Assessing community vulnerability to initial COVID-19 spread in Florida ZIP Codes using Shapley additive explanations with random forest modeling

doi:10.21203/rs.3.rs-4907544/v1

Download PDF

Research Article

Assessing community vulnerability to initial COVID-19 spread in Florida ZIP Codes using Shapley additive explanations with random forest modeling

https://doi.org/10.21203/rs.3.rs-4907544/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Background

The identification of the population attributes that play important roles in the early-phase community spread of an epidemic is critical to improving our ability to prepare and develop the resilience of societies to future pandemic-potential pathogens. Our study aimed to assess the vulnerability of communities at the ZIP code-level in the state of Florida to the growth in the case incidence of the COVID-19 epidemic during its initial outbreak phase using local case and CDC/ATDSR SVI data and the application of a novel explainable machine learning model.

Methods

The COVID-19 growth rates were estimated from a log-linear regression fitted to the daily number of cases reported for the initial wave of the pandemic in each ZIP code (n = 935). A random forest model was trained to predict COVID-19 growth rates using 22 social vulnerability indicators. The trained model was interpreted with Shapley additive explanations (SHAP) to investigate the contribution of social vulnerability features to early COVID-19 spread across all ZIP codes in Florida. SHAP feature ranking and results were used to calculate a Social Vulnerability Index (SVI) for each ZIP code.

Results

Estimated COVID-19 growth rates ranged from 1 to 1.247 (mean = 1.054). The percent of single-parent households was the most important feature in predicting growth rates, followed by (in order) population density and the percentages of the population facing language barriers, living in group quarters, burdened by housing costs, and diagnosed with coronary heart disease in a ZIP code. High values of the five highest ranking features were shown to contribute positively to predicted growth rates, whereas high values of the sixth feature contributed negatively. The constructed SVI had a significant positive association (p-value < 0.0001) with the ZIP code-level epidemic growth rates.

Conclusions

The constructed ML-SHAP modeling approach and SVI can help assess the social vulnerability of communities to the early COVID-19 spread that was observed in Florida ZIP codes. They can also serve to identify high risk sub-populations and localities, which will be important for advancing development of mitigation strategies to prevent, enhance community resilience, and respond to future novel pathogens of pandemic potential.

Social vulnerability

COVID-19

Machine Learning

Explanative Artificial Intelligence

The COVID-19 pandemic dramatically highlighted the vast damage that biothreats posed by infectious diseases can cause to human societies. Apart from causing over 6.9 million deaths and continuing long-term morbidity globally [1], the pandemic also underscores how such infectious hazards can also affect the economies of societies, with studies indicating that it reduced GDP across countries in 2020 by 2–4% whilst inducing in the case of the United States GDP its worst contraction since World War II [2]. The spread and effects of the health crisis were also particularly visible in the United States: all 50 states and 90% of all counties were affected within the first four months of the virus's arrival in the country [3]. While most early studies both in the United States and internationally understandably focused on the aggregate trajectories and sizes of evolving waves of cases, hospitalizations, and deaths arising from epidemic spread within a spatial region, it was noted even at the earliest stages that these outcomes are not evenly distributed across a population [4, 5]. Indeed, COVID-19 has been called a syndemic pandemic, with studies conducted during different phases of the pandemic discovering the strong links that exist between socioeconomic inequalities and increases in case incidence, severity, and mortality within a community [6–10].

Social vulnerability to outbreaks refers to the susceptibility of communities to the negative impacts of an epidemic disease due to various factors such as socioeconomic status, access to resources, infrastructure, and governance [11–14]. The concept is premised on long evidence that community-level social disadvantage and inequalities can influence the incidence of infectious diseases and their adverse outcomes in a population [14, 15]. Studies from natural disaster management has further indicated how understanding a community’s vulnerability to a particular hazard may be used to enhance its resilience to the future consequences arising from the hazard by either reducing the risk of exposure or by modifying the risk of consequences from future exposures [14]. These considerations have led to the carrying out of a multitude of social vulnerability-focused studies by researchers not only to identify communities that are more susceptible or vulnerable to COVID-19 transmission, but also to evaluate the sociodemographic, health, and economic components that increase the risk of infection, severity, and mortality in populations [16–21].

While these studies have increased understanding of the social and economic factors associated with SARS-CoV-2 incidence and adverse effects in different communities, the vulnerability of communities to the early phase of the pandemic is less well studied [21–25]. Understanding this phase of an epidemic is important because it provides first information regarding the transmissibility of a new disease outbreak and thus the time-scale or speed with which it may spread within and between communities [26, 27]. The rate of this early, often exponential, growth in the number of cases or deaths and how this may vary between different locations will also determine the magnitude of its overall impact for a region [28]. Identification of the attributes that play important roles in the early-phase community spread of an epidemic can thus be critical to improving our ability to prepare and develop the resilience of societies to future novel pandemic-potential pathogens [14].

The Social Vulnerability Index (SVI), a metric developed by the Geospatial Research, Analysis, and Services Program (GRASP) at the Centers for Disease Control and Prevention (CDC) and the Agency for Toxic Substances and Disease Registry (ATSDR), provides a composite measure of community susceptibility to adversities in the face of health shocks [11, 12, 29], and has been widely used in carrying out COVID-19 community vulnerability studies. The 2020 CDC/ATSDR SVI is composed of several variables summarized into four main themes: 1) Socioeconomic Status, 2) Household Characteristics, 3) Racial and Ethnic Minority Status, and lastly 4) Housing Type and Transportation [29], and can be readily extended to include other social risk factors [19]. An important property of this index is that the data used are available from different sources and furthermore at different spatial scales, including at smaller geographic units, such as census tracts and ZIP codes [19]. These factors, along with the fact that the metric offers a standardized indicator of the vulnerability of communities, have resulted in not only its widespread adoption but also applicability for gaining a consistent understanding of a community’s vulnerability to various health and other disasters as a result of its socioeconomic condition.

There have been numerous studies that have examined the factors related to social vulnerability in the event of COVID-19 spread based on the variables used for constructing the CDC/ATSDR SVI [16, 18, 20, 30, 31]. Many of previous works have used either descriptive statistics or traditional statistical analysis tools [16, 18, 20, 32]. Recently, however, focus has shifted to the use of Machine Learning (ML) algorithms in order to leverage their abilities for handling a large number of predictors, address complex interactions between predictors and responses, and model non-linear associations between variables and responses as well as between the variables themselves without also making any distributional assumptions about the data [33–36]. Thus, Tiwari et al. [37] devised a COVID-19 vulnerability index by employing a Random Forest (RF) framework, and subsequently classified counties in the United States into different vulnerability groups, while Moosazadeh et al. [38] developed ML algorithms to assess the regional impact of 30 county-level social vulnerability features on the case fatality rate of COVID-19 in the United States. Tatar et al. [17] also looked at the relationship between social vulnerability components and COVID-19 cases during an initial transmission period of 30 days in counties of the United States Southeast region using XGBoost and SHapley Additive exPlanations (SHAP) as an ML-based explanative modeling framework. Additionally, Itzhak et al. [39] employed ML models to classify United States counties by levels of COVID-19 morbidity and mortality using a variety of predictors before applying SHAP to determine variable importance.

