Your Neighborhood Matters: A Machine-Learning Approach to the Geospatial and Social Determinants of Health in 9-1-1 Activated Chest Pain

Abstract

Healthcare disparities in the initial management of patients with acute coronary syndrome (ACS) exist. Yet, the complexity of interactions between demographic, social, economic, and geospatial determinants of health hinders incorporating such predictors in existing risk stratification models. We sought to explore a machine-learning-based approach to study the complex interactions between the geospatial and social determinants of health to explain disparities in ACS likelihood in an urban community.

This study identified consecutive patients transported by Pittsburgh EMS for a chief complaint of chest pain or ACS-equivalent symptoms. We extracted demographics, clinical data, and location coordinates from electronic health records. Median income was based on US census data by zip code. A random forest classifier and a regularized logistic regression model were used to identify the most important predictors of ACS likelihood.

Our final sample included 2,400 patients (age 59±17 years, 47% Females, 41% Blacks, 15.8% adjudicated ACS). In our random forest model (AUC of 0.71±0.03) age, prior revascularization, income, distance from hospital, and residential neighborhood were the most important predictors of ACS likelihood. In regularized regression (AIC = 1843, BIC = 1912, chi square = 193, df = 10, p < 0.001), residential neighborhood remained a significant and independent predictor of ACS likelihood.

Findings from our study suggest that residential neighborhood constitutes an upstream factor to explain the observed healthcare disparity in ACS risk prediction, independent from known demographic, social, and economic determinants of health, which can inform future work on ACS prevention, in-hospital care, and patient discharge.

INTRODUCTION

Healthcare disparities in the prehospital and in-hospital management of patients with symptomatic coronary artery disease (CAD) based on gender, race, distance from hospital, and the system providing care are well described in literature (Barnato, Lucas, Staiger, Wennberg, & Chandra, 2005; Coppler, Elmer, Rittenberger, Callaway, & Wallace, 2018; Davies & Rier, 2018; Ladapo et al., 2017; Meadows et al., 2011; Nguyen, Berger, Duval, & Luepker, 2008). Such disparities have become even more pronounced during the COVID-19 pandemic (Rashid et al., 2021). While most studies have focused on disparities attributable to non-modifiable demographic factors such as sex and race (Davies & Rier, 2018; Meadows et al., 2011); other studies have explored the role of more dynamic social determinants of health (SDOH), such as distance to a hospital and access to a primary care provider (Coppler et al., 2018; Schultz et al., 2018). However, while SDOH are complex, they are often represented solely by socioeconomic indicators, such as income and education. Such indicators are heavily intercorrelated and suffer from multi-collinearity. Moreover, SDOH indices often overlook the geographic associations of a given population or the factors uniquely characterizing a specific neighborhood, such as inner-city communities, which often complicate the unidimensional application of SDOH. Thus, the use of multidimensional geospatial approaches to quantify SDOH rather than the use of a singular entity (i.e., income) may better capture the complexity underlying these determinants (Kolak, Bhatt, Park, Padron, & Molefe, 2020). Yet, few studies have explored the role of geospatial factors to predict patient outcomes or to optimize emergency medical service (EMS) resource utilization (Concannon et al., 2008; Starks et al., 2017; Stassen, Olsson, & Kurland, 2020). Geospatial factors refer to data points with geographic coordinates, such as census and location data, analyzed in context with traditional data information (A. Moser, Day, Bruner, & Day, 2017). Given the complexity of exploring interactions between geospatial, demographic, social, and economic factors in relevance to dire outcomes in CAD, incorporating these determinants of health in existing risk stratification models remains a challenge. This is especially important in patients calling 9–1-1 for suspected acute coronary syndrome (ACS) due to the acuity of illness and the urgency of time-sensitive therapeutics in this vulnerable population.

