This paper is in the following e-collection/theme issue:

Descriptive Epidemiology and Population Size Estimates (109) JMIR Theme Issue: COVID-19 Special Issue (2461) Longitudinal and Cohort Studies in Public Health (316) Machine Learning (1501) Social Determinants of Health (36)

Published on in Vol 10 (2024)

Preprints (earlier versions) of this paper are available at https://preprints.jmir.org/preprint/54421, first published .
Identifying Potential Factors Associated With Racial Disparities in COVID-19 Outcomes: Retrospective Cohort Study Using Machine Learning on Real-World Data

Identifying Potential Factors Associated With Racial Disparities in COVID-19 Outcomes: Retrospective Cohort Study Using Machine Learning on Real-World Data

Identifying Potential Factors Associated With Racial Disparities in COVID-19 Outcomes: Retrospective Cohort Study Using Machine Learning on Real-World Data

Authors of this article:

Osama Dasa1, 2 Author Orcid Image ;   Chen Bai3 Author Orcid Image ;   Ruba Sajdeya1 Author Orcid Image ;   Stephen E Kimmel1 Author Orcid Image ;   Carl J Pepine2 Author Orcid Image ;   Matthew J Gurka J4 Author Orcid Image ;   Reinhard Laubenbacher5 Author Orcid Image ;   Thomas A Pearson1 Author Orcid Image ;   Mamoun T Mardini3 Author Orcid Image
Article Authors Cited by Tweetations (2) Metrics

    Original Paper

    1Department of Epidemiology, College of Public Health and Health Professions and College of Medicine, University of Florida, Gainesville, FL, United States

    2Division of Cardiovascular Medicine, Department of Medicine, University of Florida, Gainesville, FL, United States

    3Department of Health Outcomes and Biomedical Informatics, University of Florida, Gainesville, FL, United States

    4Department of Public Health Sciences, School of Medicine, University of Virginia, Charlottesville, VA, United States

    5Laboratory for Systems Medicine, Division of Pulmonary, Critical Care, and Sleep Medicine, Department of Medicine, University of Florida, Gainesville, FL, United States

    Corresponding Author:

    Osama Dasa, MD, MPH, PhD

    Department of Epidemiology

    College of Public Health and Health Professions and College of Medicine

    University of Florida

    PO Box 100238

    Gainesville, FL, 32610

    United States

    Phone: 1 3527331234

    Fax:1 3522650379

    Email: osama.dasa@gmail.com


    Background: Racial disparities in COVID-19 incidence and outcomes have been widely reported. Non-Hispanic Black patients endured worse outcomes disproportionately compared with non-Hispanic White patients, but the epidemiological basis for these observations was complex and multifaceted.

    Objective: This study aimed to elucidate the potential reasons behind the worse outcomes of COVID-19 experienced by non-Hispanic Black patients compared with non-Hispanic White patients and how these variables interact using an explainable machine learning approach.

    Methods: In this retrospective cohort study, we examined 28,943 laboratory-confirmed COVID-19 cases from the OneFlorida Research Consortium’s data trust of health care recipients in Florida through April 28, 2021. We assessed the prevalence of pre-existing comorbid conditions, geo-socioeconomic factors, and health outcomes in the structured electronic health records of COVID-19 cases. The primary outcome was a composite of hospitalization, intensive care unit admission, and mortality at index admission. We developed and validated a machine learning model using Extreme Gradient Boosting to evaluate predictors of worse outcomes of COVID-19 and rank them by importance.

    Results: Compared to non-Hispanic White patients, non-Hispanic Blacks patients were younger, more likely to be uninsured, had a higher prevalence of emergency department and inpatient visits, and were in regions with higher area deprivation index rankings and pollutant concentrations. Non-Hispanic Black patients had the highest burden of comorbidities and rates of the primary outcome. Age was a key predictor in all models, ranking highest in non-Hispanic White patients. However, for non-Hispanic Black patients, congestive heart failure was a primary predictor. Other variables, such as food environment measures and air pollution indicators, also ranked high. By consolidating comorbidities into the Elixhauser Comorbidity Index, this became the top predictor, providing a comprehensive risk measure.

    Conclusions: The study reveals that individual and geo-socioeconomic factors significantly influence the outcomes of COVID-19. It also highlights varying risk profiles among different racial groups. While these findings suggest potential disparities, further causal inference and statistical testing are needed to fully substantiate these observations. Recognizing these relationships is vital for creating effective, tailored interventions that reduce disparities and enhance health outcomes across all racial and socioeconomic groups.

    JMIR Public Health Surveill 2024;10:e54421

    doi:10.2196/54421

    Keywords

    health disparities; racial disparities; COVID-19 outcomes; social determinants of health; area deprivation index; health outcomes; machine learning; real-world data; COVID-19; SARS-CoV-2; real-world data; socioeconomic status 


    The COVID-19 pandemic has exposed significant racial disparities, with non-Hispanic Black populations experiencing higher incidence rates and worse outcomes than non-Hispanic White populations [1,2]. Previous research has suggested that these disparities might be attributed to a higher burden of comorbidities in non-Hispanic Black populations [3]. Additional hypothesized contributing factors include systemic socioeconomic disadvantages and the compounding impact of chronic exposure to social and economic stressors, as well as experiences of marginalization [4,5]. Furthermore, some studies have raised the possibility of biological and genetic determinants as contributory risk factors [6,7].

    Analyzing racial health disparities solely through isolated data can lead to misleading interpretations, such as attributing higher disease rates to inherent biological factors or perpetuating racial stereotypes, particularly concerning perceived health behaviors [4]. Similarly, dissecting COVID-19 data by geography demands caution to avoid reinforcing negative stereotypes about marginalized communities and “territorial stigmatization” [8]. In the context of the COVID-19 pandemic, a comprehensive framework is necessary that considers social determinants of health (SDOH)—social, economic, and geographical—to properly address the complexities surrounding disparities in outcomes [9-11]. Gathering data at both the neighborhood and individual level is crucial. Such an approach dispels myths rooted in racial biology and highlights the significance of socioeconomic factors, discrimination, and location-based risks [4,5].

    It has been recognized that the COVID-19 pandemic is a syndemic involving interactions of multiple factors and conditions [12] where advanced data-driven approaches are needed to capture complex underlying association patterns. Machine learning (ML) offers valuable tools for this complex analysis. Unlike traditional methods, ML can handle vast multidimensional datasets, enabling researchers to discern intricate patterns and relationships that would be impossible to discern using conventional statistical methods. Although many studies have used ML in evaluating the outcomes of COVID-19 [13-18], few have applied these techniques specifically to investigate racial disparities [9,19], which often result from complex interplay among multiple stressors. Most studies focused on a single stressor or a single domain of stressors or were purely ecological, with only county-level COVID-19 data [20].

    The purpose of this study is to provide a more comprehensive understanding of the factors contributing to these disparities by analyzing large real-world data from the OneFlorida Research Consortium, using 9-digit zip codes to link individuals to measures of SDOH and other geospatial data related to the place of residence. This will give insight into the importance or relevance of each variable (feature) in predicting worse outcomes of COVID-19. By incorporating comorbidities and SDOH, we seek to explore potential reasons behind the worse outcomes of COVID-19 experienced by non-Hispanic Black patients compared with non-Hispanic White patients and how these variables interact in an ML model.


