Interoperability among hospitals treating populations that have been marginalized

Abstract

Objective

To test whether differences in hospital interoperability are related to the extent to which hospitals treat groups that have been economically and socially marginalized.

Data Sources and Study Setting

Data on 2393 non‐federal acute care hospitals in the United States from the American Hospital Association Information Technology Supplement fielded in 2021, the 2019 Medicare Cost Report, and the 2019 Social Deprivation Index.

Study Design

Cross‐sectional analysis.

Data Collection/Extraction Methods

We identified five proxy measures related to marginalization and assessed the relationship between those measures and the likelihood that hospitals engaged in all four domains of interoperable information exchange and participated in national interoperability networks in cross‐sectional analysis.

Principal Findings

In unadjusted analysis, hospitals that treated patients from zip codes with high social deprivation were 33% less likely to engage in interoperable exchange (Relative Risk = 0.67, 95% CI: 0.58–0.76) and 24% less likely to participate in a national network than all other hospitals (RR = 0.76; 95% CI: 0.66–0.87). Critical Access Hospitals (CAH) were 24 percent less likely to engage in interoperable exchange (RR = 0.76; 95% CI: 0.69–0.83) but not less likely to participate in a national network (RR = 0.97; 95% CI: 0.88–1.06). No difference was detected for 2 measures (high Disproportionate Share Hospital percentage and Medicaid case mix) while 1 was associated with a greater likelihood to engage (high uncompensated care burden).

The association between social deprivation and interoperable exchange persisted in an analysis examining metropolitan and rural areas separately and in adjusted analyses accounting for hospital characteristics.

Conclusions

Hospitals that treat patients from areas with high social deprivation were less likely to engage in interoperable exchange than other hospitals, but other measures were not associated with lower interoperability. The use of area deprivation data may be important to monitor and address hospital clinical data interoperability disparities to avoid related health care disparities.

What is known on this topic

• Several hospital characteristics reflecting smaller scale or reduced resources have been shown to relate to lower rates of hospital engagement in the interoperable exchange of health information, an important policy goal.

• Numerous proxy measures exist for describing the patient populations hospitals treat, and measures are not closely correlated, complicating descriptions of the extent that hospitals treat populations that have been marginalized.

What this study adds

• Only one of five proxy measures of the extent to which hospitals treat populations that have been marginalized—the area social deprivation index—was associated with differences in hospital interoperability.

• Differences by social deprivation index persist when separately examining rural and urban hospitals and adjusted analysis, indicating this metric is capturing information beyond geographic divides.

1. INTRODUCTION

Interoperable exchange—the sharing and integration of health information between organizations—provides important benefits to patients. ¹ While hospital engagement in interoperable exchange and participation in national networks has steadily increased over time, ² systematic differences in hospital engagement in interoperability exist. Hospital engagement in interoperability and other advanced usages of electronic health records (EHR) have been found to be lower among smaller, rural, and independent hospitals compared to large, urban, health system‐based hospitals. ³ , ⁴ , ⁵ , ⁶ , ⁷

These gaps may contribute to health disparities if hospitals caring for more patients that have been marginalized are also less likely to be interoperable. Interoperable exchange of patient information is particularly important for these populations, who are at higher risk for fragmented care and poor care coordination. ⁸ , ⁹ For instance, interoperable exchange can ensure that information from recent treatment is available when a patient presents at an emergency department (ED) and can lead to reduced redundant diagnostic procedures, which is likely particularly relevant for populations that rely on the ED for a large proportion of their care. ¹⁰ , ¹¹ Similarly, interoperable exchange can help to ensure that all prescribed medications are available at the point of treatment, allowing for informed clinical decision making and limiting adverse drug events. ¹² Improving the interoperable exchange of patient information, therefore has the potential to reduce harm due to health care information fragmentation across settings experienced by these populations.

Policy approaches to improve interoperability and address health disparities may be more effective if they are guided by an empirical understanding of which hospitals need additional support. ¹³ , ¹⁴ Yet a digital divide related to the patient population that hospitals treat has not been systematically demonstrated. Many empirical approaches exist to identify hospitals that disproportionately treat patients that have been economically and socially marginalized, and each approach identifies a different set of hospitals. ¹⁵ , ¹⁶ , ¹⁷ Without evidence to guide which approach to select, efforts to support engagement in interoperability may not reach hospitals with the greatest need. For example, the use of urban–rural location as a proxy for the treatment of populations that have been marginalized in rural areas may overlook differences in interoperability among hospitals that treat more patients that have been marginalized regardless of location.