Novel machine learning-based approaches provide an opportunity to discover and control for otherwise unknown interactions in highly dimensional and complex datasets. Specifically, random forests (RF) have been shown to be superior to multivariate regression in handling interactions, non-linearity, and multicollinearity with low risk of overfitting (Breiman, 2001). Machine learning classifiers can compute the relative importance, or contribution, of each predictor to the classification task of a given outcome. Similar approaches, including RF, were successfully applied in cardiovascular research to identify factors contributing to morbidity and mortality in heart failure patients (Hsich, Gorodeski, Blackstone, Ishwaran, & Lauer, 2011) and to stratify risk for cardiovascular disease at a population level (Alaa, Bolton, Di Angelantonio, Rudd, & van der Schaar, 2019; Yang et al., 2020). Similarly, an array of machine learning models including random forest, gradient boosting machine, and naïve Bayes, among others, were used to both identify and risk stratify acute coronary syndrome (ACS) patients with high accuracy (S. Al-Zaiti et al., 2020; D’Ascenzo et al., 2021). The success of advanced machine learning analyses in cardiovascular studies provides an impetus to further employ these models in the analysis of the acute care of ACS. Thus, in this study, we sought to explore a machine-learning-based approach to study the geospatial and SDOH on disparity in ACS likelihood in consecutive patients calling 9–1-1 for acute chest pain or equivalent symptoms in an urban setting.

METHODS

Sample & Settings

Data for this secondary analysis comes from the parent EMPIRE study (ECG Methods for the Prompt Identification of Coronary Events) (S. Al-Zaiti et al., 2020; S. S. Al-Zaiti, Martin-Gill, Sejdic, Alrawashdeh, & Callaway, 2015). The original study extracted a database of patients with chest pain who were transferred via the City of Pittsburgh Bureau of EMS (Pittsburgh EMS) to three University of Pittsburgh Medical Center-affiliated tertiary care hospitals (UPMC Presbyterian, Mercy, and Shadyside hospitals) in Pittsburgh, PA, between the years 2013 and 2018. A total of 2,400 eligible patients were identified and extracted from a database for the purposes of the parent study. The Greater Pittsburgh Region is located in Western Pennsylvania with a population of 2,333,000 while the population of The City of Pittsburgh is approximately 300,000, with a median age of 33.3 years. The City’s population is composed of 66% White, 25% Black, 5% Asian, and 3% of residents who identify as of two or more races. This study was approved by the Institutional Review Board under a waiver of informed consent.

Data Collection

The parent study was based on a list of patients with a complaint of chest pain that we extracted from the Pittsburgh EMS electronic record system representing 12 dispatch stations around the city. This list underwent probabilistic matching process with transferred 12-lead ECG to the three study hospitals based on age, gender, and time window of patient EMS transfer. Only patients with a record of prehospital ECG transfer were identified in this study and given a study ID. Prehospital data for these responses were extracted from the EMS electronic health records (EHR) including clinical data, geographical coordinates of the response location, time, and distance of transfer. Next, in-hospital data elements were manually abstracted from the hospital EHR as defined by the American College of Cardiology (Cannon et al., 2013): demographics, past medical history, home medications, clinical presentation and course of hospitalization, laboratory tests, imaging studies, cardiac catheterization, treatments, and in-hospital complications. Data about median household income for US geographical zip codes were retrieved from the U.S. Census Bureau (2017) (U.S. Census Bureau, 2017). For this analysis, we focused on 16 predictors typically available during prehospital phase of care, including patient demographics (age and sex), clinical characteristics (history of hypertension, diabetes, COPD, smoking, heart failure, known coronary artery disease, old myocardial infarction, and prior revascularization), SDOH indicators (race and income), geospatial-level data (distance from hospital and residential neighborhood), clinical presentation (duration of symptoms), and EMS system-level metrics (total prehospital response time). Age, income, distance, duration of symptoms, and response time were continuous, and all other variables were categorical.

Study Outcome

The primary outcome of the study was the diagnosis of ACS during the primary indexed visit, as defined by the American Heart Association (AHA) Fourth Universal Definition of myocardial infarction, which includes acute myocardial infarction and unstable angina (Thygesen et al., 2018). The consensus statement established the diagnosis of ACS based on symptoms suggestive of myocardial ischemia on presentation (i.e., diffuse discomfort in the chest, upper extremity, jaw, or epigastric area for more than 20 minutes), with the presence of biomarker, nuclear, or angiographic evidence of myocardial ischemia and/or loss of viable myocardium. The study outcome was adjudicated by two independent physicians who comprehensively reviewed all medical charts while blinded from the final purposes of the study. Any disagreements regarding the outcomes were resolved by a third physician reviewer.

Data Analysis

Descriptive analyses were used to summarize the data using mean and standard deviation for continuous variables and frequencies and percentages for categorical variables. The likelihood of ACS was compared across five defined residential areas of the city (central, west, east, south, and north). Residential area descriptive census data are presented in Table 1. Groups were compared using chi-square for categorical data or analysis of variance for continuous data.

TABLE 1.