    Design and Population

    We conducted a retrospective cohort study of patients diagnosed with COVID-19 infection who were followed prospectively for outcomes. We derived patient-level data from the OneFlorida Clinical Research Network, which includes Floridians enrolled in Medicaid, and robust patient-level electronic health record (EHR) data from public and private health care systems [21]. OneFlorida partners encompass hospitals, clinic settings, and physicians, which provide care for 17 million patients across all of Florida’s 67 counties [21,22]. The data query included adult patients (≥18 years old) with laboratory-confirmed COVID-19 diagnoses between December 1, 2019, and April 28, 2021. Then, using 9-digit zip codes, we linked patients to neighborhood and geospatial variables, as established by previous research [20]. These supplementary variables were derived from several external publicly available datasets, enabling a more comprehensive understanding of the impact of socioeconomic and geospatial factors on outcomes of COVID-19.

    Study Variables (Features)

    Exposure

    We used self-reported race as our exposure variable. It is a surrogate for many interconnected socioeconomic and environmental factors. Race is recognized as a social construct, not a biological determinant [23-25].

    Outcome

    We used a combined measure of severe outcomes of COVID-19. This measure includes intensive care unit admission, intubation, and mortality at the time of the first index encounter. Through this composite outcome, we intended to capture a broader overview of the disease’s severity and short-term impacts. We used encounter, procedures, and death common data model domains in OneFlorida to assess outcomes [26].

    Covariates

    Individual-Level Variables

    At the index encounter, we collected sociodemographic data, including age, self-reported sex, race and ethnicity as non-Hispanic Black or non-Hispanic White, and previous insurance status. We also included 9-digit zip codes, BMI, smoking, alcohol, and substance use status.

    In line with the Centers for Disease Control and Prevention’s compilation of medical conditions associated with an increased risk for severe outcomes of COVID-19 [27], we examined comorbidities during the encounters from January 1, 2012, leading up to the COVID-19 index encounter. Conditions with a less than 1% prevalence in our study population were excluded. We derived COVID-19–related comorbid conditions by mapping available International Classification of Diseases, Ninth Revision and International Statistical Classification of Diseases, Tenth Revision codes to the Healthcare Cost and Utilization Project (HCUP) Clinical Classification Software definitions [28]. We also used the well-established Elixhauser Combined Comorbidity Index (ECI) to quantify the aggregate burden of comorbid conditions. This index includes 31 comorbidities and reliably predicts outcomes like in-hospital mortality, length of stay, adverse events, and hospital discharges [29-32].

    Group-Level Variables
    Area Deprivation Index

    The area deprivation index (ADI) evaluates community deprivation, impacts health outcomes, and guides policy and health care use patterns [33]. It has demonstrated that residing in a disadvantaged area can be as detrimental to health as certain chronic diseases [33]. The Centers for Medicare and Medicaid Services (CMS) leverage the ADI in their strategies [34]. During the COVID-19 pandemic, the ADI informed equitable resource distribution, emphasizing the role of socioeconomic factors in disease outcomes [35,36].

    Using the 2018 ADI, we evaluated community-level disparities. Calculated at the United States census block group level, it reflects a “neighborhood” of approximately 600 to 3000 residents [33]. This granularity offers a localized view of health-related social and geospatial determinants. We geocoded patient addresses, assigning an ADI rank based on their Florida residential census block group, sourced from their 9-digit zip codes. Higher ADI ranks signify greater social disadvantage.

    Geospatial Variables

    We used an extensive exposome-wide association study to identify additional external exposome elements that may correlate with COVID-19 mortality. This study included 337 external exposome factors encompassing 9 distinct categories [20]. It identified 4 external exposome factors at the county level significantly associated with worse outcomes of COVID-19. These included variables characterizing the natural (criteria air pollutants and air toxicants), built (food environment), and social environment (vacant land) [20]. The results of this exposome study reaffirm the importance of environmental and geospatial variables in understanding and predicting outcomes of COVID-19. Our analysis incorporated data relating to these 4 variables by linking them to patients’ places of residence in Florida (Table 1).

    Table 1. Data sources, time periods, and spatial scales of environmental measures.
    Spatial variableData sourceTime periodSpatial scaleTemporal scale
    Particulate matter 2.5Atmospheric composition analysis group, WUSTLa [37]2006-20180.01 degree in lon and lat1 year
    Nitrogen dioxide (air pollution)The center for air, climate, and energy solutions [38]2006-2015Census block group1 year
    Percent students eligible for reduced-price lunch, 2015 (food environment)Food environment atlas [39]2007-2018CountyCross-sectional
    Percent addresses in the previous quarter with no-stat currently in service (vacant land)bAggregated USPSc administrative data on address vacancies, HUDd [40]2006-2019Census tract3 months

    aWUSTL: Washington University in St Louis.

    bTotal no-stat addresses are the addresses that can be classified as “No-Stat” for many reasons, including, (1) rural route addresses that are vacant for 90 days or longer and (2) addresses for businesses or homes under construction and not yet occupied addresses.

    cUSPS: United States Postal Service.

    dHUD: US Department of Housing and Urban Development.

    This table outlines the data sources, time periods, and spatial scales of various environmental measures used in our retrospective cohort study of COVID-19 patients from the OneFlorida Clinical Research Network. These environmental variables, sourced from diverse national databases, were analyzed at different spatial and temporal scales to assess their impact on COVID-19 outcomes among Floridians. The study encompasses adult patients (≥18 years old) with confirmed COVID-19 diagnoses between December 1, 2019, and April 28, 2021.

    Statistical Analysis

    Categorical variables were reported as frequencies and percentages, while continuous variables were reported as means with SDs or medians with IQR. We used the Chi-square test for categorical variables. For continuous variables, we used the t test for normally distributed data or the Wilcoxon-Mann-Whitney U test when data were not normally distributed. Appropriate transformations were applied to certain continuous variables with skewness to achieve a normal distribution. Following this, z score standardization was used to establish a standard scale for interfactor comparison. We also evaluated mixed correlation coefficients between all variables included in our models.

    K-nearest neighbors imputation was used when missing data was present. This technique estimates the missing values based on attributes of the most similar observations, where similarity is calculated using a distance function [41]. K-nearest neighbors imputation is frequently regarded as a more robust and sensitive method for missing value estimation than conventional techniques [42].

    Machine Learning Model Development

    We used Extreme Gradient Boosting (XGBoost), a robust ML framework known for its efficiency, flexibility, and portability [43]. It is an ensemble learning algorithm based on the gradient boosting framework, in which models are built sequentially to boost (increase) the performance of the previous models by using the gradient descent algorithm to minimize errors [43]. Our selection of XGBoost is based on simplified interpretability and the inclusion of feature selection as part of the model-building process. XGBoost exhibits various advantages that render it a compelling alternative to conventional statistical techniques and other ML algorithms.

    We developed and validated 3 consecutive ML models to improve the understanding of the models’ operational dynamics and the incremental contribution of different sets of variables to the outcome. Model 1 consisted exclusively of individual-level variables, which included sociodemographic data and individual comorbidities. Following this, the model was extended to encompass variables at the group level; specifically, we incorporated ADI and environmental measures (particulate matter 2.5 [PM2.5], nitric oxide [NO2], vacant land measure, and food environment measure), resulting in the development of model 2. In our final model, model 3, instead of individual comorbidities, we incorporated the ECI while preserving the group-level variables from model 2. Implementing a unidimensional singular numerical summary of comorbidities facilitates the modeling and integration with other covariates instead of requiring the modeling and interactions between covariates and each constituent of the comorbidity score, thereby enhancing computational efficiency. All models incorporated baseline variables (age, sex, and BMI; Figure 1).