In this study, we extend these previous work by coupling two sets of factors: 1) an epidemiological variable that describes the growth of COVID-19 during the early exponential epidemic phase together with components of the CDC/ATDSR SVI, with 2) an interpretable ML model [17, 39] to determine and analyze the factors that contributed to the establishment and spread of the pandemic during its initial growth phase at the ZIP code-level across Florida. We combined a RF model with SHAP [40] to undertake this analysis with the objectives of not just to gain a better understanding of which sociodemographic and economic risk factors underscored the vulnerability of local communities to the early growth of the pandemic that occurred in these ZIP codes, but also conversely which of these variables restricted and thereby contributed to community resilience to the observed initial spread of the outbreak. We consider that focusing this analysis on Florida is particularly useful for providing insights regarding the impact of social vulnerability and resilience on early outbreak spread given its highly heterogeneous population distribution. We also chose to investigate the association of social vulnerability components and COVID-19 growth at the ZIP code-level since it represents a spatial scale that may be considered to be large enough to exhibit sufficient social heterogeneity and yet allow examination of how key social factors may increase resilience to or support the early geographic spread of a respiratory contagion across a region. Given the influence of local resources and community attributes, we determined that such a high-resolution spatial analysis will further be key to helping identify critical gaps to better inform local-level decision and policy making. This in turn will be central to the development and implementation of the appropriate preparedness activities and future responses to prevent and/or reduce the impacts of current and novel outbreak diseases.

2.1 Study setting and overview

Our study focused on assessing the vulnerability (and resistance) of communities at the ZIP code-level in the state of Florida to the growth in the case incidence of the COVID-19 epidemic during its initial outbreak phase in the state. To calculate the initial growth rate, we used the number of new daily COVID-19 cases from the date of the first recorded case to the peak of the first wave in each ZIP code. Here, the first wave was considered as the rise in cases immediately after the Florida state government lifted its stay-at-home order—which went into effect on 3 April 2020 [41] soon after community spread was perceived in the state—on 4 May 2020 [42]. As predictors, we use a total of 22 social vulnerability features inspired by the CDC/ATSDR SVI. Using this data, we first trained and validated a RF model. Then, we applied the SHAP framework to the trained model to appreciate the relationship between cumulative COVID-19 cases and important social vulnerability features across all ZIP codes in Florida. Figure 1 summarizes the methodology of our work.

2.2 Data assembly

2.2.1 Growth rate

COVID-19 daily case data for Florida ZIP codes was obtained from the Tampa Bay Times [43]. This website contains daily new cases data for Florida ZIP codes from April 6, 2020 to June 2021. For each ZIP code, the peak of the wave was identified as the day with the maximum number of cases from the day the stay-at-home order was lifted to the end of September 2020, which marks the end of the first post-lockdown wave statewide. The growth rate for each ZIP code was calculated through the fit of a log-linear regression model in the form of $\:\text{log}\left(y\right)\sim\:rt+{\beta\:}_{0}$ where $\:y$ is the number of daily cases from the date of first record to the date of the peak specific to each ZIP code, $\:r\:$is the growth rate, $\:t$ is the number of days, and $\:{\beta\:}_{0}$ is the model intercept [26, 27]. The estimated growth rates for each ZIP code are presented in Additional File 1.

2.2.2 Socioeconomic vulnerability variables

The SVI proposed by CDC and ATSDR calculates the vulnerability of an administrative area by summing the percentile ranks for all the social variables by themes that they represent [29]. For this work, we use the 16 variables included in the 2020 CDC/ATSDR SVI: people living in poverty, unemployed, burdened by housing costs, with no high school diploma, with no health insurance, aged 65 and older, aged 17 and younger, with disabilities, in single-parent households, with language barriers, identifying with a racial or ethnic minority, in group quarters; and housing that are multi-unit, mobile homes, overcrowded, and that do not have a vehicle available [29]. Table 1 lists the definitions of these variables as used in this study. Additional variables were included for population density and comorbidities that have been linked to severe outcomes of this disease [44–46] to make this index even more applicable to the COVID-19 context. The ZIP code-level data for the variables are obtained from PolicyMap [47]. The data for the socioeconomic variables extracted from this database are based on the 2018–2022 data from the U.S. Census Bureau, while comorbidities are based on 2020 data from CDC. The unemployment rates are obtained directly from the 2020 American Community Survey [48]. ZIP codes that were not found in the case and feature datasets were excluded from the study. Data assembly and visualizations were done in RStudio Version 2023.12.1 (R Version 4.3.2). Maps were generated with the ggplot2 [49] package.

Table 1

Descriptive statistics for Florida ZIP codes. SD = standard deviation
Theme	Variable	Mean (Min-Max); SD
Theme 1: Socioeconomic Status	Percent of families at less than 125% of Federal Poverty Level¹	13.49 (0-100); 9.21
	Unemployment Rate ²	5.53 (0-84.7); 3.91
	Percent of population without health insurance¹	12.06 (0-82.81); 6.45
	Percent of renters who are burdened by rent costs¹	48.02 (0-100); 15.48
	Population Density (Residents/Mi-sq)*¹	2237.94 (0.07-44976.78); 3050.41
	Percent of adults aged 25 and older with No High School Diploma¹	6.82 (0-33.35); 4.68
Theme 2: Household Characteristics	Percent of population aged 65 years and older¹	22.93 (0-100); 12.2
	Percent of population aged 5–17 years¹	13.74 (0-44.59); 4.95
	Percent of non-institutionalized population with a disability (Mental or physical)¹	14.79 (0-43.01); 5.64
	Percent of all households who are single-parent families with children¹	11.84 (0-42.86); 6.86
	Percent of population aged 5 and older who speak English “less than ‘very well’” ¹	8.59 (0-60.61); 10.01
Theme 3: Racial and Ethnic Minority Status	Percent Minority (all persons except non-Hispanic whites)¹	39.61 (0-99.18); 24.83
Theme 4: Housing Type and Transportation	Percent of housing units in buildings with 3 or more units¹	22.16 (0-97.46); 21.76
	Percent of housing units that are mobile homes ¹	13.18 (0-87.16); 16.27
	Percent of renter-occupied housing units with more than 1.5 occupants per room¹	1.44 (0-39.29); 2.54
	Percentage of households with no vehicle available¹	5.47 (0-41.15); 4.68
	Percent of population living in group quarters (institutionalized and non-institutionalized)¹	3.4 (0-100); 9.97
Theme 5: Comorbidities	Percent of Obesity among adults aged 18 and older*¹	31.25 (0-48.5); 5.18
	Percent of diagnosed Diabetes among adults aged 18 and older *¹	11.81 (0-22.8); 2.92
	Percent of Asthma among adults aged 18 and older *¹	8.78 (0-14.5); 1.05
	Percent of Coronary Heart Disease (CHD) among adults aged 18 and older*¹	7.41 (0-20.7); 2.34
	Percent of High Blood Pressure among adults aged 18 and older*¹	34.79 (0-61.8); 6.72
	Percent of current smoking adults aged 18 and older*¹	18.74 (0–35); 4.9
*These variables are added due to their relevance in COVID-19 outcomes
¹Source: PolicyMap [47]
²Source: 2020 American Community Survey [48]
Poverty line: the minimum level of income needed to secure the necessities of life as determined by the Federal Poverty Level (FPL) guidelines.
Unemployment rate: the number of unemployed people as a percentage of the labor force.
Group quarters: All people not living in housing units. Institutionalized populations live in locations such as correctional facilities, nursing homes, mental hospitals. Non-institutionalized populations live in locations such as college dorms and military quarters.