US Census Data for Designated Residential Areas in the City of Pittsburgh Explored in the Study

Characteristic	Central	South	East	West	North
Median Age (years)	27	39	39	40	41
Annual Median Income (k/year)	49	57	54	50	62
Race (% in total of residential area)
American Indian or Alaska Native	0.3%	0.1%	0.2%	0.2%	0.2%
Asian	8.7%	2.5%	4.1%	3.9%	1.7%
Black	26.9%	8.8%	29.8%	21.8%	18.3%
White	59.9%	85.9%	63.0%	71.6%	77.1%
Two or More Races	4.2%	2.7%	2.9%	2.5%	2.2%
Hispanic Ethnicity (% in total of residential area)	4.0%	1.9%	2.2%	2.2%	1.7%
Female Sex	49.9%	51.7%	53.7%	54.2%	49.0%

To account for the complex multidimensional interactions between study predictors, we developed a random forest (RF) classifier to screen for the most important predictors of ACS in our cohort (70% training with 10-fold cross validation on training and 30% testing). As per recommendations, 70% of patients (n=1680) were used to develop the model (Kagiyama, Shrestha, Farjo, & Sengupta, 2019; Xu & Goodacre, 2018). We further used 10-fold cross validation (i.e., use 10 random partitions to the data to iteratively estimate and adjust the model coefficients) to fine tune the model parameters and estimate uncertainty in the model. The final model was tested on the remaining 30% (n=720) of patients to obtain the final performance metrics. Random forests are well known machine learning models that address classification problems and have been extensively used in research (Breiman, 2001; Hsich et al., 2011; Navarini et al., 2020). Briefly, RF is an ensemble technique that uses decision trees as a base classifier. Decision trees place patients into nodes according to sequential splits by the most influential predictors in descending order, yielding branches and leaves that differ in depth and complexity from one tree to another. Each tree eventually yields a single classification for each patient (e.g., ACS vs. no ACS) according to a series of yes/no rules, which is intuitive and follows clinical reasoning. A RF combines the output of hundreds or thousands of individual decision trees based on simple vote count to provide a final classification. To make sure individual decision trees do not constantly all err in the same direction, each tree is built on a random subset of patients using a random subset of predictors, which generates a portfolio of trees that is greater than the sum of its parts. To fine tune RF hyperparameters (i.e., number of trees, node size, tree depth, etc.), we used an automated grid search algorithm to find optimal parameters and used input from clinicians to optimize the combination of various hyperparameters.

A unique attribute of RF is that it can estimate the permutation of a single predictor when it is randomly dropped from the model, which is indicative of how much the model depends on this single predictor (i.e., permutation feature importance). Using this importance ranking, we have selected a parsimonious subset of ACS predictors that are relatively important for ACS prediction. To compute a clinically useful model for estimating ACS likelihood, we entered these important features in a multivariate logistic regression model to estimate final odds ratio. Data were analyzed using Python 3.7 and SPSS version 24. The significance level was set at 0.05 for two-tailed hypothesis testing.

RESULTS

Sample Characteristics

The sample included 2,400 patients aged 59 ± 17 years, 47% females, 41% Black, and <1% Hispanic. Table 2 summarizes in detail the socio-demographic and clinical characteristics of the study sample. The median household income of neighborhood of residence was $49±19 k/year, and around 20% of patients were either uninsured or had government-issued Medicaid. On average, comorbidities were ubiquitous in this sample, with hypertension prevalent in 2 out of 3 patients and diabetes, smoking, coronary artery disease, and prior revascularizations prevalent in 1 out of 3 patients. The median duration of symptoms was 2.5 hours, and the total prehospital response time (9–1-1-call to hospital arrival) was 45 minutes. The vast majority (75%) of patients were admitted to the hospital (median length of stay was 1 day) and approximately 16% had a final confirmed diagnosis of ACS as the etiology of their chest pain.

TABLE 2.

