Lung disease is a severe problem in the United States. According to the Centers for Disease Control and Prevention (CDC), asthma is responsible for at least 3000 deaths per year and chronic obstructive pulmonary disease (COPD) is responsible for at least 150,000 deaths per year. COPD is a progressive lung disease, encompassing chronic bronchitis and emphysema, which is characterized by airflow limitation and breathing difficulties. Asthma and COPD can co-occur (asthma-COPD overlap), with increased risk of mortality [1] and diminished disease-related quality of life [2]. This is from a variety of factors, some under individual control, such as cigarette smoking, and others not under individual control, such as ambient air pollution.

Cigarette smoking has been trending downward in recent years, thanks in part to public health advertisement campaigns. Nevertheless, air quality can be dangerously poor at times, which exacerbates lung health problems [3], and the impacts can be particularly acute in populations considered vulnerable. Technologically, there are tools that help individuals avoid poor air quality. For example, there are mobile phone apps that track air quality. They notify their owners on days when air quality is dangerously poor, advising them to stay indoors or avoid strenuous outdoor exercise. The effectiveness of such apps is ambiguous thus far [4,5].

The rest of the paper is organized as follows. We first review prior work regarding the possible factors contributing to COPD in adults. We then describe our methods, including data sources for the variables of interest and descriptive statistics. Following this, we will describe and interpret the results of our multiple linear regression (MLR) and machine learning (ML) models. We conclude by describing the overall research contributions as well as limitations and future directions.

There is substantial literature on factors contributing to COPD, including a wide variety of environmental, economic, and demographic variables; the etiology of COPD is multifactorial, with smoking being the most well-known contributing factor. Furthermore, the combination of environmental pollutants and cigarette smoke has shown synergistic effects, accelerating the decline in lung function and worsening COPD [6,7]. In addition, occupational exposures, for example, to coal dust, arsenic, or diesel fumes, or to home exposures, such as gas stoves, wood stoves, kerosene heaters, and fireplaces, contribute to overall COPD outcomes. When combined with persistent ambient air pollution, the risk and severity of COPD will likely increase [8].

Pollutants and copollutants are associated with decreased lung function and can lead to COPD. The loss can range from mild, such as allergies, to severe, that is, mortality. Air quality varies widely throughout the United States because of pollutants and copollutants, and climate change may be worsening it, particularly for populations considered vulnerable [9]. Health disparities due to poor quality air and other stressors are well known [10-12]. Ambient air pollution in poorer neighborhoods tend to be exacerbated by additional copollutants, heat stress, and aeroallergens. Air quality index (AQI) includes the totality of pollutants and copollutants.

ML methods have been applied increasingly to public health and medical problems. For example, ML has been used to support the public health response to COVID-19 through surveillance, case identification, contact tracing, and evaluating interventions [13]. ML methods have been used as a supportive tool to recognize cardiac arrest in emergency calls [14]. In that study, Zicari et al [14] developed a general protocol with a collaborative team to ensure that the ML tool was domain- and context-sensitive as well as abiding by ethical guidelines, thus obtaining trustworthiness. ML has been also used to improve early and accurate stroke recognition during emergency medical calls [15].

ML methods have been used to study COPD, in particular. For example, ML methods have been used to develop a prediction system using lifestyle data, environmental factors, and patient symptoms for the early detection of acute exacerbations of COPD within a 7-day window [16]. Another study on acute exacerbations of COPD compared several ML methods and found that a decision tree classifier was best for assessing patient severity and guiding treatment strategy [17]. In another study, to improve mortality prediction from COPD, a random forest was used to identify the most important imaging features [18]. Gradient boosted trees (GBTs) have been used to predict lung function values from computed tomography images obtained from patients with COPD and those without COPD [19]. Deep learning has been effective in analyzing images diagnostic of COPD [20]. Finally, research using a generalized linear model found a complex relationship between rural living and COPD-related outcomes in US veterans [21]. Thus, a variety of ML models have been successfully applied for use in public health scenarios in general and COPD in particular. The one that ultimately works best in a given situation depends on many factors.