2.3 Statistical analysis

2.3.1 Predicting COVID-19 growth rates using Random Forest

We constructed and used a RF model to predict the initial growth rates of COVID-19 at the ZIP code-level as a function of the 22 SVI variables. RF, developed by Leo Breiman in 2001 [50], is a supervised learning model approach to perform both regression and classification tasks in ML [51, 52]. It is an ensemble learning method that combines predictions from many decision trees, with adjustments for node size, number of trees, and sampled features [53]. We began by first splitting the dataset from all 935 ZIP codes into a training subset (80% of data points) and a test subset (20% of data points). To build a RF model that can accurately predict COVID-19 growth rate based on SVI features, a grid search was then conducted to identify optimal model parameters (i.e., number of trees, maximum tree depth) based on a 5-fold cross validation scheme using the training subset as input. In 5-fold cross-validation [54], the input training subset is further split into 5 folds to serve as training and test samples. A RF model constructed with a given set of parameter values is trained using four folds. Then, the trained model is used to make predictions on the remaining fifth test fold. The model’s predictive ability is quantified by Root-mean squared error (RMSE, the square root of the resulting value of the sum of the squared differences between predicted and observed values divided by the number of observations). This process is iterated five times but, for every iteration, one fold is used as test data and the rest as training data. Note that for every iteration, data in training and test fold changes, which adds to the effectiveness of this method. This k-fold cross validation is a well-suited method to identify how generalizable the model is to unseen data and to determine if the model is overfitting [55, 56]. The RMSEs of each fold are averaged to obtain the model’s cross-validated RMSE. This cross-validation process was further repeated for each unique set of model parameters. Finally, the model with the optimal set of parameters was identified as the model with the lowest cross-validated RMSE.

2.3.2 SHAP analysis

The SHAP framework was applied to the best-fitting RF model to explain the estimated association between COVID-19 growth rates and SVI features. SHAP, introduced by Lundberg and Lee in 2017 [40], utilizes game theory [57] principles to interpret the predictions of ML models, elucidating the contribution of input features to model outputs using Shapley values. A Shapely value is the average marginal contribution of a feature to the model predictions across all possible feature combinations [40]. SHAP values are computed by assessing how a model's predictions change with the inclusion or exclusion of features, and this is performed iteratively for each feature and sample in the dataset [40, 58]. In SHAP, predictions are calculated by adding the base value (i.e., the outcome that would be predicted if no features were used to train the model) and the SHAP values of each feature [40]. Therefore, features with positive SHAP values lead to an increase in the predictions, while those with negative values lead to a decrease in the predicted output.

The magnitude of the respective increased or decreased impact on predictions was further used to calculate the percentage contribution made by each feature. This was done by first dividing the SHAP value of a feature by the sum of SHAP values of all features for a single prediction followed by averaging the corresponding percentage contributions of the feature across all predictions. The RF/SHAP methods as described above were carried out using the scikit-learn (Version 1.4.1) [59] and shap (Version 0.45.0) [40] packages in Python.

2.3.3 Development of SVI

To develop the SVI for early COVID-19 spread in the Florida ZIP codes, the top features and their impact on prediction results were used as inputs. Features that contributed to at least > 7.5% of the RF predictions as obtained from the SHAP feature analysis (6 in total, see Additional File 2) were employed in this estimation. The SVI of a ZIP code was calculated as a weighted function of these variables using the following equation [60]:

$$\:SVI\:={\sum\:}_{i=1}^{k}{W}_{i}{X}_{i}$$

where,

W_i = weighted contribution of each of 6 most important features based on mean SHAP values

X_i = percentile score of each feature in a ZIP code

i = 1,2, …… k number of features (k = 6 in the present case)

The SVI score for each ZIP code is calculated and presented on a scale of 0–1, with 1 indicating high vulnerability and 0 indicating low vulnerability.

3.1 Descriptive Statistics

A total of 935 Florida ZIP codes were found to present in all the case and feature datasets, and therefore were included in the analysis. The maximum number of new cases was 416 cases reported by ZIP 33177 in Miami-Dade County. The average maximum number of daily cases was 34.95 cases. The mean, standard deviation, and range for each of the features included in the study are summarized in Table 1. The statistics indicate that sufficient information was available to investigate the ZIP code level associations between these features and initial epidemic spread in the state.

The initial growth rate, $\:r$, estimated using the log-linear regression model described above during the first COVID-19 wave that occurred in each Florida ZIP code is displayed in Fig. 2. The results indicate that the COVID-19 growth rate during this wave varied significantly between the Florida ZIP codes, with the estimated rates ranging from approximately 1 to as high as 1.247, with an average of 1.054. The largest growth rate (1.247) was found to occur in ZIP code 32321 in Liberty County. Furthermore, there appeared to be distinct clusters exhibiting high initial pandemic spread in the state with many ZIP codes in the middle, northwestern and northeastern and southernmost parts of the state in particular showing the highest levels of initial spread (Fig. 2).

3.2 RF results

The optimal parameters of the best-fitting RF model comprised a maximum tree depth of 7 and 350 estimator trees. Default values were acceptable for the rest of the parameters. The final trained model had a RMSE of 0.0245, the low value indicating that the differences between observed growth rates and those predicted by the RF model were acceptable and that the model is capable of predicting the growth rate for each ZIP code accurately. The features determined to be most important by the RF model for predicting COVID-19 growth rates for our time period were, in order: the percentage of the population speaking English “Less than ‘very well’”, the percentage of households with children that are headed by a single parent, Population Density, and the percent of the population living in group quarters.

3.3 SHAP results

We employed SHAP summary and violin plots to illustrate the order of global feature importance and their positive and negative influences on the output or predictions of the RF model (Fig. 3). The feature importance plot shows, quantitatively, how features contribute to the final model output, while the summary plot depicts the relationship between the value of a feature and its positive/negative influence or impact on predictions. The importance of the feature assigned by SHAP is the average of absolute Shapley values for that feature across the input dataset. The features are then plotted as bars of decreasing importance, providing a better understanding of the most important factors (Fig. 3(a)). Based on Shapely values and the percentage contribution to predictions (Additional File 2), the percent of households with children that are headed by a single parent was found to be the most important feature followed by population density, the percent of the population speaking English “Less than ‘very well’”, the percent of the population living in group quarters, the percentage of renters burdened by housing costs, and the percentage of the population in the ZIP code diagnosed with coronary heart disease (CHD). All of these factors had individual contributions of over 7.5% of the growth rate predictions, with percent single-parent households, population density and percent speaking English “Less than very well” each contributing to > 11%. It is also evident from the results that the SHAP analysis has generally led to making the same conclusions as those generated by the RF model importance function.

SHAP violin plots, illustrated in Fig. 3(b), combine feature importance with feature effects, and thus allow examination of the positive and negative impacts of each of the presently investigated SVI features on the COVID-19 growth rates observed during the early outbreak phase in the Florida ZIP codes. The results show that across the five highest ranking variables, higher values of these variables resulted in overall positive contributions to the predicted growth rate, while lower values had a negative impact. The largest absolute contributions are, however, negative produced at lower values of the top three features (percentage single-parent households, population density and percentage of population speaking English “Less than very well”) as can be seen by the blue data points to the left of the plot (Fig. 3(b)). Most of the data points on the other hand, particularly those of higher feature values, have positive contributions as showcased by the dense congregation of purple-to-pink points to the right of the plot. By contrast, levels of CHD diagnosis in the ZIP codes had an overall negative impact on growth rate predictions, indicating that complex associations between co-morbidities and COVID-19 infection rates may occur locally, at least during the initial growth phase of the pandemic. The rest of the variables beyond the sixth position, while still important by rank, had moderate to small effects on the predicted growth rate.

Figure 4 portrays the dependency of Shapely values on the distribution of each of the six most important features among the Florida ZIP codes. For all six features there is a sharp rise in SHAP values as the ZIP code feature values start to increase. For the percentage of single-parent households, SHAP values increase as the percentage of this variable increases and starts to plateau into positive contributions after approximately 5% of households are headed by single parents. In the case of the population density feature, after a sharper rise, SHAP values begin to decrease after a density of approximately 5000 persons per square mile to reach lower and negative values after 8000 people per square mile. The features pertaining to persons with a language barrier and persons living in group quarters exhibit a trend of SHAP values similar to that of the percentage of households headed by a single parent, but with sharper rises in SHAP values before plateauing. The SHAP values for these two features start to plateau very soon after crossing the threshold value of 0. As with the case for population density, SHAP values show a more complex and nonlinear association for the features corresponding to percentage burdened by housing costs and for CHD. For these features, SHAP values increase positively as feature values increase from 0 but following peaks they decline to become negative at the higher-most feature values, with this effect most dramatically observed for CHD (Fig. 4).