Policy makers may rely on proxy measures related to their programs, such as Medicaid caseload, Medicare Disproportionate Share Hospital (DSH) Index, Uncompensated Care burden, and Critical Access Hospital (CAH) designation to target supportive policies. For instance, past policy initiatives have leveraged these measures to reach hospitals including the Centers for Medicare & Medicaid Services (CMS) State Innovation Models and the Office of the National Coordinator for Health Information Technology's regional extension centers program. ¹⁴ , ¹⁵ Most prominently, the Medicaid component of the Electronic Health Record Incentive Program, which ended in 2021, allowed for additional incentive payments to hospitals that served at least 10% of Medicaid case volume. ¹⁶ While these measures capture important information about the insurance status of the patients that hospitals serve, they each identify unique hospitals and may not capture broader social dynamics related to the extent to which the populations hospitals serve have been marginalized.

Beyond programmatic measures, there is growing interest in measures designed to capture population‐level socioeconomic risk for health disparities, avoidable care, and chronic health condition at small area levels. One widely used example of which is the Social Deprivation Index (SDI). The SDI is a composite index of sociodemographic data elements universally available at the census level from U.S. census sources that have been validated to predict a host of health outcomes better than poverty alone, and which are intended to capture the relative deprivation of a geographic area and its residents. ¹⁸ The SDI and similar measures based on the areas in which individuals reside might reflect dimensions of the populations that hospitals serve not well captured by programmatic measures and that may lead hospitals to be less likely to engage in interoperability. Policy makers have begun to use these measures to adjust payments to providers, such as through CMS's Accountable Care Organization Realizing Equity, Access, and Community Health program. ¹⁹

In this study, we examined how different measures of marginalized populations, both programmatic measure proxie and the SDI, are associated with gaps in interoperability, which may inform the potential selection and use of such measures for guiding policies designed to improve interoperability for hospitals serving marginalized populations. Our primary hypothesis was that hospitals that serve more marginalized patients, as measured by established proxies, are less likely to engage in interoperable exchange. We then sought to identify whether the choice of proxy impacted whether we found support for this hypothesis; whether differences persisted when separately examining rural and urban hospitals; and whether differences in interoperability persisted when hospital characteristics previously shown to relate to interoperable capabilities were accounted for.

2. METHODS

2.1. Data sources

We combined data on hospital information technology (IT) adoption from the American Hospital Association (AHA) IT Supplement fielded April–June 2021 with data on hospitals and the populations they serve from four sources. First, we identified information on hospital financial status from the Medicare Hospital Cost Reports. Second, we identified the volume of patients from each ZIP code hospitals treated using the Medicare 2019 Hospital Service Area File (HSAF). Third, the 2019 SDI, received by request from HealthLandscape and the Robert Graham Center, was combined with HSAF data to estimate hospital service area population deprivation. ¹⁸ , ²⁰ The SDI is an index of census variables including income, education, employment status, housing, household characteristics, automobile ownership, and the prevalence of high‐needs populations. Finally, we identified key hospital characteristics from the 2020 American Hospital Association Annual Survey, which was also fielded in 2021.

This study was completed following a Strengthening of the reporting of observational studies in epidemiology (STROBE) checklist for cross‐sectional studies. This study, based on survey data, was determined to be exempt from review by the Office of the National Coordinator for Health IT.

2.2. Interoperable IT

We first measured interoperable exchange as engagement in all four primary domains of information exchange necessary to achieve interoperability as described by the Office of the National Coordinator of Health Information Technology and measured in prior literature. ³ , ²¹ Those four domains include finding, sending, receiving information, and integrating that information into the EHR. These domains are captured on the AHA IT Supplement. We describe how the measure is derived in more detail in the technical appendix (Data [Link], [Link]).