Different races and ethnicities may have different baseline rates of disease due to various factors, including historical misdiagnosis and mistreatment of various racial or ethnic groups, which leads to differential outcomes [22]. There may be outcome, equity, and counseling differences by gender as well as race or ethnicity in the diagnosis and treatment of COPD [23,24].

We had three general expectations of COPD in our models:

This paper used MLR and ML methods to predict COPD at the core-based statistical area (CBSA) level [25]. At the time of this study, there were 388 metropolitan and 541 micropolitan statistical areas in the United States. The data sources were obtained from data repositories of 3 official US agencies, specifically from the CDC. We gathered, integrated, and checked them for data quality. By combining different variables from this variety of data sources, we aimed to obtain a uniquely high accuracy model, while simultaneously reducing biases or flaws that may be attributable to individual data sources. We further checked for missing values (ie, NULL or NA) in every variable. We checked for data correctness by checking the plots of the distributions for every variable, looking for impossible or outlying values. Table 1 shows the data sources used.

Data were collected for all CBSAs that were available from 2016 to 2019. All data obtained from the CDC were contributed voluntarily at the individual level and aggregated to remove all personally identifiable information [29].

The COPD rates are for 2019, whereas all the predictor variables are averaged over the timespan from 2016 to 2018. As such, the models obtained are predictive over time. The data collection result was 517 (56%) of the 929 CBSAs, with proportionally more from the 388 metropolitan statistical areas than from the 541 micropolitan statistical areas. The response variable is the percentage of the CBSA having COPD. We modeled all factors as random variables directly contributing to COPD, which is measured as the proportion (percentage) of the population having COPD. Race or ethnicity was also modeled as percentage of the population rather than as categorical variables. All variables in Table 2 are averaged as mean, except for household income, which was averaged as median.

In Figure 1, we observe that some variables (ie, population, gross domestic product [GDP], GDP per capita, and median household income) are skewed in their distribution.

Therefore, we made a log transformation of these variables (ie, logPopl, logGDP, logGDPpc, and logHHI) to make them less skewed, and we show a heat map of correlations of them with the other variables in Figure 2.

We see a range of correlations, from very negative (green) to negative (orange) to positive (purple) to very positive (pink). In the rightmost column, we see the correlations between the response variable, COPD rate, and the other variables, ranging from very positive (smoking rate) to moderately positive (poverty and unemployment rates) to moderately negative (education and logged household income) to slightly negative (log of GDP, log of population, log of GDP per capita, and AQI). Given these correlations, we are likely to find good predictive models, but we need to check for multicollinearity in any linear model that we identify.

To understand and model COPD, one has to consider the consistently largest contributing factor: cigarette smoking. Research tends to either control for cigarette smoking or exclude it entirely. In this paper, we chose to include cigarette smoking, accounting for it in our models, but also to examine other factors to compare the magnitudes of influence among the various factors. We aimed to model a variety of factors, including cigarette smoking, to arrive at the model that predicts COPD with the greatest accuracy.

This research did not involve human subjects at the individual level and therefore did not require institutional review board approval. Our data were collected from CDC sources at the level of CBSA. All sources were free of personal identifying information, because the CDC is legally required to ensure the protection of the data. All data were collected and aggregated in a non–personal identifying information manner. The results of our analysis do suggest communicating with different racial and ethnic groups differently, tailoring the implications directly to patients as well as indirectly to their families, communities, and health care providers in a race- or ethnicity-sensitive manner.

We had three general expectations, which were largely met:

Consistent with the literature, we found that smoking remains the most significant risk factor for COPD, with research consistently demonstrating a strong association between smoking status and COPD prevalence. In our MLR, we found that smoking rate is the strongest predictor of COPD rate. We found the same result in our GBT but also found that the smoking rate has a curvilinear, almost exponential, relationship with COPD. The Rotterdam study, a large-scale population-based cohort study, found that current and former smokers had a substantially higher risk of developing COPD compared to never smokers [46]. A nationwide population-based cohort study in South Korea demonstrated that smoking cessation after COPD diagnosis was associated with lower all-cause and cause-specific mortality [47].