3.4 SVI

The top 6 features that contributed to > 7.5% of the prediction results were used to create a SVI score for quantifying the level of vulnerability exhibited by the Florida ZIP codes to early COVID-19 spread (Additional File 3). Validation of the developed SVI was carried out in three ways. First, in Fig. 5, we compare the SVI scores of each ZIP code with their corresponding initial pandemic growth rates. The significant positive association (linear regression: F = 97.96, df = 1,934, p-value < 0.0001) between the two variables indicates that the SVI developed here can indeed reliably identify the Florida ZIP codes most vulnerable to high initial COVID-19 spread. Second, Fig. 6 shows ZIP codes with high vulnerability (in the range 0.6 to 1) to initial spread of the pandemic (shown in darker shades), while those with low vulnerability (in the range 0-0.2) are shown in lighter shades. The ZIP code wide distribution of vulnerability to the initial growth of COVID-19 that was observed in Florida showed a clear spatial pattern with ZIP codes, revealing high SVIs occurring in the central, southeastern, northeastern, and some parts of the northwestern regions of the state. This spatial pattern overall corresponds well with the raw growth rate map as portrayed in Fig. 2. Finally, Table 2 lists 10 randomly selected ZIP codes of high and low SVI scores, along with the corresponding values for each of the top 6 features found in this study to be most important for underlying the initial growth of COVID-19. The results show that ZIP codes that exhibited high vulnerability to initial pandemic spread in general had higher levels of single-parent households, population density, and populations who live with language barriers, lived in group quarters and who are burdened by housing costs. By contrast, ZIP codes with lower SVI scores had populations presenting with generally higher levels of CHD (Table 2). This result matches the findings from the SHAP analyses above with respect to the features found to be important for predicting the ZIP code level initial pandemic growth that was observed in Florida.

Table 2

Randomly selected ZIP codes of high and low SVI alongside their respective values for the top 6 features and SVI.
ZIP Code	% Single-parent households	% Speaking English "Less than very well"	Population density	% in Group quarters	% Burdened by housing costs	Coronary Heart Disease	SVI
Low SVI
32134	4.97	0.95	26.43	0.16	32.11	11.3	0.03
32136	9.11	1.28	471.08	0	53.22	9.7	0.17
32190	7.97	4.88	47.76	0	28.57	9.2	0.10
32359	0	0	5.48	0	0	12.3	0.01
32430	4.85	0	8.9	0	0	9	0.01
32568	0.56	0	17.58	0	24.43	7.6	0.00
32643	6.75	2.02	103.43	0.09	45.77	7.5	0.08
32776	4.97	3.22	231.64	0	45.99	5.8	0.07
32976	2.29	1.2	383.27	0	48.07	14.7	0.06
33981	4.13	1.37	315.34	0.1	52.51	9.5	0.10
High SVI
32703	18.05	11.1	1776.57	0.99	59.81	5.9	0.82
32812	16.9	12.99	4206.4	0.33	59.69	5.8	0.87
33020	20.06	23.1	7463.83	1.98	61.71	6.2	0.98
33064	13.59	26.11	5878.98	1.11	59.92	7	0.92
33147	21	23.79	6523.37	0.58	56.55	7.5	0.96
33180	12.27	23.04	10080.61	0.99	63.73	6.5	0.91
33462	15.72	19.21	3766.9	1.03	59.63	7.7	0.92
33712	23.01	2.06	3935.56	1.45	63.87	6.6	0.86
33916	21.03	16.77	2543.83	0.95	53.24	6.6	0.89
34744	13.85	21.26	1834.04	0.89	73.25	5.6	0.82

The COVID-19 pandemic highlighted the need to gain a better understanding of the community attributes that govern a population’s susceptibility to epidemic transmission and spread [17, 21, 61]. A focus on population factors that govern early epidemic transmission is particularly key, as determining what makes some regions more vulnerable to pandemic transmission than others at this stage could enhance the development of institutional and societal preparedness measures to prevent or significantly mitigate the spread of future outbreaks [14]. In this research, we combined a RF model with a ML model agnostic explanative framework in conjunction with components of the CDC/ATDSR SVI data and number of COVID-19 cases that occurred during the early phase of the outbreak (defined as the period from reportage of first case to peak of the first wave) in order to determine the social vulnerability components that were most predictive of initial COVID-19 growth rate in Florida at the ZIP code-level. This work extends previous studies on this topic in several important ways. First, in the context of socioeconomic vulnerability and COVID-19 initial outbreaks, past studies have used cumulative cases [17, 22, 23] and incidence [21, 24, 25] as outcomes. While these studies have provided information on how SVI factors may impact the rise in early cases, they furnish less information on how these factors may affect the speed of epidemic growth [27]. Understanding the impact of vulnerability factors on the growth rate of an epidemic during its early phase on the other hand can convey information about the time scale and impact of disease spread, including how quickly the epidemic will grow, when it might peak, its overall magnitude, and how long it will last, all of which constitute important knowledge for guiding epidemic management [26, 28, 62]. Impact of SVI variables on COVID-19 growth rates have been analyzed, but at larger spatial scales and using a limited number of socioeconomic variables as predictors [32, 63]. In contrast, this study estimated and analyzed growth rates of the pandemic as a function of more than 20 socioeconomic variables at the fine spatial scale of ZIP codes (Table 1). Understanding the impact of these factors on the growth of an epidemic at this local spatial scale will allow a better appreciation of variations in the nature, scale, timing and speed of disease propagation across a region, which will be useful for the development and implementation of better defined and even more spatially-focused preparatory, control and prevention efforts. Given the variability in the way outbreak preparedness is operationalized across the US, the development of such analytical frameworks for assessing local-level variability to epidemic outbreaks and spread has been highlighted as a critical need for preparing for future epidemics [14]. Lastly, the use of ML-based covariate assessment methods using interpretative tools, such as SHAP, provides an enhancement via the ability of such frameworks for carrying out and explaining complex multidimensional data analysis [64, 65].

Our study first addressed the epidemic growth rates associated with the first wave of the COVID-19 pandemic observed in the ZIP codes of Florida. The results, obtained via fits of a simple log-linear regression model [27], indicate that this initial growth phase varied significantly between the ZIP code jurisdictions in Florida, with distinct hotspots of heightened transmission found along the central, southern and northern coastal and northwest regions of the state (Fig. 2). This immediately underscores the need to better understand the local vulnerability/resilience factors that underpinned the occurrence of such spatial variation, so that gaps in current preparatory measures can be addressed to prevent future impact of an outbreak on Florida communities. The RF model we trained to predict these growth rates using 22 unique ZIP code-level features, comprising socioeconomic, demographic, transportation and building type, and health-related factors, in this study was found to perform accurately to provide significant initial insights into this topic. The first finding of importance here was that our SHAP analysis identified six important factors from among the 22 factors investigated that was most associated with the observed early growth rate of the pandemic at the ZIP code level in Florida (Fig. 3, Table S1). The SHAP plots revealed that ZIP code features, viz. percentage single-parent households, population density and percentage of population speaking English less than fluently, percentage living in group quarters and percentage of the population burdened by housing costs, generally had a positive impact on epidemic growth prediction rates and thus making the ZIP codes with higher levels of these features vulnerable to heightened early transmission of the pandemic. By contrast and less intuitively, higher levels of CHD had an overall negative impact on epidemic growth predictions making ZIP codes with higher CHD levels less vulnerable to the early spread of the pandemic (see Fig. 4). The positive association between the first factors and incidence of COVID-19 cases has been reported previously at various spatial levels, from census tract to county and state levels [17, 18, 20, 66–69], albeit this is the first time that this positive relationship has also been documented for initial growth of the pandemic. Various explanations, ranging from the financial struggles faced by single parents [70–72], the impact of population density in enhancing contacts between susceptible and infected sub-groups in a community [68, 73–75], the overrepresentation of individuals less fluent in English in occupations deemed to be “essential” [10, 76, 77], their inability to fully understand safety guidelines, and propensity to live in multi-generational crowded homes [78, 79], and the unique exposure of individuals living in group quarters, including in nursing homes [80, 81] and correctional facilities [82, 83] have been proposed for this association and may similarly underlie the impacts of these factors in making the ZIP codes with high levels of these features also vulnerable to early heightened propagation of COVID-19.