    ‎
    Figure 1. Study flowchart and machine learning analytical framework. ADI: area deprivation index; AUC: area under the receiver operating characteristic curve; EHR: electronic health record; KNN: k-nearest neighbors; SHAP: Shapley Additive Explanations.

    The figure illustrates the study flowchart and the machine learning analytical framework used in the retrospective cohort study of COVID-19 patients from the OneFlorida Clinical Research Network. It provides a visual summary of the study design, data sources, patient selection criteria, and the ML models used to analyze the data, capturing adult patients (≥18 years old) diagnosed with COVID-19 between December 1, 2019, and April 28, 2021.

    In addition, to ensure a comprehensive understanding of our models’ performance, we initiated our analysis by constructing a global model that includes the entire population. Subsequently, we created individualized models for non-Hispanic White and non-Hispanic Black populations across all 3 models. This allowed us to compare the relative importance of features within and between each demographic group.

    Model Performance and Evaluation

    We used nested cross-validation (CV) to evaluate our ML models with 5 outer and inner folds (55 nested CV). A nested CV can help ensure rigor and enhance confidence in model generalizability and scalability [44] (Methods S1 in Multimedia Appendix 1 [44-47]). We calculated and reported the following measures of the models’ performance: the average value and SD of the area under the receiver operating characteristic curve (AUC), balanced accuracy, sensitivity, and specificity from the 5 outer folds (Figure 1). AUC measures the overall discriminative ability, balanced accuracy gauges fair assessment across classes, and sensitivity and specificity offer insight into the model’s prediction accuracy for each class.

    Model Interpretation and Feature Ranking

    Interpreting complex models like XGBoost can be challenging due to their reliance on numerous decision trees, making it difficult to intuit the relationships between features and outcomes. To address this, we used Shapley Additive Explanations (SHAP) values, a method rooted in coalitional game theory [48], which decomposes individual predictions to quantify the influence of each feature (Methods S2 in Multimedia Appendix 1 [44-47]) [45].

    Finally, to assess the robustness of our findings, we conducted a sensitivity analysis using a complete-cases-only approach, excluding cases with missing data and without resorting to data imputation (model 4). We also assessed the potential for divergent feature importance rankings with alternative models using light gradient boosting machine (LightGBM) algorithm [49]. It is another high-performance, gradient-boosting framework that uses tree-based learning algorithms (Figure 1) [49]. Data analyses were performed using (R version 3.6.1; R Core Team) and Python (Version 3.7; Python Software Foundation). The data preprocessing, imputation, and grid search steps were implemented using the Python Sklearn package. The XGBoost algorithm was implemented using the XGBoost package. We followed Enhancing the Quality and Transparency of Health Research Network guidelines for reporting ML analyses in observational studies [50,51].

    Ethical Considerations

    This study used a deidentified electronic health record dataset provided by the OneFlorida Clinical Research Consortium. The data were deidentified before analysis to ensure participant privacy and confidentiality, with no identifying information accessible to the researchers. The study was reviewed and approved by the Institutional Review Board at the University of Florida (Institutional Review Board 202001531), which determined that informed consent was not required as this research involved secondary analysis of existing, deidentified data. No compensation was provided to participants as the study did not involve direct interaction with individuals. In addition, no images or supplementary materials in this manuscript could potentially identify individual participants.


    Our initial data collection included 49,461 patients, of which 35% (n=17,311) were non-Hispanic White and 22.8% (n=11,277) were non-Hispanic Black. After limiting our study to only non-Hispanic White and non-Hispanic Black populations, the final sample size for this analysis was 28,943 patients. Compared with non-Hispanic White patients, non-Hispanic Black patients tended to be younger; had higher BMI; had fewer outpatient visits; and had higher rates of no insurance, visits to the emergency department, and hospital admissions (Table 2). In the non-Hispanic Black group, the burden of comorbidities and the total ECI scores were significantly higher, especially hypertension, coronary artery disease, congestive heart failure (CHF), and chronic kidney disease. Geospatial and neighborhood variables revealed that non-Hispanic Black populations resided in regions with higher (hence less favorable) ADI rankings, NO2, and PM2.5 concentrations, as well as lower vacant land and food environment measures. Finally, the non-Hispanic Black group experienced the highest rates of the composite primary outcome (Table 2). Correlation analysis revealed moderately high correlations when insurance status was compared with PM2.5 and NO2 concentrations. Unsurprisingly, there were also high correlations between ECI scores and most comorbidities (Figure S1 in Multimedia Appendix 1 [44-47]).

    Table 2. Baseline characteristics and outcomes of the study population stratified by racea. The table presents the demographic characteristics of adult patients (≥18 years old) diagnosed with COVID-19 within the OneFlorida Clinical Research Network between December 1, 2019, and April 28, 2021.

    		Non-Hispanic White (n=17,651)Non-Hispanic Black (n=11,293)Total (N=28,944)
    Age, (years), mean (SD)52.82 (20.24)48.16 (18.22)51 (19.61)
    Sex (female), n (%)9490 (53.8)6791 (60.1)16,281 (56.3)
    BMI, kg/m2, mean (SD)29.96 (7.66)33.19 (9.22)31.31 (8.50)
    Insurance status, n (%)

    		Medicare or private15,494 (87.8)8591 (76.1)24,085 (83.2)

    		Medicaid or no insurance1566 (8.9)2191 (19.4)3757 (13.0)

    		Unknown589 (3.3)511 (4.5)1100 (3.8)
    Encounter type, n (%)

    		Outpatient7476 (42.4)2712 (24)10,188 (35.2)

    		EDb4551 (25.8)4525 (40.1)9076 (31.4)

    		Inpatient4884 (27.7)3344 (29.6)8228 (28.4)

    		ICUc740 (4.2)712 (6.3)1452 (5.0)
    Comorbidities

    		ECId, median (IQR)2 (0-5)3 (1-6)2 (0-6)

    		Hypertension, n (%)7645 (43.3)6123 (54.2)13,768 (47.6)

    		Hyperlipidemia, n (%)5989 (33.9)3924 (34.7)9913 (34.2)

    		Diabetes mellitus, n (%)5273 (29.9)4615 (40.9)9888 (34.2)

    		Coronary artery disease, n (%)3089 (17.5)1902 (16.8)4991 (17.2)

    		Congestive heart failure, n (%)1970 (11.2)1656 (14.7)3626 (12.5)

    		Stroke or TIAe, n (%)1655 (9.4)1328 (11.8)2983 (10.3)

    		Dementia, n (%)1437 (8.6)845 (7.7)2282 (8.2)

    		Chronic liver disease, n (%)2424 (13.7)1878 (16.6)4302 (14.9)

    		CKD-Vf/ESRDg, n (%)272 (1.5)565 (5.0)837 (2.9)

    		Respiratory disorders, n (%)9862 (59.1)7436 (67.7)17,298 (62.5)

    		Smoking history, n (%)3115 (17.6)1833 (16.2)4948 (17.1)

    		Substance use, n (%)4103 (23.2)3133 (27.7)7236 (25.0)

    		Alcohol use, n (%)902 (5.1)747 (6.6)1649 (5.7)

    		Mental health disorders, n (%)5331 (32.0)3227 (29.4)8558 (30.9)