Notably, 3 of the 4 next most important variables, in terms of impact in our GBT, are socioeconomic variables: household income (rank 2), education rate (rank 4), and unemployment rate (rank 5). In the MLR, we found that household income (logged) had the second highest impact. In the GBT, household income had the second highest impact, but the tipping point was around US $40,000, after which higher income had a linear, negative relationship with COPD. Education rate had a strongly negative, curvilinear relationship with COPD. Unemployment rate had a sharply positive relationship with COPD, but then peaked at 5% unemployment, after which it plateaued.

These results are largely consistent with the literature on socioeconomic factors and smoking behavior, suggesting an indirect relationship with COPD via smoking. A study examining smoking among adolescents in 6 European cities found that disposable income was positively associated with smoking [48]. Conversely, lower socioeconomic status was associated with higher COPD prevalence because in addition to lower education and income, there may be environmental pollutants, occupational hazards, or barriers to COPD screening, diagnosis, and treatment [49]. In contrast with the literature, our SHAP plots show mostly nonlinear relationships with COPD. Household income showed a tipping point at US $40,000, after which the negative relationship with COPD was nearly linear.

Ethnic or racial variables accounted for 3 of the top 7 variables in the GBT: American Indian or Alaska Native percentage (rank 3), Black percentage (rank 6), and Hispanic percentage (rank 7). The greater the size of those minority populations, the lower the COPD rate. Our SHAP plots show significant tipping points (nonlinearities) for American Indian or Alaska Native percentage and Black percentage and a mostly linear relationship for Hispanic percentage. Consistent with the literature, all 3 variables show a strongly negative association with COPD.

The regression and GBT models show that in addition to strongly protective impacts for lower cigarette smoking and higher household income, there are protective impacts for larger American Indian or Alaska Native and Hispanic populations as well as a nonlinear impact on larger Black populations. Higher education rate and lower unemployment rate are also protective, whereas AQI shows mixed effects. These results have implications for private health care practitioners, public health care officials, and health care policy makers who aim to reduce COPD rates. Such policies and programs should not assume high digital literacy [50,51]. System designers could use SMS text messaging, social media, and interactive voice response systems. This would be appropriate for those with lower household income or lower education levels. To design culturally appropriate visual cues and messaging to different racial or ethnic groups, members of the various communities should be included in the design process [52,53]. In sum, the user interface should exhibit high ease-of-use—using gamification, storytelling, and peer support—consistent with cultural norms.

Several studies have identified ethnic and racial disparities in COPD prevalence and risk among smokers. One study found that racial and ethnic minority individuals, particularly African Americans and Hispanics, had a lower prevalence of airflow obstruction than non-Hispanic White individuals, even after adjusting for smoking status and other risk factors [54]. This finding was supported by another study that observed lower COPD risk in ethnic minority groups compared to White individuals, despite similar smoking intensities [55]. A larger minority population means a larger peer support network for prevention and cessation of smoking and a larger peer community to recommend COPD screening, diagnosis, and treatment, which is particularly useful in a health care system that has implicit racial or ethnic bias [50,56].

There are varying levels of patient trust and implicit bias in health care practitioners themselves [57], which contributes to health outcome differences. From a population communication perspective, messaging regarding the risks of COPD—particularly the avoidance or cessation of cigarette smoking—should be sensitive to community context, engaging trusted local authorities to optimize the chances of patient engagement [58]. Health care practitioners could partner with trusted local authorities and community leaders regarding smoking prevention and cessation as well as respiratory health in general to decrease COPD risk. Health care practitioners and educators should communicate to different populations in culturally sensitive ways [59,60].

Educational materials and behavior change strategies may need to be customized according to different risk factors, beliefs, preferences, and technographics of different subpopulations [50,51]. On a basic level, people with lower levels of education or household income could be directed via phone geolocation to their local health care and to their community leaders for in-person guidance and support. Local leaders could then inform them about local smoking cessation programs and apps or websites that monitor air quality in their community. Trusted local authorities are helpful entry points in those communities, after which peer support and network effects spread the information.