The SHAP dependency plots shown in Fig. 4 provide additional insights regarding the relationships of the identified six most important factors or features and the RF predictions for early epidemic growth rates at the ZIP code-level. The results indicate a generally sharp positive increase in growth rate predictions as the percentage of single-parent households and percent of populations that have a language barrier or live in group quarters in a ZIP code increase. More complex functional associations were however identified for early epidemic growth predictions and increases with covariates, such as population density, percentage of population burdened by housing costs and percent of population in a ZIP code with CHD (Fig. 4). Indeed, the results show that at the highest levels of these variables at the ZIP code level, predictions for growth rates turned negative, with the most negative association at high covariate values found for CHD. Further, the non-linear form of the relationships indicates that the highest vulnerability to high epidemic growths during the early phase of the pandemic were exhibited by ZIP codes with intermediate levels of increase in the values of these covariates. While these unexpected findings suggest the need for further research, it is possible to speculate at least for population density and for the observed non-linear association of epidemic growth with CHD levels that the negative predictions at the highest levels for each of these variables may be linked to the fact that public services and possibly population compliance with protective measures (mask wearing, observing safe distances) may be higher in densely populated centers, such as major cities. Higher compliance by CHD patients to observing protective measures may similarly also underlie the negative association between high CHD levels and initial COVID-19 growth rates. These findings highlight how applying SHAP analysis in combination with ML modeling can facilitate the unearthing of novel insights into the impact of factors that drive epidemic transmission whilst also serving as a foundation for guiding future investigations.

The SVI metric developed and evaluated in this work using the top 6 important ZIP code-level features that contributed to at least > 7.5% of the RF model predictions revealed the Florida ZIP codes that were prone to high levels of COVID-19 transmission during the initial phase of the pandemic in the state. The SVI distribution showed a positive correlation with initial ZIP code-level epidemic growth indicating that the developed social vulnerability metric is able to reliably identify local ZIP code jurisdictions that were most sensitive to high initial outbreak conflation or those that were less susceptible to initial epidemic transmission. Moreover, it can be seen that such a metric can also be useful for assessing spatial variations in local vulnerability to the initial spread of an epidemic within a larger region, such as a state (Fig. 5). This finding underscores the importance of considering local-level community vulnerability to the spread of an epidemic in order to better understand the interplay between social determinants of outbreaks and development of more targeted mitigation efforts to curtail the emergence and spread of epidemics in a region [17, 20, 61, 67, 84]. In the latter regard, it is pertinent to note that our analysis has demonstrated that several specific community vulnerabilities, viz. percentage single-parent households, population density, percentage of population speaking English less than fluently, percentage living in group quarters and percentage of the population burdened by housing costs, were most salient in determining the rapid initial local spread of COVID-19 in Florida. Taken together, these results show how indexes such as the SVI can help not only to identify communities most susceptible to epidemic flare-ups but also by further identifying the specific vulnerability factors provide information that could be used to effectively allocate resources that could contain or mitigate epidemic spread across a region [11, 60]. These should include policy that protects and supports “essential” workers with workplace social distancing and mask-wearing mandates, access to diagnostic tests, health insurance, paid sick leave, supplementary income, and education regarding the disease’s transmission and protective measures in the native language of each worker [85–88]. Similarly, single parents with dependent children need a financial security net, which can be based on provision of supplementary income, paid sick leave, and affordable childcare, counseling services and health insurance [70–72]. Specific policy that addresses transmission in group quarters will also require to be developed going forward based on the nature of each type of quarter [83, 89–91]. This study indicates that focused policies guided by metrics such as the ML-based SVI developed here will help to minimize and slow early community spread which will be crucial to preventing the formation of the severe pandemics with major societal impacts as observed in the case of COVID-19.

Our analysis and findings must be interpreted considering the study’s limitations. Firstly, it is possible that the number of reported daily COVID-19 may have been understated, affecting the data on our outcome. If the number of cumulative cases was underreported differently by the socioeconomic factors included as our predictive features, our results might be biased. Furthermore, even though our independent variables comprised 22 distinct factors inspired by the 2020 CDC/ATSDR SVI themes, it is plausible that the incorporation of additional confounding factors, such as the number of persons using public transportation and movement of individuals between ZIP codes, would be beneficial. Such connectivity between ZIP codes may be particularly important as a ZIP code’s reported COVID-19 cases might be a function of epidemic transmission in neighboring communities. Our analysis also did not explicitly consider the intensity of the initial lockdowns that were implemented between April to early May and the actual compliance of each ZIP code’s population to such lockdowns owning to paucity of local-level data on these variables. ML-based methods can make better predictions with large amounts of training data, which our study did not have due to the spatial extent and focus of our study to considering only Florida. Therefore, our conclusions require to be supported by further investigations using more complete data and other types of analytical methodologies than used presently. Such research must also consider other regions of the United States if a more generalizable account of the vulnerability and factors underlying such vulnerability of local communities to epidemics across the country is to be realized.

Our findings show that coupling a RF model with interpretation of the output of the ML model using SHAP in conjunction with ZIP code-level early case and 2020 CDC/ATSDR SVI data can help assess the local vulnerability or resilience of communities to the initial spread of epidemics, such as COVID-19. Not only can this approach be used to identify local communities at high risk of early epidemic propagation but the determination of the specific community vulnerabilities most salient to early epidemic spread afforded by the combination ML-SHAP method can further be useful in guiding policy efforts that could target such vulnerabilities to improve local and global preparedness, community resilience and response to the arrival of the next novel pandemic-potential pathogens with infectious characteristics similar to COVID-19. The SVI to early epidemic spread generated here using the top 6 important ZIP code-level features based on the SHAP analysis indicate how our methods can be used to assess the level of community vulnerability, including identification of hot spots of high social vulnerability, to epidemic establishment and initial transmission across a region. Vulnerability evaluations based on such SVI metrics could thus be used to advance mitigation strategies that consider multiple factors affecting sub-populations in different regions whilst also focusing specifically on measures that can best contain and dampen spatial spread from hot spot localities. Finally, while this research has highlighted the usefulness of ML methods for assessing the social vulnerability of communities to initial spread of epidemics, additional work in other regions than Florida will be required to validate and improve the methods and outcomes of our study. A key need here will be the provision of the relevant local case and features data needed in order to carry out such evaluations.

ATSDR

Agency for Toxic Substances and Disease Registry

CDC

Centers for Disease Control and Prevention

CHD

Coronary Heart Disease

Machine Learning

Random Forest

Standard Deviation

SHAP

SHapley Additive exPlanations

SVI

Social Vulnerability Index

Acknowledgements: We acknowledge and thank Dr. Kenneth Newcomb and Dr. Shakir Bilal for assistance in writing Python code for extracting ZIP code-level case data.