Second, we measured whether hospitals participated in one of three prominent national networks that support interoperability (CommonWell Health Alliance, Carequality, or eHealth Exchange), as measured on the IT survey. These networks have emerged relatively recently as an important method to enable information exchange at scale. ⁴ , ²² , ²³

2.3. Measures of marginalization

2.3.1. Weighted hospital service area social deprivation index

We adopted a geographic retrofitting approach to define geographic areas served by each hospital. ²⁴ For each hospital, we identified the Zone Improvement Plan (ZIP) codes of all beneficiaries the hospital treated and combined this data with the ZIP code tabulation area (ZCTA) SDI. We then calculated the proportion of all cases each hospital treated that were from each ZIP code. Using this proportion, we created a case‐weighted mean of the hospital service area's SDI. We describe this approach, and validation of it compared to data from the Healthcare Cost and Utilization Project in one state, in the technical appendix (Data [Link], [Link]).

2.3.2. Medicaid caseload

Hospitals that treat a greater proportion of Medicaid patients disproportionately serve populations that have been economically marginalized. Using the CMS Cost Report, we identified hospitals' Medicaid Case Load as the percent of each hospital's discharges accounted for by Title XIX (Medicaid) patients divided by the total number of discharges.

2.3.3. Medicare's DSH index

We identified each hospital's DSH index from the CMS Cost Reports. The DSH index is a combination of the proportion of Medicare days comprised by Medicare Supplemental Security Income days and the proportion of all inpatient days comprised by Medicaid, non‐Medicare inpatient days out of all inpatient days, and is calculated for prospective payment system hospitals. ²⁵ The DSH index is used by CMS for a variety of payment policies and is intended to reflect care for groups that have been economically marginalized.

2.3.4. Burden of uncompensated care

Hospitals providing more uncompensated care are likely providing care to groups that have been economically marginalized. As in recent work, we identified the total burden of uncompensated care as the cost of charity care plus non‐Medicare and non‐reimbursable Medicare bad debt divided by the hospital's total operating expenses, identified from the hospital cost report. ¹³ , ¹⁶

2.3.5. Critical access hospital designation

Medicare CAHs meet specific criteria and are eligible for several benefits including cost‐based reimbursement. ²⁶ Generally, CAHs serve small, isolated rural communities that may have low socioeconomic status.

2.4. Analytic approach

We first evaluated the concordance of measures of marginalization by comparing pairwise Spearman rank correlations.

We next examined bivariate associations between measures of marginalization and interoperable exchange. For each measure of marginalization other than CAH status, we divided hospitals into those in the highest 20 percent of marginalization (i.e., the top quintile) and compared them to all other hospitals to assess for differences in interoperable exchange. Because CAH status is a binary determination, we retained the binary indicator alone.

We then divided hospitals into those in metropolitan (urban) versus micropolitan and rural core‐based statistical areas (rural) based on rural–urban commuting areas. We repeated the bivariate tests of association between marginalization measures and interoperable exchange within urban and rural hospital cohorts.

Next, we created multivariate Poisson models to assess the relative risk of interoperable exchange and national network participation based on each measure of marginalization. This approach allows for the direct estimation of relative risk, rather than odds ratios produced using logistic regression. ²⁷ In these models, we included additional hospital characteristics that might be correlated with both treating populations that have been marginalized and IT adoption, following prior studies on health IT and interoperability adoption. ²⁸ , ²⁹ These measures include hospital size (small [<100 beds], medium [100 to <400 beds] and large [400+ beds]), teaching status (non‐teaching, medical school, residency program), ownership (non‐profit, for‐profit or government owned), urban/rural location, multi‐hospital system membership, and whether the hospital was a CAH. The 95% confidence intervals (CI) were estimated using robust standard errors.

We performed a series of robustness tests. First, we replicated our main models except that we included only one measure of marginalization in each model to ensure results were not biased by correlations between measures. Second, we replicated our main models while constraining estimates to hospitals that received a DSH index to evaluate whether changes in coefficients for models including DSH were based on changing sub‐samples or correlations between measures. Third, we replicated our primary model while including each quintile of each measure of marginalization to examine a potential dose–response effect, where hospitals that treat more patients that have been marginalized were increasingly less likely to engage in interoperable exchange. Finally, we included each hospital's primary EHR developer as a covariate in the regression model. EHR developer was excluded from the primary model because of concerns that the choice of EHR was on the causal pathway between measures of marginalization and engagement in interoperable exchange; we, therefore, sought to understand the relationship between these measures and interoperable exchange after accounting for this dynamic.