AQI was not significant in the MLR, but it was significant in the GBT, albeit not as strongly as we expected. It could be that the AQI is more of a diffuse, macrolevel environmental factor that fluctuates over time, making some CBSAs worse on average, but with wide volatility, for example, as weather and wind directions change [61,62]. Therefore, AQI could have more of an indirect or interaction effect with other variables. Combining campaigns on smoking prevention with campaigns on air quality could create a holistic public health strategy, particularly—as our findings suggest—in communities considered vulnerable, that is, communities with lower education, higher unemployment, and lower household income. Subsidies for households in communities considered vulnerable to convert to more efficient, cleaner home heating and cooling methods would improve their home’s air quality at a lower cost [63]. Research suggests that engaging communities in targeting their air quality issues can lead to more positive outcomes in both air quality and public health [64-66].

There is a small but growing body of research that uses ML models in health care and medicine. There is recognition that the models can be highly accurate, but there is no consensus yet on how to interpret the results in a way that meshes seamlessly with clinical practice. The following examples provide an overview.

Elshawi et al [67] compared model-agnostic explanations using 2 techniques, LIME and Shapley values, to interpret a ML model for predicting hypertension risk. LIME uses small subsets of the data, which may be idiosyncratic, to provide intuitive explanations, that is, rules. Shapley values are more theoretically sound and global, using all the available data, and are, therefore, less idiosyncratic than LIME, but they do not provide LIME’s simple, linear explanations [67].

Hakkoum et al [68] conducted an extensive literature review of ML interpretability in medicine published between 1994 and 2020. The review found that there was no consensus on evaluation metrics or frameworks to assess the quality and utility of the interpretability methods [68]. The highest performing ML models did not translate easily into clinical rules.

Meng et al [69] reviewed the interpretability and fairness evaluation of deep learning models on MIMIC-IV data set, a large, publicly available benchmark for developing and evaluating the interpretability of high-performing ML models that use sensitive demographic features. The review found that existing interpretation methods, for example, variable importance rankings, provide partial explanations without fully elucidating the model’s complex decision logic.

In sum, there is no consensus on the best way to interpret high-performing ML models in health care. There are always trade-offs between accuracy and interpretability or explainability. We chose to use Shapley values because they represent the frontier in explainability, and they are similar to interpreting a multiple regression, interpreting 1 variable at a time, without the assumptions of linear models. In addition, Shapley values allow for nonlinear relationships between each independent (predictor) variable and the dependent variable. Variable importance plots in conjunction with Shapley values help us to identify the most important variables and characterize their relationships with COPD.

Our best MLR model had variance explained of 81.1%, MAE of 0.591, and SMAPE of 9.666. Our best ML model was the GBT, with variance explained of 85.7%, MAE of 0.456, and SMAPE of 6.956. The GBT explains most of the variance—4.6% more than the best MLR—with far less predictive error. The GBT’s SMAPE (6.956) was 28% lower than that of the MLR’s SMAPE (9.666). Similarly, the GBT’s RMSE was 26% lower than the MLR’s RMSE, and its MAE was 23% lower than that of the MLR. Real-world predictive accuracy should be similar to that found in the test data set because the test data were never used in the GBT’s model development.

Our GBT performed strongly on the test data, with very little performance deterioration on the test data versus performance on the training and validation data. This demonstrates that the GBT model does not overfit the data. To interpret the GBT, we used a variable importance plot [34,70,71] and SHAP plots [72,73]. SHAP plots are useful for interpreting the strength of the pairwise relationships between predictor variable and COPD rate, showing the added nuances of the curvilinear plots. By doing so, we rendered transparent the “black-box model” [74-76], thus preserving interpretability and actionability, in addition to adding nonlinear nuance.

This research has a few limitations. The data were obtained from 517 (56%) of the 929 CBSAs. We assumed that this was an adequate sample and that the remaining CBSAs that did not report the data were similar to those that did. Alternatively, it could be that the CBSA that did not report COPD rates did so because the rates were low, that is, COPD was not considered a major problem by the local public health officials. Data covering additional demographic variables, such as gender and age, in addition to occupational exposures and physical exercise, could be gathered [77-79]. Future research could develop separate models stratified by demographic variables such as race or ethnicity, assuming there are sufficient data for each categorical class. There could also be geopolitical variations in terms of population density as well as demographics, psychographics [80], and technographics [81,82].