    		Common solid cancer, n (%)1074 (6.1)586 (5.2)1660 (5.7)

    		Hematologic malignancies, n (%)291 (1.7)147 (1.3)438 (1.6)
    Geospatial variables, median (IQR)

    		Florida ADIh5 (3-7)8 (6-9)6.0 (4-8)

    		Mean PM2.5i, μg/m37.35(6.88-7.77)7.45(7.05-7.96)7.41 (6.94-7.86)

    		Mean NO2j, ppbk2.73 (1.61-4.63)3.17(1.88-4.62)2.87 (1.69-4.63)

    		Food environmentl4.37 (1.85-4.56)3.68 (1.85-4.25)4.10 (1.85-4.25)

    		Vacant landm0.22 (0.11-0.48)0.16 (0.07-0.31)0.19 (0.10-0.40)
    Unadjusted outcomes, n (%)

    		Composite outcome1423 (8.1)1233 (10.9)2656 (9.2)

    		Death at index encounter251 (1.4)238 (2.1)489 (1.7)

    aAll P values were <.001 (except CAD and history of smoking (P=.148 and .002, respectively). Independent sample t test, Wilcoxon-Mann-Whitney U test, or Pearson chi-square test were used wherever appropriate. Values are presented as means (SD), n (%), or median (IQR).

    bED: emergency department.

    cICU: intensive care unit.

    dECI: Elixhauser Comorbidity Score.

    eTIA: transient ischemic attack.

    fCKD-V: chronic kidney disease stage 5.

    gESRD: end-stage renal disease.

    hADI: area deprivation index.

    iPM2.5: particulate matter 2.5.

    jNO2: nitrogen dioxide.

    kppb: parts per billion.

    lPercent students are eligible for reduced-price lunch, 2015.

    mPercent addresses in the previous quarter with “no-stat” currently in service. “Total No-Stat Addresses” are addresses that can be classified as “No-Stat” for many reasons, including, (1) rural route addresses that are vacant for 90 days or longer and (2) addresses for businesses or homes under construction and not yet occupied addresses.

    When examining the outcomes from the XGBoost modeling, the first model, which solely included individual-level variables, had an AUC value of 0.80, indicating reasonably good model performance (Table 3). The balanced accuracy, sensitivity, and specificity were all around the 0.72-0.74 range, suggesting a well-balanced model capable of predicting both positive and negative outcomes with similar accuracy. The outcomes of model 1 subgroups constructed for non-Hispanic White and non-Hispanic Black populations separately exhibited a comparable performance to the overall model 1, suggesting the racial factor has not significantly influenced these particular model outcomes. In model 2, the performance slightly improved when group-level variables were added. The AUC for the general and race-specific models increased to 0.81-0.83. Likewise, the balanced accuracy also increased. There was a marginal reduction in sensitivity for the non-Hispanic White demographic, implying this model predicted fewer true positives for this subgroup. Finally, model 3, which included the ECI (instead of individual comorbidities) and group-level variables from model 2, demonstrated the most optimal performance across all subgroups, with an AUC of 0.82-0.83. Balanced accuracy was also highest for these models, particularly in sensitivity, which indicates these models are better at predicting true-positive outcomes (Table 3).

    Table 3. Comparative performance metrics of predictive models by racea.
    ModelsAUCbBalanced accuracySensitivitySpecificity
    Model 1, mean (SD)

    		All0.80 (0.01)0.73 (0.01)0.75 (0.02)0.71 (0.01)

    		Whites0.80 (0.01)0.72 (0.01)0.74 (0.03)0.71 (0.01)

    		Blacks0.80 (0.01)0.73 (0.01)0.74 (0.01)0.72 (0.01)
    Model 2, mean (SD)

    		All0.83 (0.01)0.74 (0.01)0.74 (0.02)0.75 (0.01)

    		Whites0.82 (0.01)0.73 (0.01)0.72 (0.03)0.75 (0.01)

    		Blacks0.82 (0.01)0.74 (0.01)0.74 (0.01)0.74 (0.01)
    Model 3, mean (SD)

    		All0.84 (0.01)0.77 (0.01)0.81 (0.01)0.72 (0)

    		Whites0.84 (0.01)0.76 (0.01)0.81 (0.02)0.72 (0.01)

    		Blacks0.84 (0.01)0.76 (0.01)0.81 (0.01)0.71 (0.01)

    aThe table presents comparative performance metrics of predictive models by race, derived from a retrospective cohort study of COVID-19 patients within the OneFlorida Clinical Research Network. The data encompasses adult patients (≥18 years old) diagnosed with COVID-19 between December 1, 2019, and April 28, 2021.

    bAUC: the area under the receiver operating characteristic curve.

    SHAP value analysis in model 1 provided insight into the impact of individual-level variables on predicting COVID-19 outcomes (Figures S2 and S3 in Multimedia Appendix 1 [44-47]). Overall, comorbid conditions such as diabetes mellitus type 2, hypertension, CHF, respiratory disorders, and chronic liver disease consistently ranked the highest across the models and contributed positively to predicting the outcome. Sociodemographic factors such as age and sex ranked consistently high in all models; old age and being female contributed positively to predicting the outcome. BMI and smoking history were important predictors but ranked lower than age and sex. Notably, comorbid conditions such as CHF, rheumatological disorders, and chronic kidney disease stage 5/end-stage renal disease were more significant predictors in the non-Hispanic Black population model, with CHF being the leading predictor.

    The incorporation of group-level variables in model 2 yielded additional nuanced findings. Discernible patterns were noted within specific racial subgroups. Comorbid conditions such as diabetes mellitus type 2, hypertension, CHF, and respiratory disorders remained highly ranked and contributed positively to predicting the outcome. Sociodemographic factors such as age and sex remained important predictors, but BMI and smoking history became less important. Group-level variables such as the food environment measure and air pollution measure, PM2.5, were among the top predictors. Notably, CHF remained the top predictor for the non-Hispanic Black population model. In addition, Florida ADI ranked relatively higher in the overall and the non-Hispanic White population models compared with the non-Hispanic Black population model. However, the food environment measures ranked higher in the non-Hispanic Black population model than in other models (Figures S4 and S5 in Multimedia Appendix 1 [44-47]).

    Model 3 incorporated comorbidities previously distributed across individual conditions in models 1 and 2, consolidating them into the ECI. This provided a more comprehensive risk measure, where the ECI emerged as the topmost predictor, followed by age, sex, air pollution, and food environment measures. Finally, race generally ranked relatively lower than other predictors in all models (Figures 2 and 3).

    ‎
    Figure 2. SHAP feature importance model results for outcomes of the COVID-19 pandemic (model 3). ADI: area deprivation index; NO2: nitrogen dioxide; PM2.5: particulate matter 2.5; SHAP: Shapley Additive Explanations.
    ‎
    Figure 3. Heat map ranking of Shapley Additive Explanations features importance for outcomes of the COVID-19 pandemic in model 3. ADI: area deprivation index; NO2: nitric oxide; PM2.5: particulate matter 2.5.

    The diagrams represent feature importance, arranged from top to bottom. Higher priority features are displayed at the top due to their greater influence on model prediction (having higher predictive power). Color coding indicates the value of the feature: red signifies a higher feature value, while blue represents a lower one. Points to the right of the 0 midline suggest a higher likelihood of the predicted outcome, while those to the left indicate a lower probability of the outcome. Figure 2A displays results for the whole study population. Figure 2B displays results for the non-Hispanic White population. Figure 2C displays results for the non-Hispanic Black population.

