UNSTRUCTURED
Introduction
The COVID-19 pandemic exhibits an uneven geographic spread which leads to a locational mismatch of testing, mitigation measures and allocation of healthcare resources (human, equipment, and infrastructure).(1) In the absence of effective treatment, understanding and predicting the spread of COVID-19 is unquestionably valuable for public health and hospital authorities to plan for and manage the pandemic.
While there have been many models developed to predict mortality, the authors sought to develop a machine learning prediction model that provides an estimate of the relative association of socioeconomic, demographic, travel, and health care characteristics of COVID-19 disease mortality among states in the United States(US).
Methods
State-wise data was collected for all the features predicting COVID-19 mortality and for deriving feature importance (eTable 1 in the Supplement).(2) Key feature categories include demographic characteristics of the population, pre-existing healthcare utilization, travel, weather, socioeconomic variables, racial distribution and timing of disease mitigation measures (Figure 1 & 2).
Two machine learning models, Catboost regression and random forest were trained independently to predict mortality in states on data partitioned into a training (80%) and test (20%) set.(3) Accuracy of models was assessed by R2 score.
Importance of the features for prediction of mortality was calculated via two machine learning algorithms - SHAP (SHapley Additive exPlanations) calculated upon CatBoost model and Boruta, a random forest based method trained with 10,000 trees for calculating statistical significance (3-5).
Results
Results are based on 60,604 total deaths in the US, as of April 30, 2020. Actual number of deaths ranged widely from 7 (Wyoming) to 18,909 (New York).CatBoost regression model obtained an R2 score of 0.99 on the training data set and 0.50 on the test set. Random Forest model obtained an R2 score of 0.88 on the training data set and 0.39 on the test set.
Nine out of twenty variables were significantly higher than the maximum variable importance achieved by the shadow dataset in Boruta regression (Figure 2).Both models showed the high feature importance for pre-existing high healthcare utilization reflective in nursing home beds per capita and doctors per 100,000 population. Overall population characteristics such as total population and population density also correlated positively with the number of deaths.Notably, both models revealed a high positive correlation of deaths with percentage of African Americans. Direct flights from China, especially Wuhan were also significant in both models as predictors of death, therefore reflecting early spread of the disease.
Associations between deaths and weather patterns, hospital bed capacity, median age, timing of administrative action to mitigate disease spread such as the closure of educational institutions or stay at home order were not significant. The lack of some associations, e.g., administrative action may reflect delayed outcomes of interventions which were not yet reflected in data.
Discussion
COVID-19 disease has varied spread and mortality across communities amongst different states in the US. While our models show that high population density, pre-existing need for medical care and foreign travel may increase transmission and thus COVID-19 mortality, the effect of geographic, climate and racial disparities on COVID-19 related mortality is not clear.
The purpose of our study was not state-wise accurate prediction of deaths in the US, which has already been challenging.(6) Location based understanding of key determinants of COVID-19 mortality, is critically needed for focused targeting of mitigation and control measures.
Risk assessment-based understanding of determinants affecting COVID-19 outcomes, using a dynamic and scalable machine learning model such as the two proposed, can help guide resource management and policy framework.