Future data collection could focus on understanding racial or ethnic disparities. By collecting data more intensively from the minority populations, we could go deeper into understanding how their rates of COPD drop so dramatically. Is it related to active peer recommendations for better self-care in a predominantly White health care system and population? Is it related to successfully tailored smoking prevention or cessation programs? Data pertaining to answering these more specific questions could be collected to enhance our understanding of how best to tailor communications to different demographic or ethnographic groups.

All our models were structured as direct effects. We applied MLR and ML methods with data from CBSAs, which have significant variation in terms of health care access and quality. Using these models as a foundation, we should recognize the interconnectedness (ie, direct, indirect, and interactive) of pollutants and copollutants to fully understand COPD’s complex etiology. Future research could model interaction, moderating, or mediating effects, perhaps with a structural equation model, to identify the direct and indirect effects of COPD, for example, showing how asthma may lead to COPD or to asthma-COPD overlap [77].

There are many research knowledge gaps in the health impacts of extreme air pollution, including the effects of interactions between temperature and air pollution on respiratory health due to climate change [83]. Future research directions could focus on modeling the direct and indirect links between environmental exposures and COPD. On the basis of those results, we could design interventions, such as air quality warning systems, to mitigate their impact. The findings would underscore the opportunities for public health regulations, public-private sector partnerships, private company entrepreneurship, and global initiatives to reduce environmental exposures.

Greenhouse gas emissions may exacerbate overall air quality [84-88], contributing indirectly to COPD. Future research could collect data on new, additional variables pertaining to climate change [89]. Wildfires, which are increasingly common, produce more carcinogens in the air, including high levels of particulate matter. This can directly decrease air quality or copollute with other ambient pollutants [90]. These problems have been shown to increase the odds of lung cancer [91], and it is plausible that they can also contribute to COPD.

The association between COPD and environmental pollutants, including tropospheric ozone, nitrogen dioxide, sulfur dioxide, and occupational exposures, has been extensively investigated [8,91-94]. Coarse, fine, and ultrafine particulate matter have been studied extensively and linked to systemic oxidative stress, inflammation [95], atherosclerosis [96], and mortality [97] in the United States [98,99] and China [100-102]. Tropospheric ozone exposure by itself has been linked to impaired lung function and increased COPD-related hospital admissions [103-105]. Similarly, elevated levels of nitrogen dioxide and sulfur dioxide, which are common in cities and industrial work sites, have been linked to an increased risk of COPD in the general population [106,107] and older adults [108]. In sum, data pertaining to ambient pollution, for example, particulate matter, sulfur dioxide, and carbon monoxide, could be useful additional copollutant data to include in future models [6,86-88,91,109-111].

Our novel contributions in this paper include the following: (1) integration of multiple publicly available CDC data sources, (2) development of highly accurate models using linear and nonlinear methods, and (3) interpretation of the variable impacts for the best model. Smoking was the number 1 variable impacting the COPD rate, which was expected. Household income was the second most influential predictor variable. Four economic factors spanned the full range of influence, from large (household income) to moderate (education rate) to small (unemployment rate and GDP). The race or ethnicity variable also had a range of impacts, from moderately high (American Indian or Alaska Native percentage) to moderate (Black or Hispanic percentage) to small (White, Asian, or Native Hawaiian or Pacific Islander percentage).

This research demonstrates the power of ML methods in general and a GBT, which produced a highly accurate model of COPD rates. The computational complexity of a GBT enables it to obtain high accuracy, but health care policy makers may be reluctant to adopt it unless they can obtain a rule-based explanation. Furthermore, clinicians typically want to be able to explain, justify, and communicate results to others in an intuitive manner. Finally, there may be legal, auditing, or regulatory requirements concerning transparency. If the method is audited, and it cannot be clearly explained, there may be serious legal or financial consequences [72]. Consequently, it is important to have explainable models to open the “black box,” rendering them interpretable and actionable [75,76]. This research shows that it is possible to do so.