All analyses included weights derived from the probability that a hospital included in the AHA annual survey—a near census of hospitals—responded to the IT supplement survey based on the hospital characteristics listed above. These weights adjust for any observed difference between respondents and the population from which they are drawn.

3. RESULTS

The analytic sample was comprised of 2393 non‐federal acute care hospitals in the 50 United States and the District of Columbia, representing 53.5% of non‐federal acute care hospitals listed in the AHA Annual Survey. Respondents significantly differed from non‐responding hospitals: they were more often large, members of multihospital systems, and non‐profit (Appendix Table 1).

Correlations across measures of marginalization varied substantially (Table 1), with the highest correlation, 0.52, between the mean service area SDI and Medicare DSH index, and the lowest correlation, −0.10, between Medicare CAH Status and mean service area SDI.

TABLE 1.

Relationship between measures of marginalization.

Panel A. Rank correlations between measures of marginalization of populations hospitals treat
	Mean SDI	Medicaid caseload	Burden of uncompensated care	Medicare DSH index
Medicaid caseload	0.21	‐	‐	‐
Burden of uncompensated care	0.19	0.10	‐	‐
Medicare DSH index	0.52	0.23	0.10	‐
Critical access hospital	−0.10	−0.06	0.16	N/A

Panel B. Number and percent of hospitals in the top quintile of each measure of marginalization
	Mean SDI	Medicaid caseload	Burden of uncompensated care	Medicare DSH index
Medicaid caseload	106 (4.4%)	‐
Burden of uncompensated care	126 (5.3%)	80 (3.3%)
Medicare DSH index	139 (9.4%)	94 (6.3%)	78 (5.3%)	‐
Critical access hospital	70 (2.9%)	96 (4.0%)	160 (6.7%)	N/A

Note: Of the 2393 hospitals included, 1486 were prospective payment system hospitals for which the Medicare DSH index is applicable, therefore correlations and percentage calculations related to the DSH index were limited to these 1486 and excluded CAHs. As a result, the correlation between the Medicare DSH index and CAH is not applicable.

Abbreviations: DSH, disproportionate share hospital; SDI, social deprivation index.

3.1. Differences in hospital characteristics by measures of marginalization

Variation in the hospital characteristics associated with each marginalization measure was considerable, as shown in Table 2, creating different sets of hospitals labeled as serving populations that have been marginalized.

TABLE 2.

Difference in hospital characteristics by measures of marginalization.

	Lower 80 percent SDI (%)	Top quintile SDI (%)	Lower 80 percent DSH (%)	Top quintile DSH (%)	Lower 80 percent Medicaid case percent (%)	Top quintile Medicaid case percent (%)	Lower 80 percent uncomp‐ensated care (%)	Top quintile uncomp‐ensated care (%)
	(n = 2052)	(n = 358)	(n = 1199)	(n = 287)	(n = 1926)	(n = 467)	(n = 1894)	(n = 499)
Size
Small (<100 beds)	52	45 a	35	4 a	52	46 a	49	57 a
Medium (100–399 beds)	37	44 a	52	62 a	38	40	40	34 a
Large (400+ beds)	10	11	13	34 a	10	13 a	11	9
Critical access hospital	28	21 a			29	21 a	25	34 a
Teaching status
Non‐teaching	55	54	43	19 a	57	47 a	53	62 a
Residency program	40	41	52	62 a	39	47 a	42	36 a
Medical school	5	5	6	20 a	5	7	6	2 a
Rural	38	39	29	9 a	36	46 a	38	39
Multi‐hospital system member	67	63	77	77	67	66	63	78 a
Ownership
Non‐profit	62	46 a	68	57 a	60	58	61	55 a
For‐profit	18	26 a	18	24	20	15 a	17	23 a
Government	20	28 a	13	19 a	20	27 a	21	22

Abbreviations: DSH, disproportionate share hospital; SDI, social deprivation index.

^a Difference between top quintile and others was statistically significant at a p‐value of 0.05. Confidence intervals for each difference are available as an appendix (Data [Link], [Link]).

