Unraveling racial disparities in asthma emergency department visits using electronic healthcare records and machine learning

Abstract

Objective: Hospital emergency department (ED) visits by asthmatics differ based on race and season. The objectives of this study were to investigate season- and race-specific disparities for asthma risk, and to identify environmental exposure variables associated with ED visits among more than 42,000 individuals of African American (AA) and European American (EA) descent identified through electronic health records (EHRs).

Methods: We examined data from 42,375 individuals (AAs = 14,491, EAs = 27,884) identified in EHRs. We considered associated demographic (race, age, gender, insurance), clinical (smoking status, ED visits, FEV1%), and environmental exposures data (mold, pollen, and pollutants). Machine learning techniques, including random forest (RF), extreme gradient boosting (XGB), and decision tree (DT) were used to build and identify race- and -season-specific predictive models for asthma ED visits.

Results: Significant differences in ED visits and FEV1% among AAs and EAs were identified. ED visits by AAs was 32.0% higher than EAs and AAs had 6.4% lower FEV1% value than EAs. XGB model was used to accurately classify asthma patients visiting ED into AAs and EAs. Pollen factor and pollution (PM2.5, PM10) were the key variables for asthma in AAs and EAs, respectively. Age and cigarette smoking increase asthma risk independent of seasons.

Conclusions: In this study, we observed racial and season-specific disparities between AAs and EAs asthmatics for ED visit and FEV1% severity, suggesting the need to address asthma disparities through key predictors including socio-economic status, particulate matter, and mold.

Keywords: EHR; Race disparities; electronic healthcare records; machine learning; seasonal variation.

Figure 1.. Schematic representation of the workflow for clinical, and socio-environmental exposure datasets.

We used Python, a computer programming language, to program a data merging routine. We used Pandas, a python code package, to run Python code. We then used R, a statistical coding language, within R Studio, for asthma severity classification and ED Visit frequency. The previous steps were followed by machine learning algorithms through R packages (caret, Healthcare.ai), to produce the predictive models and prioritize the best predictive model. Finally, the most important variables that contributed to the best performing model were used for prediction. This pipeline is applied for both African American and European American datasets.

Figure 2.. Variation of ED visit and FEV1% by age.

Patients are grouped by age less than 20, 20-29, 30-39, 40-49, 50-59, 60-69, 70-79, and 80+. Outliers are shown by dots. The color-coded line segment show the line of best fit with the age. ED visit shows positive correlation for both races (r_black = 0.29, p_black < 2E-16; r_white = 0.19, p_black < 2E-16) and FEV1% shows negative correlation in both races (r_black = −0.1215, p_black < 2E-16; r_white = −0.1517, p_black < 2E-16). AA = African American, EA = European American.

Figure 3.. Percent disparities of AAs relative to EAs.

Figure showed the racial disparities in environmental exposures (pollen, mold, and air pollution), age, ED visit and FEV1% between AAs and EAs. The negative percent disparity measure of FEV1% showed that AA asthmatics had lower FEV1% compare to EA asthmatics. Similarly, AA were slightly younger than the EA as shown by the negative percent disparity for age. The ED visit had a positive mean percentage disparity which showed more frequent ED visits by AAs asthmatics than the EAs. AA = African American, EA = European American.

Figure 4.. Circle plot showing significant differences in variables between AA and EA.

Significant differences between races exist in asthma-related exposure of environmental (mold, pollen, and pollution) and clinically relevant factors (age, gender, smoking, FEV1%). P-value is adjusted using The Benjamini-Hochberg procedure. Each circle in the outer later represent a variable. Darker the color shade represents the strength of p-value. Color coded connecting lines show the seasonal correlation (|cor| ≥ 0.4) between the variables. Several of the environmental variables are in moderate to high correlation across the seasons. AA = African American, EA = European American.

Figure 5.. Graphical presentation of prediction accuracy rate of (A) ED visit and (B) asthma severity for both AA and EAs.

The figure shows the accuracy rate under XGB, RF, DT, and LR classification models for both AA and EAs across seasons. Dashed line at the middle of each bar shows the best guess accuracy rate. XGB = extreme gradient boosting, RF = Random forest, DT = Decision tree, LR = Logistic regression; AA = African American, EA = European American. *** = P-value < 10-5, ** = P-value < 10-3, * = P-value < 0.05.