Baseline Demographic and Clinical Characteristics of the Sample

Characteristic	All Patients (n=2,400)
Socio-Demographics
Age (years)	59 ± 17
Sex (female)	1119 (47%)
Race (Black)	988 (41%)
Ethnicity (Hispanic)	8 (0.3%)
Annual Median Income [US Census] (k/year)	49 ± 19 (22−98)
Uninsured / Medicaid [US Census]	477 (20%)
Distance from Hospital (miles)	2.2 ± 1.3 (0.1−7.2)
Residential Neighborhood
Central	591 (25%)
South	717 (30%)
East	704 (29%)
West	102 (4%)
North	207 (9%)
Past Medical History
Hypertension	1684 (70%)
Diabetes	689 (29%)
Chronic Obstructive Pulmonary Disease	566 (24%)
Current Smoking	743 (31%)
Coronary Artery Disease	851 (36%)
Old Myocardial Infarction	627 (26%)
Prior Coronary Revascularization	665 (28%)
Chronic Heart Failure	433 (18%)
Prehospital Data
Duration of Symptom (hours)	2.5 [1.0 – 8.0]
Prehospital Response Time (min)	45 ± 12 (15 – 120)
Cath Lab Activation (STEMI)	70 (3%)
Transport Destination
UPMC Presbyterian Hospital	475 (20%)
UPMC Shadyside Hospital	721 (30%)
UPMC Mercy Hospital	1204 (50%)
In-hospital Data
Length of stay (days)	1.1 [0.3 – 2.7]
Admitted to Hospital	1807 (75%)
Confirmed ACS	378 (15.8%)

Values are mean ± SD; or n (%). Abbreviations: Min: Minute; STEMI: ST-Segment Elevation Myocardial Infarction; ACS: Acute Coronary Syndrome.

Geospatial Distribution

Figure 1 shows the geospatial distribution of EMS-attended prehospital chest pain encounters in the city of Pittsburgh. The density of distributions of these encounters (blue dots, Fig. 1A) follows the residential areas around the unique geography of the city: central, south, north, east, and west areas, which are all separated by three regional rivers. To better understand the geospatial distribution of ACS events in the data (red dots, Fig. 1A), we used the sheer number of encounters in each area as a denominator to compute the likelihood of an EMS encounter being an ACS event based on a given residential area (Figure 1B). As seen in this figure, the likelihood of ACS ranged from 11.7% to 19.1%, with significant variations in distributions across these residential areas (chi square 14.15, df = 4, p = 0.007).

Figure 1: Geospatial Distribution of EMS-Attended Chest Pain in the City of Pittsburgh.

This figure shows the geospatial distribution of incidents of EMS-attended chest pain encounters in the City of Pittsburgh. Panel A shows the density of distributions of indexed encounters, with red dots resembling ACS cases and blue dots resembling non-ACS controls. To assure anonymity, dots were jittered to cover areas larger than single households. Panel B shows the likelihood of ACS as per each residential neighborhood. The likelihood was computed as number of ACS cases divided by the sheer number of encounters in that neighborhood. P value was based on Chi-square test.

Determinants of ACS Likelihood

Using the 16 predictors under study, the area under the receiver operator characteristics (AUROC) curve for the final RF model to classify ACS was 0.71±0.03. An ROC-optimized cutoff yielded sensitivity 69%, specificity 64%, positive predictive value 25%, and negative predictive value 92%. These metrics suggest a good classification performance and potential role of these predictors as a screening tools during initial encounter.

Figure 2 shows the variable permutation importance rank of the predictors used in the model. Older age and history of prior revascularization were the most important determinants for ACS likelihood classification, followed immediately by income, distance from the hospital, and residential area. Sex and race were also relatively important predictors of ACS likelihood, whereas existing comorbidities like hypertension, diabetes, smoking, old myocardial infarction, and known heart failure were the least important for ACS classification. After dropping these 5 predictors, we entered the remaining 11 important predictors in a multivariate logistic regression model of which 8 remained as significant predictors (AIC = 1843, BIC = 1912, chi square = 193, df = 10, p < 0.001). Figure 3 shows the z-score standardized odds ratio with ± 2 standard error bars of these significant and important predictors. Older age, prior revascularization, white race, male sex, and residing in south area were associated with excess ACS likelihood, whereas having COPD, longer duration of symptoms, and longer total EMS response time were associated with less likelihood of ACS. The common result of both the RF model and the logistic regression model revealed that sex, race, and residential area were important and significant social and geospatial determinants of ACS in this population.

Figure 2: Variable Importance Rank for Classifying Confirmed ACS events.

The figure shows the variable permutation importance rank of the variables used in the random forest model to classify ACS likelihood. Abbreviations: CAD: Coronary Artery Disease; COPD: Chronic Obstructive Pulmonary Disease; MI: Myocardial Infarction

Figure 3: Regularized Logistic Regression for Predicting ACS events.

This figure shows the z-score standardized odds ratio with ± 2 standard error bars of predictors retained in the final model (AIC = 1843, BIC = 1912, chi square = 193, df = 10, p < 0.001, pseudo R square = 0.096).