Hospitals in the top quintile of mean SDI differed from other hospitals in ownership status (e.g., were less likely non‐profit (46% [95% CI: 41%–52%] vs. 62% [95% CI: 60%–64%])) and size (e.g., were less likely small in size (45% [95% CI: 40%–51%] vs. 52% [95% CI: 50%–54%])) but were similar in terms of location, system membership and teaching status.

In contrast, hospitals in the top quintile on DSH were more likely over 400 beds (34% [95% CI: 29%–40%] vs. 13% [95% CI: 12%–15%]), more likely major teaching hospitals (20% [95% CI: 16%–24%] vs. 6% [95% CI: 5%–7%]) and less likely rural (9% [95% CI: 6%–13%] vs. 29% [95% CI: 26%–32%]). Hospitals in the top quintile of Medicaid cases were more likely government‐owned, rural, and large, and less likely Critical Access. Hospitals in the top quintile of uncompensated care burden were more likely for‐profit, non‐teaching, multi‐hospital system members, Critical Access, and small.

3.2. Association between marginalization and interoperable exchange

The top quintile SDI hospitals and CAHs were significantly less likely to engage in interoperable exchange and to participate in national networks than were other hospitals. Top quintile SDI hospitals were 33% less likely to participate in interoperable exchange compared to other hospitals (Relative Risk (RR) = 0.67; 95% Confidence Interval: 0.58–0.76; Figure 1) and were 24% less likely to participate in national networks (95% CI: 0.66–0.87). Across the various measures of marginalization, detectable differences in interoperability were greatest when using mean SDI scores. While CAHs were less likely to engage in interoperability (RR = 0.76 relative to non‐CAHs; 95% CI: 0.69–0.83) they were not less likely to participate in national networks.

FIGURE 1.

Relationship between measures of marginalization and hospital interoperability. DSH, disproportionate share hospital; SDI, social deprivation index. N = 2393 for associations that do not include the Medicare DSH index and 1486 for those that do include the Medicare DSH index. The DSH index is not calculated for critical access hospitals. Confidence intervals calculated using heteroskedastic robust standard errors. [Color figure can be viewed at wileyonlinelibrary.com]

Hospitals with high DSH index and a high proportion of Medicaid cases were not less likely to engage in interoperable exchange (RR = 0.92 and 95% CI: 0.83–1.01 in both cases) or participate in national networks. Surprisingly, hospitals with high uncompensated care burden were 20% more likely to engage in interoperable exchange than other hospitals (95% CI: 1.11–1.30); however, they were no more or less likely to participate in national networks.

3.3. Association between marginalization and interoperable exchange among urban and rural hospitals

Overall, hospitals in urban areas were significantly more likely to support interoperable exchange than hospitals in rural areas: 65% (95% CI: 63%–68%, Figure 2) of urban hospitals and 47% (95% CI: 43%–50%) of rural hospitals engaged in interoperable exchange while 55% (95% CI: 52%–57%) of urban hospitals participated in national networks compared to 48% (95% CI: 44%–51%) of rural hospitals.

FIGURE 2.

Unadjusted associations between measures of marginalization and hospital interoperability in rural and urban areas. DSH, disproportionate share hospital; SDI, social deprivation index. Confidence intervals calculated using heteroskedastic robust standard errors. [Color figure can be viewed at wileyonlinelibrary.com]

Hospitals in the top quintile on mean SDI were consistently associated with lower scores on both measures of interoperable exchange in urban and rural areas: Urban hospitals in the highest SDI quintile were 31% less likely than other urban hospitals to engage in interoperable exchange (RR = 0.69, 95% CI: 0.60–0.81), while rural hospitals in the highest SDI quintile were 39% less likely than other rural hospitals (RR = 0.61; 95% CI: 0.46–0.79). Associations between the DSH index and interoperable exchange and Medicaid Case Percent and interoperable were not statistically significant.

While the 133 CAHs located in metropolitan areas were less likely to engage in interoperable exchange than other metropolitan hospitals (RR = 0.82; 95% CI: 0.70–0.97), they had a similar likelihood of participating in a national network. And Rural CAHs were no less likely to engage in interoperable exchange or participate in national networks than were other rural hospitals.

