which shows higher absolute errors for higher-income counties.

Similarly, for Race and Ethnicity, the figure clearly depicts, there is a direct correlation between the absolute errors and death counts, and this is meaningfully reduced when the error is normalized by the death count, causing the confidence intervals to overlap.

Interpretable Sequence Learning with Covid19 datasets

This kind of ML framework proposed how different compartments (composed of different direct and indirect factors that affect prediction coefficients) evolve. It uses interpretable encoders to incorporate covariates and improve model performance. The performance of the model has been further analyzed for different subgroups based on the subgroup distributions within the counties.

The model is based on an extension to the standard SEIR (susceptible–exposed–infectious–removed) model that includes additional compartments for undocumented cases and hospital resource usage. The end-to-end modeling framework can infer meaningful estimates for undocumented cases even if there is no direct supervision for them.

• Masked supervision from partial observations,
• Partial teacher-forcing to minimize error propagation,
• Regularization, and
• Cross-location information-sharing

• Introduction of compartments for undocumented infected and recovered cases
• Introduction of hospitalized, ICU, and ventilator compartments
• Partial immunity
• No death from undocumented infected cases
• Invariant populationExplainable AI for Covid19

Explainable AI for Covid19

The most important characteristics of Interpretable Sequence-Learning works on the basis of modeling the compartments explicitly to provide an understanding of disease evolution. The below figure demonstrates how the fitted curves can be used to infer important insights on where the peaking occurs or the current decay trends.

The ratio of undocumented to documented infected at different phases is computed, as well as the amount of increase/decrease for each compartment is analyzed. For intervention covariates, the largest weights (with significant changes of disease spread) is noticed after a lag of a few days, suggesting their effectiveness after some lag. The positive weights of the mobility index, and negative weights of public interventions are also clearly observed.

The above figure demonstrates the Learned weights of the time-varying covariates for β (Average contacts of doc. infected/undoc. infected), for 7-day state-level forecasting models for three weeks starting from 24th May 2020 to 7th June 2020. It is observed that the mobility index consistently has a highly positive impact on β while gathering bans, school closures and shelter-in-place interventions have highly negative effects. In addition, the weight magnitude of the interventions gets larger after a lag of few days.

Impact of non-pharmaceutical interventions (NPIs) to reduce COVID-19 mortality and healthcare demand

The exponential rise of COVID-19 cases and the number of deaths have forced governments in different countries to introduce interventions too early. However, it possesses the risk of allowing the transmission to return once they are lifted (if insufficient herd immunity has developed).

 – Mitigation strategy scenarios for GB showing critical care (ICU) bed requirements. The black line shows the unmitigated epidemic. The green line shows a mitigation strategy incorporating closure of schools
and universities; the orange line shows case isolation; the yellow line shows case isolation and household quarantine; and the blue line shows case isolation, home quarantine, and social distancing of those aged over 70. The blue shading shows the 3-month period in which these interventions are assumed to remain in place.

AI/ML Model-based COVID-19 vaccine prioritization

This method involves building a mathematical SEIR model (susceptible, exposed, infectious, recovered) to compare five age-stratified prioritization strategies. The prioritization strategies remain consistent across countries, transmission rates, vaccination rollout speeds, and estimates of naturally acquired immunity. In addition, this ML-based framework allows comparing the impacts of prioritization strategies across contexts.

Figure A demonstrates age-dependent vaccine efficacy shows a decrease from 90% baseline efficacy to 50% efficacy among individuals aged 80+ years, beginning at age 60. Figures (B and C) Percent reduction in deaths in comparison with an unmitigated outbreak for transmission-blocking all-or-nothing vaccines with either constant 90% efficacy for all age groups (solid lines) or age-dependent efficacy. 

Conclusion

References

• COVID-19 Forecasting using Spatio-Temporal Graph Neural Networks 
• Contribute at: 
• Impact of non-pharmaceutical interventions (NPIs) to reduce COVID-19 mortality and healthcare demand – 
• Mobility network models of COVID-19 explain inequities and inform reopening
• A review on COVID-19 forecasting models 
• New machine learning model predicts who may benefit most from COVID-19 vaccination
• Prioritize vaccine delivery – 
• A multi-stage SEIR model to predict the potential of a new COVID-19 wave in KSA after lifting all travel restrictions 
• Predicting Coronavirus Pandemic in Real-Time Using Machine Learning and Big Data Streaming System 
• Concept Drift and the Impact of COVID-19 on Data Science