DISCUSSION

This study sought to explore the role of, and interaction between, geospatial and SDOH factors in predicting the likelihood of ACS in EMS-attended chest pain encounters in the City of Pittsburgh. Using a diverse middle-age population, we show that age, significant coronary history, sex, race, and residential neighborhood are the most important and independent predictors of ACS likelihood in this population, controlling for chronic comorbidities and other risk factors. Income and distance from the hospital were also found to be very important predictors but were likely accounted for by the residential neighborhood. Our findings support the growing evidence that health disparities in patients with symptomatic coronary disease are strongly correlated with geospatial determinants of health, which has important implications for EMS resource utilization and regionalization of advanced cardiac care (Stassen et al., 2020).

The two most important predictors for ACS prediction were older age and a previous history of coronary revascularization. This highlights, that while multiple factors can help predict an ongoing ACS, older age, and a history of revascularization overwhelmingly remain the most important factors to assess for during initial patient evaluation. This has been historically well established, with all chest pain risk stratification scores universally incorporating age and a past history of revascularization as major independent predictors of ACS in their parsimonious clinical decision aids (Poldervaart et al., 2017).

The role of residential area was evident in the likelihood of having ACS despite controlling for patients’ characteristics and area income. While our results support the notion that income and distance from the hospital are well known determinants for the outcomes in the residents of a neighborhood (Agarwal, Garg, Parashar, Jaber, & Menon, 2014), some other unmeasured variables of a living space can influence the eventual outcome in chest pain patients. For instance, variables such as walkability of a neighborhood can influence cardiac outcomes (Howell, Tu, Moineddin, Chu, & Booth, 2019), and the relationship between walkability and income is in the City of Pittsburgh widely varies (Bereitschaft, 2019). Few previous studies explored the influence of residential neighborhoods in the City of Pittsburgh on health outcomes. This includes an analysis of opioid overdose patterns in urban Pittsburgh, highlighting the southern part as one of the most affected areas of the city (Dodson, Enki Yoo, Martin-Gill, & Roth, 2018). Our study found similar results, with the southern part of the city having the highest likelihood for ACS. Previous studies reported a similar neighborhood disparities in conditions such as out-of-hospital cardiac arrest (Chan, McNally, Vellano, Tang, & Spertus, 2020), incidence of congenital heart disease (Peyvandi et al., 2020), and 1-year mortality following admissions for CVD (Villanueva & Aggarwal, 2013). The repeated emergence of neighborhoods as an independent predictor of adverse events across multiple studies further highlights that neighborhood-level predictors of health should often be examined at a population-level. Furthermore, these findings emphasize that neighborhood level variables should be adjusted for in predictive models, in addition to the individual level indicators that are often examined. For instance, although income and race are strong SDOH, they did not account for the dimensionality of neighborhood-level metrics, where a wide range of incomes and demographics share a similar disparity encompassed by the shared characteristics of their living spaces, which transcends beyond a patient’s residential address (Bhatnagar, 2017; Kolak et al., 2020).

Interestingly, race played an important role in predicting outcomes in our study, with White patients being more likely to present with ACS. Among the plausible reasons for this finding include the higher likelihood for White patients to use EMS for acute MI, while among Black patients, a higher proportion utilize it for non-ACS causes of acute chest pain, such as congestive heart failure exacerbations (Hanchate et al., 2019). Similar findings were previously reported in stroke patients utilizing EMS, with White women being more likely to use EMS compared to Black women (Mochari-Greenberger et al., 2015). It is worth noting that the relationship between race and outcomes remain heavily confounded by socioeconomic status (Bucholz, Ma, Normand, & Krumholz, 2015), neighborhood characteristics, perceived racism, environmental exposures, access to health care, and other SDOH (Mensah, 2019). Therefore, while our study is designed to account for interactions between predictors, the effect of residual confounding cannot be discounted, and while race remains an independent predictor, it is a multi-facet theme that encompasses factors beyond race despite all efforts to control for confounding. Future work should aim to investigate the reason for the consistent difference in EMS utilization for ACS (Frisch et al., 2019; Mathews et al., 2011; D. K. Moser et al., 2006) using a mixed methods approach to further identify the underpinnings of such known disparities.