In urban areas, hospitals with greater uncompensated care burdens were again more likely to engage in interoperable exchange. In rural areas, hospitals with greater uncompensated care burdens were not more or less likely to engage in interoperable exchange.

3.4. Association between marginalization and interoperable exchange accounting for hospital characteristics

In multivariate models, those in the highest SDI quintile remained significantly less likely to engage in interoperable exchange (RR = 0.70; 95% CI: 0.62–0.79; Table 3) and to participate in a national network (RR = 0.88; 95% CI: 0.78–0.99). Hospitals with the highest uncompensated care burden remained significantly more likely to engage in interoperable exchange (RR = 1.21; 95% CI: 1.13–1.30) but were not more likely to participate in a national network. After controlling for other hospital characteristics, CAHs were no longer less likely than non‐CAH hospitals to engage in interoperable exchange or participate in national networks. Independent of all marginalization measures, rural hospitals also remained less likely than urban to participate in national networks.

TABLE 3.

Adjusted associations between marginalization and interoperable exchange accounting for hospital characteristics.

	(1)	(2)	(3)	(4)
	Relative risk engaging in interoperable exchange (95% CI)	Relative risk engaging in interoperable exchange including DSH percentile (95% CI)	Relative risk national interop. network participation (95% CI)	Relative risk national interop. network participation including DSH percentile (95% CI)
Highest 20% SDI	0.70**	0.78**	0.88*	0.87+
	(0.62–0.79)	(0.68–0.89)	(0.78–0.99)	(0.75–1.01)
Highest 20% DSH percentile		0.90+		1.04
		(0.81–1.00)		(0.93–1.17)
Highest 20% Medicaid case percent	0.96	1.01	0.95	1.00
	(0.87–1.05)	(0.93–1.10)	(0.87–1.04)	(0.91–1.09)
Highest 20% uncompensated care burden	1.21**	1.16**	1.00	0.99
	(1.13–1.30)	(1.07–1.25)	(0.91–1.09)	(0.89–1.10)
Critical access hospital	1.03		0.96
	(0.92–1.16)		(0.85–1.08)
Size (Ref: Small)
Medium	1.14*	1.07	0.87*	0.91
	(1.03–1.26)	(0.97–1.19)	(0.78–0.97)	(0.81–1.02)
Large	1.25**	1.16*	0.96	0.99
	(1.10–1.41)	(1.03–1.32)	(0.84–1.09)	(0.86–1.14)
Teaching status (Ref: Non‐teaching)
Residency program	1.11*	1.07	1.13**	1.07
	(1.01–1.20)	(0.98–1.16)	(1.03–1.24)	(0.96–1.19)
Medical school	1.29**	1.19**	1.34**	1.23**
	(1.13–1.47)	(1.05–1.36)	(1.17–1.54)	(1.06–1.41)
Rural	0.96	0.84**	0.86**	0.82**
	(0.87–1.06)	(0.74–0.94)	(0.77–0.95)	(0.73–0.93)
Multi‐hospital system member	1.79**	1.41**	1.41**	1.32**
	(1.59–2.01)	(1.24–1.61)	(1.25–1.58)	(1.16–1.49)
Ownership status (Ref: Non‐Profit)
For‐profit	0.82**	1.01	0.13**	0.10**
	(0.74–0.91)	(0.93–1.11)	(0.08–0.19)	(0.06–0.16)
Government	0.71**	0.91	0.79**	0.91
	(0.62–0.81)	(0.79–1.05)	(0.70–0.89)	(0.79–1.04)
Constant	0.37**	0.50**	0.56**	0.60**
	(0.32–0.43)	(0.43–0.59)	(0.48–0.65)	(0.51–0.71)
Observations	2393	1486	2280	1437

Note: Confidence intervals and statistical significance calculated using heteroskedastic robust standard errors.

Abbreviations: DSH, disproportionate share hospital; SDI, social deprivation index.

^* p < 0.05;

^** p < 0.01.