Author’s contributions: EM, YA, and WZ designed the study. YA, RG and WZ acquired the feature data used in the study and performed the analysis. YA and EM prepared the manuscript. EM supervised the study. All authors reviewed and approved the manuscript.

Availability of data and materials: Data generated during this study are included in this published article and its supplementary information files. Feature data are not publicly available due to restricted use by sources, but are available through PolicyMap (www.policymap.com) [47] as accessed by the authors in this article. Unemployment rate data is publicly available through the U.S. Census Bureau [48].

Funding: Part of the funding for this work was obtained from a COVID-19 grant awarded by Hillsborough County, Florida (BOCC No: 22-0680).

Competing interests: The authors have no conflicts of interest to declare.

Ethics approval and consent to participate: This study was a secondary analysis based on data aggregated at the ZIP code-level. No human participants or experimental animals were directly involved in the study. Therefore, ethics approval was not required for this project.

Williams BA, Jones CH, Welch V, True JM. Outlook of pandemic preparedness in a post-COVID-19 world. Npj Vaccines. 2023. 10.1038/s41541-023-00773-0.
International Monetary Fund. World Economic Outlook Update: A Crisis Like No Other, An Uncertain Recovery. In: World Economic Outlook. 2020. https://www.imf.org/en/Publications/WEO/Issues/2020/06/24/WEOUpdateJune2020. Accessed 18 Jun 2024.
Badr HS, Du HR, Marshall M, Dong ES, Squire MM, Gardner LM. Association between mobility patterns and COVID-19 transmission in the USA: a mathematical modelling study. Lancet Infect Dis. 2020. 10.1016/S1473-3099(20)30553-3.
Bambra C, Riordan R, Ford J, Matthews F. The COVID-19 pandemic and health inequalities. J Epidemiol Commun H. 2020. 10.1136/jech-2020-214401.
Stokes EK, Zambrano LD, Anderson KN, Marcler EP, Raz KM, Felix SE, et al. Coronavirus Disease 2019 Case Surveillance - United States, January 22-May 30, 2020. Mmwr-Morbid Mortal W. 2020. 10.15585/mmwr.mm6924e2.
Chung RYN, Dong D, Li MNM. Socioeconomic gradient in health and the COVID-19 outbreak. BMJ. 2020. 10.1136/bmj.m1329.
Magesh S, John D, Li WT. Disparities in COVID-19 Outcomes by Race, Ethnicity, and Socioeconomic Status: A Systematic Review and Meta-analysis. Jama Netw Open. 2021. 10.1001/jamanetworkopen.2021.44237.
Agyemang C, Richters A, Jolani S, Hendriks S, Zalpuri S, Yu E, et al. Ethnic minority status as social determinant for COVID-19 infection, hospitalisation, severity, ICU admission and deaths in the early phase of the pandemic: a meta-analysis. Bmj Glob Health. 2021. 10.1136/bmjgh-2021-007433.
Mude W, Oguoma VM, Nyanhanda T, Mwanri L, Njue C. Racial disparities in COVID-19 pandemic cases, hospitalisations, and deaths: A systematic review and meta-analysis. J Glob Health. 2021. 10.7189/jogh.11.05015.
Do DP, Frank R. Unequal burdens: assessing the determinants of elevated COVID-19 case and death rates in New York City's racial/ethnic minority neighbourhoods. J Epidemiol Commun H. 2021. 10.1136/jech-2020-215280.
Flanagan BE, Gregory EW, Hallisey EJ, Heitgerd JL, Lewis B. A Social Vulnerability Index for Disaster Management. J Homel Secur Emerg. 2011. 10.2202/1547-7355.1792.
Flanagan BE, Hallisey EJ, Adams E, Lavery A. Measuring Community Vulnerability to Natural and Anthropogenic Hazards: The Centers for Disease Control and Prevention's Social Vulnerability Index. J Environ Health. 2018;80(10):34–6.
Rogers TN, Rogers CR, VanSant-Webb E, Gu LY, Yan B, Qeadan F. Racial Disparities in COVID-19 Mortality Among Essential Workers in the United States. World Med Health Pol. 2020. 10.1002/wmh3.358.
Rogers CJ, Cutler B, Bhamidipati K, Ghosh JK. Preparing for the next outbreak: A review of indices measuring outbreak preparedness, vulnerability, and resilience. Prev Med Rep. 2023. 10.1016/j.pmedr.2023.102282.
Healthy People 2030, U.S. Department of Health and Human Services, Office of Disease Prevention and Health Promotion. Social Determinants of Health. https://health.gov/healthypeople/priority-areas/social-determinants-health. Accessed 15 Jul 2024.
Biggs EN, Maloney PM, Rung AL, Peters ES, Robinson WT. The Relationship Between Social Vulnerability and COVID-19 Incidence Among Louisiana Census Tracts. Front Public Health. 2021. 10.3389/fpubh.2020.617976.
Tatar M, Faraji MR, Wilson FA. Social vulnerability and initial COVID-19 community spread in the US South: a machine learning approach. Bmj Health Care Info. 2023. 10.1136/bmjhci-2022-100703.
Oates GR, Juarez LD, Horswell R, Chu S, Miele L, Fouad MN, et al. The Association Between Neighborhood Social Vulnerability and COVID-19 Testing, Positivity, and Incidence in Alabama and Louisiana. J Commun Health. 2021. 10.1007/s10900-021-00998-x.
Karmakar M, Lantz PM, Tipirneni R. Association of Social and Demographic Factors With COVID-19 Incidence and Death Rates in the US. Jama Netw Open. 2021. 10.1001/jamanetworkopen.2020.36462.
Tortolero GA, Otto MD, Ramphul R, Yamal JM, Rector A, Brown M, et al. Examining Social Vulnerability and the Association With COVID-19 Incidence in Harris County, Texas. Front Public Health. 2022. 10.3389/fpubh.2021.798085.
Mukherji N. The Social and Economic Factors Underlying the Incidence of COVID-19 Cases and Deaths in US Counties During the Initial Outbreak Phase. Rev Reg Stud. 2022; 52(1):127 – 50. ISSN: 0048-749x.
Gorris ME, Shelley CD, Valle SYD, Manore CA. A time-varying vulnerability index for COVID-19 in New Mexico, USA using generalized propensity scores. Health Policy Open. 2021. 10.1016/j.hpopen.2021.100052.
Jackson SL, Derakhshan S, Blackwood L, Lee L, Huang Q, Habets M, et al. Spatial Disparities of COVID-19 Cases and Fatalities in United States Counties. Int J Env Res Pub He. 2021. 10.3390/ijerph18168259.
Laefer DF, Protopapas D. Essential Business Visits and Social Vulnerability during New York City's Initial COVID-19 Outbreak. Epidemiologia-Basel. 2022. 10.3390/epidemiologia3040039.
Mehta M, Julaiti J, Griffin P, Kumara S. Early Stage Machine Learning-Based Prediction of US County Vulnerability to the COVID-19 Pandemic: Machine Learning Approach. Jmir Public Hlth Sur. 2020. 10.2196/19446.
Polonsky JA, Baidjoe A, Kamvar ZN, Cori A, Durski K, Edmunds WJ, et al. Outbreak analytics: a developing data science for informing the response to emerging pathogens. Philos T R Soc B. 2019. 10.1098/rstb.2018.0276.
Ma JL, Dushoff J, Bolker BM, Earn DJD. Estimating Initial Epidemic Growth Rates. B Math Biol. 2014. 10.1007/s11538-013-9918-2.