Clinical and Policy Implications

Our study has several nursing implications for hospitals that care for patients with ACS. Patients with ACS are high consumers of health care resources during both their acute event and their long-term care. In the acute care phase of health care delivery, nursing staff ratios and hospital bed availability are of utmost importance to consider. These numbers are evaluated at regular intervals by nursing administration and unit level management during every nursing shift. For hospitals at locations caring for a high density of complex ACS patients, this level of planning can be even more challenging, as the need for expert nurses increases. Nurses with expertise in high alert medication administration, invasive procedures, life-sustaining medical equipment management, and complex discharge planning should be staffed appropriately at locations with high rates of these events. Importantly, nurses are also extremely influential in discharge teaching, hospital readmission prevention, and long-term health optimization. These findings should inform nurses that while advising about preventative measures, such as exercising and a heart healthy diet, are generally important, knowing where patients live is equally important in providing an individualized long term care plan. Residential areas differ in terms of ease of access to primary health care providers, walkability, access to healthy foods, literacy about symptoms of an ACS, and access to EMS in case of a perceived health emergency. For instance, a recent study showed that neighborhood-related metrics, such as rate of violent crime and perceived neighborhood safety, are strong predictors of missed medical appointments in a large urban society (Chou et al., 2021). Such disparities in neighborhood health care accessibility should be well known by nurses and considered throughout the continuum of care for ACS.

Moreover, our study has important policy implications for EMS planning and accessibility evaluation. Perceptions towards when to utilize EMS and recognition of proper use of ACS differs across communities (Fong, Anantharaman, Lim, Leong, & Pokkan, 2001; Ong, Ang, Chan, & Yap, 2004) and should be addressed on a neighborhood level basis by EMS organizations. Considering that national guidelines recommend calling 9–1-1 for symptoms of ACS, qualitative work needs to be done across neighborhoods, genders, races, and socioeconomic status to identify lived experiences and possible barriers that lead to differences in utilizing EMS in case of acute chest pain. For instance, it is well established that Black patients’ previous experiences with discrimination and perceived racism pose a major barrier to accessing care (Carnethon et al., 2017). It has been recently noted that implementing patient level interventions have improved Door to CT time in ischemic stroke for Hispanic patients, yet no similar improvement has been observed in Black patients (Polineni et al., 2021), indicating the more patient level interventions are needed. Finally, considering the level of prehospital care required for ACS patients, EMS agencies should identify neighborhoods with high likelihood of ACS to ensure proper staffing of EMS personnel with advanced life support capabilities and to optimize locations of EMS agencies accordingly. Furthermore, our work in identifying the most important predictors would inform triage by EMS paramedics in the field regarding who and where they should have the highest index of suspicion for a diagnosis of ACS.

Study Limitations

This study has a few limitations. First, the racial demographics of Pittsburgh do not include underrepresented minorities other than Black, hence race in our study was categorized as a dichotomous variable. Second, our geospatial analysis was based on location coordinates and Zip codes of EMS pickup location, not necessarily the exact home address. It is plausible to assume the activity space where the encounter happened is within patient’s neighborhood. Yet, it is possible that some of these EMS-attended chest pain encounters also occurred outside the patient’s neighborhood. Third, median income was based on census data for a given zip code, not necessarily corresponding to a patient’s individual level of income. Furthermore, it is worth noting that the Pittsburgh EMS system includes 12 dispatch stations and serves both UPMC and non-UPMC affiliated hospitals. Although our study examined consecutive patients calling 9–1-1, our study population included only patients who were eventually transported to a UPMC facility. As an estimate, an EMS log inquiry for transports during the study period shows that 69% (95% CI 61%−77%) of chest pain-activated encounters were transported to a UPMC facility. It is possible that sicker patients in a certain neighborhood were disproportionately transported to a UPMC facility and vice versa, which might constitute a potential sampling bias. Finally, considering the study recruited a population from a single city within a specific healthcare system, we are unable to generalize our findings, including identified predictors, to other populations. Specifically, any biases existing in the study data will be further propagated during the training phase of the machine learning model, limiting the study’s external validity.

CONCLUSION

Using a diverse middle-age population, this study shows that age, significant coronary history, race, and residential neighborhood are the most important and independent predictors of likelihood of ACS in patients transported by EMS for chest pain in a large Urban community, controlling for chronic comorbidities and other risk factors. Our study suggests that residential neighborhood constitutes an upstream factor to explain the observed healthcare disparity across the city, independent from known demographic, social, and economic determinants of health. Future work should explore the role of neighborhood characteristics in predisposing individuals to higher risk of disease and consider such characteristics throughout the continuum of care for ACS prevention, in-hospital care, and patient discharge.