Our estimates were generally consistent in a series of robustness tests. We examined whether multivariate models in which only one measure of the extent to which hospitals treated patients that have been marginalized was related to interoperability when other measures of marginalization are not included in these models (Appendix Table 2). Results were generally consistent with one exception: when included in a model without other proxies for marginalization, the DSH percentile became statistically significant at conventional thresholds RR = 0.86, RR = 0.78–0.93 rather than of marginal significance in our primary results (RR = 0.90, RR = 0.81–1.00). We next examined whether differences in coefficients in models including the DSH index were driven by the population of hospitals eligible for DSH payments or collinearity between the DSH index and other measures by re‐estimating that model constrained to hospitals eligible for DSH payments but omitting the DSH index variable. We found that differences largely emanated from the sub‐population of hospitals rather than collinearity (Appendix Table 3). We also examined whether the relationship between measures of marginalization exhibited a dose–response effect by including each quintile of each measure in regression models. We found evidence consistent with a dose response for SDI and uncompensated care burden, wherein the group highest in each measure exhibited the greatest difference in relative risk. For Medicaid Case Percent we observed a consistent difference between the lowest quintile in each measure and all other quintiles (Appendix Table 4). Finally, we replicated our primary models including the primary EHR developer as a covariate. We observed few differences in the estimated association between measures of marginalization and interoperability in our primary models and models including EHR developer (Appendix Table 5). Some coefficients decreased in magnitude consistent with a suppression effect. (i.e., the top quintile SDI hospitals in models including primary EHR developers had a 0.80 relative risk of engaging in interoperable exchange versus a relative risk of 0.70 in our primary models) but these differences were not statistically significant.

4. DISCUSSION

The seamless exchange of health information, often supported by interoperability, is an important part of person‐centered, effective health care. ³⁰ Many individuals living in areas with high social deprivation may be at‐risk due to systemic structural inequities in their communities including, as identified here, in the hospitals they depend on for care. ¹⁸ , ³¹ We found that the selection of measures to identify marginalized populations, and their association with gaps in interoperability is complex. We identified a substantial digital divide in hospital engagement in interoperable exchange and participation in national interoperability networks for the 20% of hospitals that served patients from areas with the highest levels of area social deprivation (mean SDI). However, we found that marginalization measures other than SDI did not consistently convey the risk of a digital divide.

One important message from these findings is that the choice of proxy measure influences conclusions about whether a disparity exists in hospital interoperability. We found a robust correlation between SDI and interoperability that persisted within urban and rural hospitals (i.e., was independent of previously noted urban/rural disparities) and when controlling for other factors, but other proxy measures did not consistently identify disparities. This finding should perhaps not be surprising: recent work has highlighted the numerous definitions of “safety net” hospitals and that common proxies do not identify the same hospitals. ¹³ , ³² , ³³ , ³⁴ Others have demonstrated the challenge inherent in effectively targeting programs intended to assist safety‐net hospitals. ³⁵ Nevertheless, our findings highlight the importance of considering multiple proxies in this and related domains: choice of proxy may have implications for identifying disparities in other hospital processes, technologies, quality, and outcomes. In particular, measures like uncompensated care burden and other financial metrics, which may relate to hospital and health system accounting practices in addition to patient treatment patterns, may not effectively identify hospitals that treat marginalized populations. Recognition of the multidimensional challenge of caring for individuals that have been marginalized, and of measuring that marginalization, has motivated substantial equity‐centered work including efforts by CMS and others to include payment adjustments for social risk factors. ³⁶ , ³⁷

While our data cannot directly speak to why a persistent digital divide was identified by one measure but not by others, we believe there are two likely explanations. One possibility is that the SDI measure more effectively identifies hospitals that treat communities that have been marginalized by capturing broader dimensions of the area's socioeconomic condition. These facets of the area hospitals serve may impact their engagement in interoperability by driving other pressures on hospitals' performance and resources, increasing the importance of competing priorities, or leading to greater difficulty establishing strong information technology capabilities while serving patients that have been marginalized. Further investigation of the mechanisms driving the relationship between the social deprivation of an area that a hospital serves and the hospital's ability to invest in initiatives like interoperability seems warranted and may involve the use of related measures such as the area deprivation index. ³⁸

A second, and potentially complementary, explanation for the inconsistent relationship between programmatic measures of marginalization and interoperability is the success of past programs, including the Medicaid and Medicare EHR incentive program and regional extension centers; which used these programmatic proxies to target support to hospitals and were effective at increasing EHR adoption across hospitals. ³⁹ , ⁴⁰ Ultimately these efforts may have facilitated interoperability engagement among these hospitals and maintained parity. Similarly, DSH funds may have been at least partly successful in providing targeted hospitals with financial support to update IT systems, though there are independent concerns about the effective targeting of those payments. ⁴¹ However, because proxy measures are weakly correlated, programs targeted using one proxy may not have maintained parity between hospitals that disproportionately serve populations that have been marginalized as captured by other proxies.