    The heat map ranks features in model 3 (all population, White population, and Black population) by importance, with number 1 being the most important feature. The food environment measure is the percentage of students eligible for reduced-price lunch, 2015; vacant land measure is the percentage of addresses in the previous quarter with “no-stat” currently in service.

    Results from the sensitivity analysis (complete cases only without imputation) showed minor differences compared with the primary analysis. ECI, age, and sex remained the top predictors. In the White population, ADI and vacant land measure ranked higher than PM2.5, while in the Black population, no change was noted in feature ranking (Figures S6 and S7 in Multimedia Appendix 1 [44-47]). Similarly, when retraining the model using LightGBM, minimal variations were noted compared with the XGBoost model (Figures S8 and S9 and Table S1 in Multimedia Appendix 1 [44-47]).


    Principal Findings

    The COVID-19 pandemic exposed substantial racial disparities. Our study explored the intricate relationship between comorbidities, SDOH, and poor outcomes of COVID-19 in non-Hispanic Black patients. We analyzed real-world data using explainable ML methods to better understand these disparities while acknowledging the need for further causal inference and statistical testing to fully substantiate our findings.

    Comorbid conditions and ECI were the most important predictors of outcomes of COVID-19, demonstrating the critical role of baseline comorbidities in predicting poor outcomes of COVID-19, irrespective of race. Mainly, CHF was the most important predictor in non-Hispanic Blacks; the reasons for that are multifaceted and complex [52,53]. Patients with CHF who are diagnosed with COVID-19 infection are often older, predominately Black or Hispanic, have a higher prevalence of diabetes and kidney disease, and use more health care resources. This may serve as a proxy for broader health disparities; Black patients, for example, are generally diagnosed with CHF at later stages and experience worse outcomes despite similar symptoms to their White counterparts [54]. Such disparities can be traced back to limited access to preventive care, higher prevalence of CHF risk factors like hypertension and diabetes, and socioeconomic barriers that hinder effective CHF management among Black communities [54].

    Similarly, the food environment measure (county-level percent of students eligible for reduced-price lunch) [39] emerged as one of the top 3 predictors of adverse outcomes among Black individuals. This finding aligns with earlier research [20] and may be partly attributed to historically discriminatory US policies that have resulted in lasting economic and ethnic segregation, manifesting as present-day health disparities [55]. Such inequities are evident in the prevalence of “food deserts,” areas with limited access to healthy food, and the increased risk of repeated hospitalizations, including those due to CHF [56]. Despite adjusting for traditional cardiovascular risk factors, residents of these deprived, racially segregated neighborhoods still faced heightened risk for CVD and CHF [57,58]. Contributory factors include limited recreational facilities [59]; poor walkability [60,61]; and scarce availability of fresh, nutritious foods, particularly in low income and predominantly Black neighborhoods [62].

    The feature importance analysis indicated that “race” has the lowest ranking, which may seem counterintuitive at first glance. Nonetheless, this implies that the racial disparities in outcomes of COVID-19 are predominantly linked to inequalities in health conditions, socioeconomic status, and geospatial factors rather than solely to racial identity [2]. This finding should not be construed as minimizing the importance of race; instead, it suggests that the disparities observed across races may be primarily attributable to the unequal distribution of these conditions across racial groups [63]. Race is accepted as a social construct rather than a biological one [64,65]. Health disparities, such as with COVID-19 infection, often arise from enduring inequalities affecting racial and ethnic minorities, notably non-Hispanic Black individuals [63]. Societal and structural dynamics, more than biological distinctions, reinforce these disparities [63]. These discrepancies remain even after considering biological factors and personal health behaviors [2]. The primary drivers are likely the socioeconomic and environmental conditions experienced by diverse racial and ethnic groups. Our model underscores the dominant role of socioeconomic and environmental factors in health outcomes and disparities.

    Building on the understanding of race as a key determinant, PM2.5 exposure emerged as a notable predictor of health outcomes, ranking fourth for the White population and third for the Black population, ahead of food environment metrics. Previous research links maintained exposure to air pollutants, like PM2.5 and NO2, with increased COVID-19 mortality [66-68]. Originating mainly from fossil fuel burning, PM2.5 could signify heightened pollution exposure in certain demographics, exacerbating outcomes of COVID-19. Environmental findings report greater short- [69] and long-term [70] PM2.5 exposure in racial minorities, especially Black individuals, than in White individuals. Following closely is the food environment measure, emphasizing the combined effects of socioeconomic challenges and prolonged PM2.5 exposure on health. It highlights the multifaceted roots of health disparities, emphasizing the urgency to tackle both socioeconomic and environmental factors for enhanced public health [71]. Such a holistic perspective is pivotal in understanding spatial variations in outcomes of COVID-19 due to interconnected biological, clinical, socioeconomic, and environmental factors [9,72].

    Interestingly, the ADI ranked last in the Black-only model, whereas it came sixth in the White-only model (model 3). This discrepancy does not inherently imply that ADI is inconsequential for the black population but rather that it may interact with other determinants in complex ways. It suggests that although socioeconomic factors are crucial in determining health outcomes, their significance may vary by race. These differences emphasize the importance of considering race-specific factors when analyzing outcomes of COVID-19 and the need for individualized interventions. The higher ranking of the ADI for the White population reflects the well-established relationship between socioeconomic status and health outcomes [34,35]. As measured by the ADI, lower socioeconomic status may contribute to poor health outcomes in this group due to limited access to health care, poorer education, and increased exposure to environmental pollutants and stressors [34,35]. However, the lower ranking of the ADI in the model for the Black population suggests that other factors may be more predictive of outcomes of COVID-19 in this racial group. Another possible explanation is that there is less variation in ADI among Black populations and a greater proportion of Black populations with higher ADIs. The varying ADI values across different socioeconomic levels enable ADI to serve as a strong predictor of outcomes in White populations. In contrast, the higher concentration of individuals with higher ADIs limits the discriminatory power of ADI in this model for Black populations.

    Furthermore, in the model focusing on Black individuals, the food environment measure, indicative of food security and nutrition, ranked higher than ADI, a typical socioeconomic marker. This suggests that factors like food insecurity and neighborhood conditions may have a heightened influence on health outcomes in predominantly Black communities [56]. Research also points to racial differences in stress responses, further influencing health outcomes [73]. Importantly, the persisting impact of systemic racism could contribute to poorer health in Black communities [74], irrespective of socioeconomic factors, which may account for the lower ranking of ADI in the Black-specific model.

    Limitations

    The study has several limitations. The findings in this study should be considered as part of a broader discussion on racial disparities in health. We acknowledge the need for further causal inference and statistical testing to fully substantiate our observations. The data might not necessarily apply to other states with different demographics or health disparities than Florida. Using EHR data for research is subject to informatic challenges and disadvantages [75]. Race, as used in our study, was self-reported and extracted from EHR. While this method of racial identification is standard in epidemiological research, it comes with inherent challenges that may affect the accuracy and interpretability of the findings. The study could not assess the outcomes of patients who did not require hospitalization or experienced mortality outside the clinical setting. Furthermore, although the XGBoost algorithm has a low risk of overfitting, the lack of an external validation cohort undermines the generalizability of our model.