Newcomb K, Bilal S, Michael E. Combining predictive models with future change scenarios can produce credible forecasts of COVID-19 futures. PLoS ONE. 2022. 10.1371/journal.pone.0277521.
Centers for Disease Control and Prevention, Agency for Toxic Substances and Disease Registry. CDC/ATSDR SVI Documentation 2020. In: Place and Health. 2022. https://www.atsdr.cdc.gov/placeandhealth/svi/documentation/SVI_documentation_2020.html. Accessed 20 Nov 2023.
Freese KE, Vega A, Lawrence JJ, Documet PI. Social Vulnerability Is Associated with Risk of COVID-19 Related Mortality in US Counties with Confirmed Cases. J Health Care Poor U. 2021. 10.1353/hpu.2021.0022.
Khazanchi R, Beiter ER, Gondi S, Beckman AL, Bilinski A, Ganguli I. County-Level Association of Social Vulnerability with COVID-19 Cases and Deaths in the USA. J Gen Intern Med. 2020. 10.1007/s11606-020-05882-3.
Clouston SAP, Natale G, Link BG. Socioeconomic inequalities in the spread of coronavirus-19 in the United States: A examination of the emergence of social inequalities. Soc Sci Med. 2021. 10.1016/j.socscimed.2020.113554.
Zollanvari A. Machine Learning With Python: Theory and Implementation. Switzerland AG: Springer Cham; 2023.
Lucas B, Vahedi B, Karimzadeh M. A spatiotemporal machine learning approach to forecasting COVID-19 incidence at the county level in the USA. Int J Data Sci Anal. 2023. 10.1007/s41060-021-00295-9.
Ak C, Chitsazan AD, Gönen M, Etzioni R, Grossberg AJ. Spatial Prediction of COVID-19 Pandemic Dynamics in the United States. ISPRS Int J Geo-Inf. 2022. 10.3390/ijgi11090470.
Ward OM, Johnsen ALAE, Ng TL, Chollet RO. Forecasting SARS-CoV-2 transmission and clinical risk at small spatial scales by the application of machine learning architectures to syndromic surveillance data. Nat Mach Intell. 2022. 10.1038/s42256-022-00538-9.
Tiwari A, Dadhania AV, Ragunathrao VAB, Oliveira ERA. Using machine learning to develop a novel COVID-19 Vulnerability Index(C19VI). Sci Total Environ. 2021. 10.1016/j.scitotenv.2021.145650.
Moosazadeh M, Ifaei P, Charmchi AST, Asadi S, Yoo C. A machine learning-driven spatio-temporal vulnerability appraisal based on socio-economic data for COVID-19 impact prevention in the US counties. Sustain Cities Soc. 2022. 10.1016/j.scs.2022.103990.
Itzhak N, Shahar T, Moskovich R, Shahar Y. The Impact of US County-Level Factors on COVID-19 Morbidity and Mortality. J Urban Health. 2022. 10.1007/s11524-021-00601-7.
Lundberg SM, Lee S-I et al. A Unified Approach to Interpreting Model Predictions. In: Guyon I, Luxburg UV, Bengio S, Wallach H, Fergus R, Vishwanathan S, editors. Advances in Neural Information Processing Systems 30: Curran Associates, Inc.; 2017. pp. 4765-74.
EXECUTIVE ORDER NUMBER 20–91: Essential Services and Activities During COVID-19 Emergency. State of Florida Office of the Governor. 2020. https://www.flgov.com/2020-executive-orders/
EXECUTIVE ORDER NUMBER 20–112. Phase 1: Safe. Smart. Step-by-Step. Plan for Florida's Recovery. State Fla Office Gov. 2020. https://www.flgov.com/2020-executive-orders/
Murray E, Taylor L. How coronavirus is spreading in Florida. Tampa Bay Times. 2020. https://projects.tampabay.com/projects/data/coronavirus/en/zipcode/. Accessed 1 Feb 2024.
Villalobos NVF, Ott JJ, Klett-Tammen CJ, Bockey A, Vanella P, Krause G, et al. Effect modification of the association between comorbidities and severe course of COVID-19 disease by age of study participants: a systematic review and meta-analysis. Syst Rev-London. 2021. 10.1186/s13643-021-01732-3.
Thakur B, Dubey P, Benitez J, Torres JP, Reddy S, Shokar N, et al. A systematic review and meta-analysis of geographic differences in comorbidities and associated severity and mortality among individuals with COVID-19. Sci Rep-Uk. 2021. 10.1038/s41598-021-88130-w.
Mattey-Mora PP, Begle CA, Owusu CK, Chen C, Parker MA. Hospitalised versus outpatient COVID-19 patients' background characteristics and comorbidities: A systematic review and meta-analysis. Rev Med Virol. 2022. 10.1002/rmv.2306.
PolicyMap. PolicyMap. www.policymap.org. Accessed 16 Feb 2024.
Census Bureau US. Table S2301: EMPLOYMENT STATUS. American Community Survey, ACS 5-Year Estimates Subject Tables. 2020. https://data.census.gov/table/ACSST5Y2020.S2301?q=employmentstatus&g=040XX00US12$8600000&y=2020
Wickham H. ggplot2: Elegant Graphics for Data Analysis. Springer Cham; 2016.
Breiman L. Random forests. Mach Learn. 2001. 10.1023/A:1010933404324.
Talekar B, Agrawal S. A Detailed Review on Decision Tree and Random Forest. Biosci Biotech Res C. 2020. 10.21786/bbrc/13.14/57.
Roy MH, Larocque D. Robustness of random forests for regression. J Nonparametr Stat. 2012. 10.1080/10485252.2012.715161.
Biau G. Analysis of a Random Forests Model. J Mach Learn Res. 2012;13:1063–95.
Geisser S. The Predictive Sample Reuse Method with Applications. J Am Stat Assoc. 1975. 10.1080/01621459.1975.10479865.
Cawley GC, Talbot NLC. On Over-fitting in Model Selection and Subsequent Selection Bias in Performance Evaluation. J Mach Learn Res. 2010;11:2079–107.
Gitte V, Hendrik B. Look before you leap: Some insights into learner evaluation with cross-validation. Proceedings of the Workshop on Statistically Sound Data Mining at ECML/PKDD; 2015/11/27; Nancy, France: PMLR; 2015. pp. 3–20.
Langsetmo L, Schousboe JT, Taylor BC, Cauley JA, Fink HA, Cawthon PM, et al. Advantages and Disadvantages of Random Forest Models for Prediction of Hip Fracture Risk Versus Mortality Risk in the Oldest Old. Jbmr Plus. 2023. 10.1002/jbm4.10757.
Strumbelj E, Kononenko I. Explaining prediction models and individual predictions with feature contributions. Knowl Inf Syst. 2014. 10.1007/s10115-013-0679-x.
Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, et al. Scikit-learn: Machine Learning in Python. J Mach Learn Res. 2011;12:2825–30.
Kumar V, Sznajder KK, Kumara S. Machine learning based suicide prediction and development of suicide vulnerability index for US counties. NPJ Mental Health Res. 2022. 10.1038/s44184-022-00002-x.
Dasgupta S, Bowen VB, Leidner A, Fletcher K, Musial T, Rose C, et al. Association Between Social Vulnerability and a County's Risk for Becoming a COVID-19 Hotspot - United States, June 1-July 25, 2020. Mmwr-Morbid Mortal W. 2020. 10.15585/mmwr.mm6942a3.
Shearer FM, Moss R, McVernon J, Ross JV, McCaw JM. Infectious disease pandemic planning and response: Incorporating decision analysis. Plos Med. 2020. 10.1371/journal.pmed.1003018.
Mishra S, Dash TK, Panda G. Effect of Socio-economic and Environmental Factors on the Growth Rate of COVID-19 with an Overview of Speech Data for Its Early Diagnosis. In: Dash S, Pani SK, Rodrigues J, Majhi B, editors. Deep Learning, Machine Learning and IoT in Biomedical and Health Informatics: Techniques and Applications. Biomedical Engineering: Techniques and Applications. 1st Edition ed. Boca Raton, FL: CRC Press; 2022.