Regardless of the underlying drivers of the observed inconsistency, our data highlight that approaches that focus on a subset of measures risk overlooking dimensions of marginalization most relevant to disparities in interoperable exchange or other outcomes of interests. As the Medicare Promoting Interoperability Program continues and the Medicaid program ends, it will be important to assess how changes in public support for health IT differentially impacts hospitals serving populations that had been marginalized using diverse measures. For example, it is possible that the parity in interoperability that we observe in Medicaid case volume in 2021 will not persist after the Medicaid program's conclusion.

In a similar vein, it is possible that digital divides will emerge, worsen, or ameliorate as new initiatives and technologies emerge that improve the value and usability of health IT. For instance, the Office of the National Coordinator for Health Information Technology has recently launched the Trusted Exchange Framework and Common Agreement (TEFCA), which aims to establish a universal policy and technical floor for nationwide interoperable exchange between existing exchange networks, including the national networks examined here. ⁴² , ⁴³ TEFCA has the potential to reduce the costs and complexity associated with interoperable exchange, making it easier for a broader range of organizations to exchange. Similarly, the growing use of application programming interfaces based on the HL7® Fast Healthcare Interoperability Resources (FHIR®) standard and the United States Core Data for Interoperability (USCDI) standard, could make engaging in interoperable exchange simpler and more valuable. ⁴⁴ , ⁴⁵ , ⁴⁶ Given that both of these policies likely require EHR updates or changes and that neither explicitly include support for providers that disproportionately treat populations that have been marginalized, they have the potential to contribute to digital divides. However, to the extent that these programs simplify exchange, they may lessen divides. It will be important to ensure that the potential benefits of these approaches reach hospitals and other organizations that care for communities that have been marginalized and that if they do not, additional support for those providers is considered.

Our measures by no means capture all dimensions of marginalization. It is possible that other dimensions exhibit yet larger digital divides, and additional work on enhanced and publicly available measures will be important to reliably detect and monitor progress towards greater interoperability equity. ⁴⁷ For instance, our measures did not include race/ethnicity differences and we employ only one of several available measures of area deprivation. Simultaneously, measures related to the treatment of populations that have been marginalized are likely to improve as interoperable information technology itself facilities standardized documentation and sharing of patient characteristics associated with disparities, such as through the uptake of USCDI. ⁴⁴

4.1. Limitations

Our study is subject to several limitations. First, our analysis is cross‐sectional and does not indicate the causes of disparities in interoperable engagement, which are likely rooted in deeper structural inequities than we capture here. While this approach is well‐suited to identify the presence of disparities it does not directly indicate how those disparities emerged. In addition, our measure of SDI is at the ZIP code level probabilistically relates to the sociodemographic status of patients treated at hospitals. Smaller geographies, such as Census tracts, are closer geographic proxies for “neighborhoods” than ZIP codes and may improve the precision of that relationship. However, measurement error associated with the use of larger geographies would likely bias towards the null and despite that concern, our data show a robust association between SDI and interoperable exchange. Finally, we identified a threshold (top quintile) to categorize hospitals for each measure; additional investigation should seek to identify the most informative threshold for each measure.

5. CONCLUSION

The last decade has seen substantial progress towards widespread engagement in interoperability by many hospitals, while other hospitals and their patients are still not benefitting. This progress raises the question of which patients may not be included in widespread data sharing. Our findings suggest wide variation across metrics in their ability to identify digital divides across populations. Only one measure—an area‐based measure capturing social deprivation—consistently identified hospitals facing a substantial digital divide. Further evaluation of this and other proxy measures will be critical if policy makers and planners are to identify hospitals that disproportionately care for marginalized groups and reduce remaining disparities in interoperable exchange.

FUNDING INFORMATION

No funding to report.

Supporting information

Data S1. Supporting Information.

Data S2. Supporting Information.