    Similarly, interpreting features’ importance from ML models, such as XGBoost, presents inherent challenges compared with traditional statistical approaches. These outputs do not offer conventional statistical significance measures and can be highly sensitive to model specification and training data characteristics. Also, the models’ evaluation metrics have limitations. The AUC might not fully capture the model’s performance when there is a significant class imbalance or different types of misclassifications vary. Balanced accuracy might not always reflect the practical significance of prediction errors. Sensitivity and specificity can provide a misleading picture of model performance if not considered together, especially in datasets with imbalanced class distributions.

    Finally, associations obtained from part of the data at the county level may not reflect individual associations.

    In conclusion, our study demonstrates the critical role of model constructs and assumptions in estimating health-related associations, advocating for frameworks that better account for data behaviors. Using a comprehensive ML approach that integrates individual- and group-level exposomic health associations, we used sequential modeling and universal Shapley effect plots for objective comparisons. Our findings emphasize the complexities of health inequalities, particularly persistent racial disparities, and stress the need for multidimensional strategies to address them. Interpretable ML serves as a valuable adjunct to traditional statistical methods, revealing nuanced patterns that can inform resource allocation and policy development for outcomes of COVID-19. Further research is required to clarify the influence of these variables and their contributions to racial health disparities.

    Data Availability

    The datasets analyzed during this study are not publicly available due to institutional review board protections and privacy considerations. However, the data may be accessed upon reasonable request through the OneFlorida Research Consortium. Information on the data request process is available on the OneFlorida Clinical Research Network’s website.

    Conflicts of Interest

    None declared.

    Multimedia Appendix 1

    Supplementary material.