Linardatos P, Papastefanopoulos V, Kotsiantis S, Explainable AI. A Review of Machine Learning Interpretability Methods. Entropy-Switz. 2021. 10.3390/e23010018.
Tursunalieva A, Alexander DLJ, Dunne R, Li JM, Riera L, Zhao YC, et al. Making Sense of Machine Learning: A Review of Interpretation Techniques and Their Applications. Appl Sci-Basel. 2024. 10.3390/app14020496.
Bauer C, Zhang KH, Lee M, Fisher-Hoch S, Guajardo E, McCormick J, et al. Census Tract Patterns and Contextual Social Determinants of Health Associated With COVID-19 in a Hispanic Population From South Texas: A Spatiotemporal Perspective. Jmir Public Hlth Sur. 2021. 10.2196/29205.
Rocklöv J, Sjödin H. High population densities catalyse the spread of COVID-19. J Travel Med. 2020. 10.1093/jtm/taaa038.
Jamal Y, Gangwar M, Usmani M, Adams A, Wu CY, Nguyen TH, et al. Identification of Thresholds on Population Density for Understanding Transmission of COVID-19. Geohealth. 2022. 10.1029/2021GH000449.
Arcaya MC, Nidam Y, Binet A, Gibson R, Gavin V. Rising home values and Covid-19 case rates in Massachusetts. Soc Sci Med. 2020. 10.1016/j.socscimed.2020.113290.
Radey M, Langenderfer-Magruder L, Speights JB. I don't have much of a choice: Low-income single mothers' COVID-19 school and care decisions. Fam Relat. 2021. 10.1111/fare.12593.
Parolin Z, Lee EK. Economic Precarity among Single Parents in the United States during the COVID-19 Pandemic. Ann Am Acad Polit Ss. 2022. 10.1177/00027162221122682.
Alonzi S, Park JE, Pagán A, Saulsman C, Silverstein MW. An Examination of COVID-19-Related Stressors among Parents. Eur J Invest Health. 2021. 10.3390/ejihpe11030061.
Lima FT, Brown NC, Duarte JP. Understanding the Impact of Walkability, Population Density, and Population Size on COVID-19 Spread: A Pilot Study of the Early Contagion in the United States. Entropy-Switz. 2021. 10.3390/e23111512.
McLaughlin JM, Khan F, Pugh S, Angulo FJ, Schmitt HJ, Isturiz RE, et al. County-level Predictors of Coronavirus Disease 2019 (COVID-19) Cases and Deaths in the United States: What Happened, and Where Do We Go from Here? Clin Infect Dis. 2021. 10.1093/cid/ciaa1729.
Wong DWS, Li Y. Spreading of COVID-19: Density matters. PLoS ONE. 2020. 10.1371/journal.pone.0242398.
Kerwin D, Nicholson M, Alulema D, Warren R. US Foreign-Born Essential Workers by Status and State, and the Global Pandemic. Center for Migration Studies. 2020. https://cmsny.org/publications/us-essential-workers/. Accessed 15 Jul 2024.
Rho HJ, Brown H, Fremstad S, Center for Economic and Policy Research. A Basic Demographic Profile of Workers in Frontline Industries. 2020. https://cepr.net/a-basic-demographic-profile-of-workers-in-frontline-industries/. Accessed 15 Jul 2024.
Swendener A, Rydberg K, Tuttle M, Yam H, Henning-Smith C. Crowded Housing and Housing Cost Burden by Disability, Race, Ethnicity, and Rural-Urban Location. University of Minnesota Rural Health Research Center. 2023. https://rhrc.umn.edu/publication/crowded-housing-and-housing-cost-burden-by-disability-race-ethnicity-and-rural-urban-location/
Cohn D, Menasce Horowitz J, Minkin R, Fry R, Hurst K. The demographics of multigenerational households. In: Financial Issues Top The List of Reasons US Adults Live in Multigenerational Homes. Pew Research Center. 2022. https://www.pewresearch.org/social-trends/2022/03/24/the-demographics-of-multigenerational-households/. Accessed 15 Jul 2024.
McGarry BE, Grabowski DC. Nursing Homes and COVID-19: A Crisis on Top of a Crisis. ANNALS Am Acad Political Social Sci. 2021. 10.1177/00027162211061509.
Barnett ML, Grabowski DC. Nursing Homes Are Ground Zero for COVID-19 Pandemic. JAMA Health Forum. 2020. 10.1001/jamahealthforum.2020.0369.
Schnepel KT. COVID-19 in U.S. State and Federal Prisons: December 2020 Update. Washington, D.C.: Council on Criminal Justice; 2020.
Rubin R. The Challenge of Preventing COVID-19 Spread in Correctional Facilities. Jama-J Am Med Assoc. 2020. 10.1001/jama.2020.5427.
Karaye IM, Horney JA. The Impact of Social Vulnerability on COVID-19 in the US: An Analysis of Spatially Varying Relationships. Am J Prev Med. 2020. 10.1016/j.amepre.2020.06.006.
Waltenburg MA, Victoroff T, Rose CE, Butterfield M, Jervis RH, Fedak KM et al. Update: COVID-19 Among Workers in Meat and Poultry Processing Facilities - United States, April-May 2020. Morbidity and Mortality Weekly Report. 2020; 10.15585/mmwr.mm6927e2
Xiuhtecutli N, Shattuck A. Crisis politics and US farm labor: health justice and Florida farmworkers amid a pandemic. J Peasant Stud. 2021. 10.1080/03066150.2020.1856089.
Chicas R, Xiuhtecutli N, Houser M, Glastra S, Elon L, Sands JM, et al. COVID-19 and Agricultural Workers: A Descriptive Study. J Immigr Minor Healt. 2022. 10.1007/s10903-021-01290-9.
Akil L, Barner YM, Bisht A, Okoye E, Ahmad HA. COVID-19 Incidence and Death Rates in the Southern Region of the United States: A Racial and Ethnic Association. Int J Env Res Pub He. 2022. 10.3390/ijerph192113990.
Klann EM, Rich SN, Hendrick SC, Felipe A, Sathish S, Prins C, et al. Multi-Faceted Approach to COVID-19 Surveillance at a Large University Campus in the Southeastern United States. Disaster Med Public. 2022. 10.1017/dmp.2022.109.
Ouslander JG, Grabowski DC. COVID-19 in Nursing Homes: Calming the Perfect Storm. J Am Geriatr Soc. 2020. 10.1111/jgs.16784.
Henry BF. Social Distancing and Incarceration: Policy and Management Strategies to Reduce COVID-19 Transmission and Promote Health Equity Through Decarceration. Health Educ Behav. 2020. 10.1177/1090198120927318.

No competing interests reported.

Download PDF

Reviewers invited by journal
08 Oct, 2024
Editor invited by journal
17 Aug, 2024
Editor assigned by journal
16 Aug, 2024
Submission checks completed at journal
16 Aug, 2024
First submitted to journal
13 Aug, 2024

You are reading this latest preprint version

Assessing community vulnerability to initial COVID-19 spread in Florida ZIP Codes using Shapley additive explanations with random forest modeling

Status:

Version 1

Abstract

Background

Methods

Results

Conclusions

Figures

1 BACKGROUND

2 METHODS

2.1 Study setting and overview

2.2 Data assembly

2.2.1 Growth rate

2.2.2 Socioeconomic vulnerability variables

2.3 Statistical analysis

2.3.1 Predicting COVID-19 growth rates using Random Forest

2.3.2 SHAP analysis

2.3.3 Development of SVI

3 RESULTS

3.1 Descriptive Statistics

3.2 RF results

3.3 SHAP results

3.4 SVI

4 DISCUSSION

5 CONCLUSIONS

Abbreviations

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1