    DOCX File , 17990 KB
    1. Disparities in COVID-19-associated hospitalizations. Centers for Disease Control and Prevention. 2022. URL: https:/​/www.​cdc.gov/​coronavirus/​2019-ncov/​community/​health-equity/​racial-ethnic-disparities/​disparities-hospitalization.​html [accessed 2022-06-14]
    2. Magesh S, John D, Li W, Li Y, Mattingly-App A, Jain S, et al. Disparities in COVID-19 outcomes by race, ethnicity, and socioeconomic status: a systematic-review and meta-analysis. JAMA Netw Open. 2021;4(11):e2134147. [FREE Full text] [CrossRef] [Medline]
    3. Clark A, Jit M, Warren-Gash C, Guthrie B, Wang HHX, Mercer SW, et al. Centre for the Mathematical Modelling of Infectious Diseases COVID-19 working group. Global, regional, and national estimates of the population at increased risk of severe COVID-19 due to underlying health conditions in 2020: a modelling study. Lancet Glob Health. 2020;8(8):e1003-e1017. [FREE Full text] [CrossRef] [Medline]
    4. Chowkwanyun M, Reed AL. Racial health disparities and COVID-19 - caution and context. N Engl J Med. 2020;383(3):201-203. [CrossRef] [Medline]
    5. Geronimus AT, Hicken M, Keene D, Bound J. "Weathering" and age patterns of allostatic load scores among blacks and whites in the United States. Am J Public Health. 2006;96(5):826-833. [CrossRef] [Medline]
    6. Vinciguerra M, Greco E. Sars-CoV-2 and black population: ACE2 as shield or blade? Infect Genet Evol. 2020;84:104361. [FREE Full text] [CrossRef] [Medline]
    7. McCoy J, Wambier CG, Vano-Galvan S, Shapiro J, Sinclair R, Ramos PM, et al. Racial variations in COVID-19 deaths may be due to androgen receptor genetic variants associated with prostate cancer and androgenetic alopecia. are anti-androgens a potential treatment for COVID-19? J Cosmet Dermatol. 2020;19(7):1542-1543. [FREE Full text] [CrossRef] [Medline]
    8. Wacquant L. Territorial stigmatization in the age of advanced marginality. Thesis Elev. 2016;91(1):66-77. [FREE Full text] [CrossRef]
    9. Ren X, Mi Z, Georgopoulos PG. Socioexposomics of COVID-19 across New Jersey: a comparison of geostatistical and machine learning approaches. J Expo Sci Environ Epidemiol. 2024;34(2):197-207. [FREE Full text] [CrossRef] [Medline]
    10. Andrianou XD, Pronk A, Galea KS, Stierum R, Loh M, Riccardo F, et al. Exposome-based public health interventions for infectious diseases in urban settings. Environ Int. 2021;146:106246. [FREE Full text] [CrossRef] [Medline]
    11. Martin-Sanchez F, Atienza-Maderuelo M, Lopez-Campos G, Collado P. Use of informatics to characterise the exposome of COVID-19. BMJ Health Care Inform. 2021;28(1):e100371. [FREE Full text] [CrossRef] [Medline]
    12. Horton R. Offline: COVID-19 is not a pandemic. Lancet Lond Engl. 2020;396(10255):874. [FREE Full text] [CrossRef] [Medline]
    13. Mei X, Lee HC, Diao K, Huang M, Lin B, Liu C, et al. Artificial intelligence-enabled rapid diagnosis of COVID-19 patients. Nat Med. 2020;26:1224-1228. [FREE Full text] [CrossRef] [Medline]
    14. Zhang K, Liu X, Shen J, Li Z, Sang Y, Wu X, et al. Clinically applicable AI system for accurate diagnosis, quantitative measurements, and prognosis of COVID-19 pneumonia using computed tomography. Cell. 2020;182(5):1360. [FREE Full text] [CrossRef] [Medline]
    15. Ucar F, Korkmaz D. COVIDiagnosis-Net: deep bayes-squeezeNet based diagnosis of the coronavirus disease 2019 (COVID-19) from X-ray images. Med Hypotheses. 2020;140:109761. [FREE Full text] [CrossRef] [Medline]
    16. Shi F, Xia L, Shan F, Song B, Wu D, Wei Y, et al. Large-scale screening to distinguish between COVID-19 and community-acquired pneumonia using infection size-aware classification. Phys Med Biol. 2021;66(6):065031. [CrossRef] [Medline]
    17. Wang P, Zheng X, Li J, Zhu B. Prediction of epidemic trends in COVID-19 with logistic model and machine learning technics. Chaos Solitons Fractals. 2020;139:110058. [FREE Full text] [CrossRef] [Medline]
    18. Cheng FY, Joshi H, Tandon P, Freeman R, Reich DL, Mazumdar M, et al. Using machine learning to predict ICU transfer in hospitalized COVID-19 patients. J Clin Med. 2020;9(6):1668. [CrossRef] [Medline]
    19. Mathur P, Sethi T, Mathur A, Maheshwari K, Cywinski JB, Khanna AK, et al. Explainable machine learning models to understand determinants of COVID-19 mortality in the United States. medRxiv. 2020. [FREE Full text] [CrossRef]
    20. Hu H, Zheng Y, Wen X, Smith SS, Nizomov J, Fishe J, et al. An external exposome-wide association study of COVID-19 mortality in the United States. Sci Total Environ. 2021;768:144832. [FREE Full text] [CrossRef] [Medline]
    21. Clinical and translational science institute university of florida. OneFlorida+ Clinical Research Network. URL: https://www.ctsi.ufl.edu/ctsa-consortium-projects/oneflorida-clinical-research-network/ [accessed 2022-11-17]
    22. Fleurence RL, Curtis LH, Califf RM, Platt R, Selby JV, Brown JS. Launching PCORnet, a national patient-centered clinical research network. J Am Med Inform Assoc. 2014;21(4):578-582. [FREE Full text] [CrossRef] [Medline]
    23. Brett AS, Goodman CW. First impressions - should we include race or ethnicity at the beginning of clinical case presentations? N Engl J Med. 2021;385(27):2497-2499. [CrossRef] [Medline]
    24. VanderWeele TJ, Robinson WR. On the causal interpretation of race in regressions adjusting for confounding and mediating variables. Epidemiology. 2014;25(4):473-484. [FREE Full text] [CrossRef] [Medline]
    25. Corbie-Smith G, Henderson G, Blumenthal C, Dorrance J, Estroff S. Conceptualizing race in research. J Natl Med Assoc. 2008;100(10):1235-1243. [FREE Full text] [CrossRef] [Medline]
    26. Hogan WR, Shenkman EA, Robinson T, Carasquillo O, Robinson PS, Essner RZ, et al. The OneFlorida data trust: a centralized, translational research data infrastructure of statewide scope. J Am Med Inform Assoc. 2022;29(4):686-693. [FREE Full text] [CrossRef] [Medline]
    27. Certain Medical Conditions and Risk for Severe COVID-19 Illness | CDC. URL: https://www.cdc.gov/coronavirus/2019-ncov/need-extra-precautions/people-with-medical-conditions.html [accessed 2020-08-01]
    28. Clinical Classifications Software Refined. URL: https://www.hcup-us.ahrq.gov/toolssoftware/ccsr/ccs_refined.jsp [accessed 2023-02-01]
    29. Elixhauser A, Steiner C, Harris DR, Coffey RM. Comorbidity measures for use with administrative data. Med Care. 1998;36(1):8-27. [CrossRef] [Medline]
    30. Chang HJ, Chen PC, Yang CC, Su YC, Lee CC. Comparison of elixhauser and charlson methods for predicting oral cancer survival. Medicine (Baltimore). 2016;95(7):e2861. [FREE Full text] [CrossRef] [Medline]
    31. Menendez ME, Neuhaus V, van Dijk CN, Ring D. The elixhauser comorbidity method outperforms the charlson index in predicting inpatient death after orthopaedic surgery. Clin Orthop Relat Res. 2014;472(9):2878-2886. [FREE Full text] [CrossRef] [Medline]
    32. Qeadan F, VanSant-Webb E, Tingey B, Rogers TN, Brooks E, Mensah NA, et al. Racial disparities in COVID-19 outcomes exist despite comparable elixhauser comorbidity indices between blacks, hispanics, native Americans, and whites. Sci Rep. 2021;11(1):8738. [FREE Full text] [CrossRef] [Medline]
    33. Kind AJH, Jencks S, Brock J, Yu M, Bartels C, Ehlenbach W, et al. Neighborhood socioeconomic disadvantage and 30-day rehospitalization: a retrospective cohort study. Ann Intern Med. 2014;161(11):765-774. [FREE Full text] [CrossRef] [Medline]
    34. Maroko AR, Doan TM, Arno PS, Hubel M, Yi S, Viola D. Integrating social determinants of health with treatment and prevention: a new tool to assess local area deprivation. Prev Chronic Dis. 2016;13:E128. [FREE Full text] [CrossRef] [Medline]
    35. Adjei-Fremah S, Lara N, Anwar A, Garcia DC, Hemaktiathar S, Ifebirinachi CB, et al. The effects of race/ethnicity, age, and area deprivation index (ADI) on COVID-19 disease early dynamics: Washington, DC case study. J Racial Ethn Health Disparities. 2023;10(2):491-500. [FREE Full text] [CrossRef] [Medline]
    36. Srivastava T, Schmidt H, Sadecki E, Kornides ML. Disadvantage indices deployed to promote equitable allocation of COVID-19 vaccines in the US: a scoping review of differences and similarities in design. JAMA Health Forum. 2022;3(1):e214501. [FREE Full text] [CrossRef] [Medline]
    37. Atmospheric Composition Analysis Group. Washington University in St. Louis. URL: https://sites.wustl.edu/acag/ [accessed 2022-11-16]
    38. CACES. URL: https://www.caces.us [accessed 2022-11-16]
    39. USDA ERS - Food Environment Atlas. URL: https://www.ers.usda.gov/data-products/food-environment-atlas/ [accessed 2022-11-16]
    40. HUD aggregated USPS administrative data on address vacancies. HUD USER. URL: https://www.huduser.gov/portal/datasets/usps.html [accessed 2022-11-16]
    41. Batista GEAPA, Monard MC. An analysis of four missing data treatment methods for supervised learning. Appl Artif Intell. 2003;17(5-6):519-533. [FREE Full text] [CrossRef]
    42. Hron K, Templ M, Filzmoser P. Imputation of missing values for compositional data using classical and robust methods. Comput Stat Data Anal. 2010;54(12):3095-3107. [FREE Full text] [CrossRef]
    43. Chen T, Guestrin C. GBoost: A Scalable Tree Boosting System. Association for Computing Machinery; 2016. Presented at: GBoost: A Scalable Tree Boosting System; 13 August 2016:785-794; New York, NY, USA. URL: https://doi.org/10.1145/2939672.2939785 [CrossRef]
    44. Cawley CG, Talbot NLC. On over-fitting in model selection and subsequent selection bias in performance evaluation. J Mach Learn Res. 2010;11(70):2079-2107. [FREE Full text]
    45. Lundberg S. Interpretable Machine Learning with XGBoost. Medium. 2020. URL: https://towardsdatascience.com/interpretable-machine-learning-with-xgboost-9ec80d148d27 [accessed 2023-05-25]
    46. Jiménez ÁB, Lázaro JL, Dorronsoro JR. Finding optimal model parameters by discrete grid search. In: Corchado E, Corchado JM, Abraham A, editors. Innovations in Hybrid Intelligent Systems. Advances in Soft Computing. Berlin, Heidelberg, Germany. Springer-Verlag; 2007:120-127.
    47. SHAP values. Kaggle. URL: https://kaggle.com/code/dansbecker/shap-values [accessed 2024-08-19]
    48. Arrow KJ, Barankin EW, Blackwell D, Bott R, Dalkey N, Dresher M. Contributions to the Theory of Games (AM-28), Volume II. USA. Princeton University Press; 1953.
    49. Ke G, Meng Q, Finley T, Wang T, Chen W, Ma W. LightGBM: a highly efficient gradient boosting decision tree. Curran Associates Inc; 2017. Presented at: Proc. 31st Int. Conf. Neural Inf. Process. Syst., Red Hook; 04 December 2017:3149-3157; NY, USA.
    50. Recommendations for Reporting Machine Learning Analyses in Clinical Research | EQUATOR Network. URL: https:/​/www.​equator-network.org/​reporting-guidelines/​recommendations-for-reporting-machine-learning-analyses-in-clinical-research/​ [accessed 2023-05-25]
    51. Stevens LM, Mortazavi BJ, Deo RC, Curtis L, Kao DP. Recommendations for reporting machine learning analyses in clinical research. Circ Cardiovasc Qual Outcomes. Oct 2020;13(10):e006556. [CrossRef]
    52. Italia L, Tomasoni D, Bisegna S, Pancaldi E, Stretti L, Adamo M, et al. COVID-19 and heart failure: from epidemiology during the pandemic to myocardial injury, myocarditis, and heart failure sequelae. Front Cardiovasc Med. 2021;8:713560. [FREE Full text] [CrossRef] [Medline]
    53. John KJ, Mishra AK, Ramasamy C, George AA, Selvaraj V, Lal A. Heart failure in COVID-19 patients: critical care experience. World J Virol. 2022;11(1):1-19. [FREE Full text] [CrossRef] [Medline]
    54. Nayak A, Hicks AJ, Morris AA. Understanding the complexity of heart failure risk and treatment in black patients. Circ Heart Fail. 2020;13(8):e007264. [FREE Full text] [CrossRef] [Medline]
    55. Mitchell B, Franco J. HOLC “redlining” maps: The persistent structure of segregation and economic inequality. Home Owners' loan corporation “redlining” maps: The persistent structure of segregation and economic inequality. 2018. [CrossRef]
    56. Morris AA, McAllister P, Grant A, Geng S, Kelli HM, Kalogeropoulos A, et al. Relation of living in a "Food Desert" to recurrent hospitalizations in patients with heart failure. Am J Cardiol. 2019;123(2):291-296. [FREE Full text] [CrossRef] [Medline]
    57. Akwo EA, Kabagambe EK, Harrell FE, Blot WJ, Bachmann JM, Wang TJ, et al. Neighborhood deprivation predicts heart failure risk in a low-income population of blacks and whites in the southeastern United States. Circ Cardiovasc Qual Outcomes. 2018;11(1):e004052. [FREE Full text] [CrossRef] [Medline]
    58. Diez Roux AV, Merkin SS, Arnett D, Chambless L, Massing M, Nieto FJ, et al. Neighborhood of residence and incidence of coronary heart disease. N Engl J Med. 2001;345(2):99-106. [CrossRef] [Medline]
    59. Lu Y, Chen L, Liu X, Yang Y, Sullivan WC, Xu W, et al. Green spaces mitigate racial disparity of health: a higher ratio of green spaces indicates a lower racial disparity in SARS-CoV-2 infection rates in the USA. Environ Int. 2021;152:106465. [FREE Full text] [CrossRef] [Medline]
    60. Howell NA, Tu JV, Moineddin R, Chu A, Booth GL. Association between neighborhood walkability and predicted 10-year cardiovascular disease risk: the CANHEART (Cardiovascular Health in Ambulatory Care Research Team) cohort. J Am Heart Assoc. Nov 05, 2019;8(21):e013146. [FREE Full text] [CrossRef] [Medline]
    61. Powell LM, Slater S, Chaloupka FJ, Harper D. Availability of physical activity-related facilities and neighborhood demographic and socioeconomic characteristics: a national study. Am J Public Health. 2006;96(9):1676-1680. [CrossRef] [Medline]
    62. Morland K, Wing S, Diez Roux A, Poole C. Neighborhood characteristics associated with the location of food stores and food service places. Am J Prev Med. 2002;22(1):23-29. [CrossRef] [Medline]
    63. National Academies of Sciences, Engineering, and Medicine, Health and Medicine Division, Board on Population Health and Public Health Practice, Committee on Community-Based Solutions to Promote Health Equity in the United States. The state of health disparities in the United States. In: Communities in Action: Pathways to Health Equity. Washington, DC. National Academies Press; 2017.
    64. New AMA policies recognize race as a social, not biological, construct. American Medical Association. Nov 16, 2020. URL: https:/​/www.​ama-assn.org/​press-center/​press-releases/​new-ama-policies-recognize-race-social-not-biological-construct [accessed 2023-05-27]
    65. Braveman P, Parker Dominguez T. Abandon "Race." focus on racism. Front Public Health. 2021;9:689462. [FREE Full text] [CrossRef] [Medline]
    66. Zhu Y, Xie J, Huang F, Cao L. Association between short-term exposure to air pollution and COVID-19 infection: evidence from China. Sci Total Environ. 2020;727:138704. [FREE Full text] [CrossRef] [Medline]
    67. Konstantinoudis G, Padellini T, Bennett J, Davies B, Ezzati M, Blangiardo M. Long-term exposure to air-pollution and COVID-19 mortality in England: a hierarchical spatial analysis. Environ Int. 2021;146:106316. [FREE Full text] [CrossRef] [Medline]
    68. Wu X, Nethery RC, Sabath MB, Braun D, Dominici F. Air pollution and COVID-19 mortality in the United States: strengths and limitations of an ecological regression analysis. Sci Adv. 2020;6(45):eabd4049. [FREE Full text] [CrossRef] [Medline]
    69. Collins TW, Grineski SE. Racial/Ethnic disparities in short-term air pollution exposures in the United States. Environ Health Perspect. 2022;130(8):87701. [FREE Full text] [CrossRef] [Medline]
    70. Jbaily A, Zhou X, Liu J, Lee TH, Kamareddine L, Verguet S, et al. Air pollution exposure disparities across US population and income groups. Nature. 2022;601(7892):228-233. [FREE Full text] [CrossRef] [Medline]
    71. Lebow-Skelley E, Young L, Noibi Y, Blaginin K, Hooker M, Williamson D, et al. Defining the exposome using popular education and concept mapping with communities in Atlanta, Georgia. Front Public Health. 2022;10:842539. [FREE Full text] [CrossRef] [Medline]
    72. Koch S, Yoon L, Gils B. From the exposome to the socioexposome in COVID-19 research-a call for more multidisciplinary research. JAMA Netw Open. 2020;3(12):e2032287. [FREE Full text] [CrossRef] [Medline]
    73. Brown LL, Abrams LR, Mitchell UA, Ailshire JA. Measuring more than exposure: does stress appraisal matter for black-white differences in anxiety and depressive symptoms among older adults? Innov Aging. 2020;4(5):igaa040. [FREE Full text] [CrossRef] [Medline]
    74. Williams DR, Mohammed SA, Leavell J, Collins C. Race, socioeconomic status, and health: complexities, ongoing challenges, and research opportunities. Ann N Y Acad Sci. 2010;1186:69-101. [FREE Full text] [CrossRef] [Medline]
    75. Methods and dimensions of electronic health record data quality assessment: enabling reuse for clinical research. Journal of the American Medical Informatics Association | Oxford Academic. URL: https://academic.oup.com/jamia/article/20/1/144/2909176?login=true [accessed 2021-09-11]

    ‎
    ADI: area deprivation index
    AUC: area under the receiver operating characteristic curve
    CHF: congestive heart failure
    CV: cross-validation
    ECI: Elixhauser combined comorbidity index
    HCUP: Healthcare Cost and Utilization Project
    ML: machine learning
    NO2: nitrogen dioxide
    PM2.5: particulate matter 2.5
    SDOH: Social Determinants of Health
    SHAP: Shapley Additive Explanations
    XGBoost: Xtreme Gradient Boosting

    Edited by A Mavragani; submitted 12.11.23; peer-reviewed by K Wang, S El Khamlichi, S Jonathan; comments to author 07.02.24; revised version received 01.04.24; accepted 29.05.24; published 26.09.24.

    Copyright

    ©Osama Dasa, Chen Bai, Ruba Sajdeya, Stephen E Kimmel, Carl J Pepine, Matthew J Gurka J, Reinhard Laubenbacher, Thomas A Pearson, Mamoun T Mardini. Originally published in JMIR Public Health and Surveillance (https://publichealth.jmir.org), 26.09.2024.

    This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Public Health and Surveillance, is properly cited. The complete bibliographic information, a link to the original publication on https://publichealth.jmir.org, as well as this copyright and license information must be included.