Data-driven modeling of diabetes care teams using social network analysis

Abstract

Objective

We assess working relationships and collaborations within and between diabetes health care provider teams using social network analysis and a multi-scale community detection.

Materials and Methods

Retrospective analysis of claims data from a large employer over 2 years was performed. The study cohort contained 827 patients diagnosed with diabetes. The cohort received care from 2567 and 2541 health care providers in the first and second year, respectively. Social network analysis was used to identify networks of health care providers involved in the care of patients with diabetes. A multi-scale community detection was applied to the network to identify groups of health care providers more densely connected. Social network analysis metrics identified influential providers for the overall network and for each community of providers.

Results

Centrality measures identified medical laboratories and mail-order pharmacies as the central providers for the 2 years. Seventy-six percent of the detected communities included primary care physicians, and 97% of the communities included specialists. Pharmacists were detected as central providers in 24% of the communities.

Discussion

Social network analysis measures identified the central providers in the network of diabetes health care providers. These providers could be considered as influencers in the network that could enhance the implication of promotion programs through their access to a large number of patients and providers.

Conclusion

The proposed framework provides multi-scale metrics for assessing care team relationships. These metrics can be used by implementation experts to identify influential providers for care interventions and by health service researchers to determine impact of team relationships on patient outcomes.

BACKGROUND AND SIGNIFICANCE

Diabetes is the seventh leading cause of death in the United States.¹ With an aging population, diabetes prevalence is expected to increase over the next 40 years, with 1 in 3 Americans projected to have diabetes by 2050.² The economic burden of diabetes affects both individuals and society. According to the American Diabetes Association, the total cost of diagnosed diabetes has increased from $245 billion in 2012 to $327 billion in 2017. Outside of the diabetes medication cost (30%), the next largest expenditure is physician office visits (13% of the total costs),³ as the management of diabetes is often complex and requires multiple visits to various types of care providers (eg, primary care, emergency physician, specialist, and nutritionist).⁴ Previous studies have highlighted the importance of health care providers’ coordination and collaboration for improving diabetes care delivery and reducing costs,^5–7 however, challenges including under- or over-utilization of health care providers and unclear definition of providers’ roles and responsibilities⁸ negatively impact effective provider teamwork.

General recommendations and guidelines exist for team roles in the care of patients with diabetes. The American Diabetes Association has recommended that a diabetes care team should consist of a primary care provider, nurse educator, registered dietitian, diabetes educator, endocrinologist, eye doctor, social worker, psychologist, podiatrist, pharmacist, dentist, and exercise physiologist.⁹ Despite the recommendations for diabetes care teams, not all teams are assembled and structured based on those recommendations. For example, health care providers may be added to or removed from the care teams by patient or patient family members, and providers’ collaborations and communications may be challenged by these changes.¹⁰ In addition, not all providers have clear descriptions of their roles, duties, and responsibilities as part of the overall care team for patients with diabetes.¹¹ Moreover, the frequency at which these ADA recommendations occur in practice is unknown.

Some major challenges to understanding care teams are the lack of data and difficulty in collecting data that measures health care providers’ working relationships.¹² There are questionnaire tools for measuring collaborations of health care providers,^13,14 however, survey-based research may be limited by a low response rate.¹⁵ Interview method is another technique which despite capturing relationships from each health care professional perspective, is limited to a small number of participants. Although they capture relationships from each health care professional’s perspective, they are typically limited to a small number of participants.¹⁶ It is important to also note that physicians and other providers form different working relationships that might be formal, such as referrals or associations with a hospital,¹⁷ or informal, like advice seeking.¹² New approaches are needed to better capture these relationships and the associated impact on providers’ decision-making and patient outcomes.

Originally developed from the social psychology¹⁸ and based on the work of Moreno,¹⁹ social network analysis (SNA) is a powerful technique that captures hidden channels of collaboration, information flow, and communication between network actors (eg, individuals, groups, organizations). In health care settings, SNA has been used to study health care providers’ collaboration, communication, and the impact of their interaction on decision-making for the patient.^20–24 These studies have primarily used surveys and interviews data^25–27 for SNA, however, application of this technique on large-scale health data can further capture the physicians’ relationships and their collaboration on a larger scale.^12,17,28,29 Physicians who provide care to the same patients form a team of providers that may have formal referral relations or informal discussion about the patients.¹² This approach helps go beyond previous utilization analysis studies that focused on individual patient health services.^30,31 Social network analysis using the patient sharing approach among health care providers allows us to look at the collaboration of health care providers, meaning: the “recurring process of working together toward common goals.”³²

SNA application on administrative data has been previously used to understand the geographic differences in the physicians’ networks across the US^10,17,33 or to assess the provider network characteristics’ impacts on patient outcomes.^34–40 Landon et al.¹⁷ used SNA and patient sharing techniques to explore variations in the professional networks of physicians across the US and showed that physicians tend to share patients with similar groups of health care providers that had similar patient panel characteristics. Others have used a similar approach to identify groups of physicians that formed working relationships and had the potential of becoming an Accountable Care Organization³³ based on physicians’ regional networks and the consistency of their relationships.¹⁰ Although applications of SNA in health care are increasing, current work with SNA primarily focuses only on providers who provide direct care to patients.^10,34,37,41 This approach is useful for assessing the relationships between individual providers like physicians, however, it neglects other members of the care team. Moreover, the many health care services that are not delivered through direct care (eg, labs) are also overlooked.

The objective of this study is to develop a framework using SNA metrics and a multi-scale community detection algorithm to assess the working relationships of care teams involved in the diabetes care process. The SNA metrics identify the global structure of the network and the influencers within the network while the community detection uncovers the health care providers that work more closely.

MATERIALS AND METHODS

Data source

This study was approved by the university’s Institutional Review Board (IRB 1511016796). The data contained de-identified administrative health claims of employees (faculty/staff) and their dependents (spouse/child) of a large university in the Midwest. The claims data included medical health utilization of the study population, insurance eligibility information, and medication usage. The eligibility data included demographic information such as age, sex, and type of compensation (hourly/salary). The medical file included International Classification of Diseases 9th edition code (ICD9), service date, cost of service and health care provider information. The medication file included medication purchase date, cost, and provider information. This study excluded the student population.

Study design and analysis

A retrospective analysis of the health claims data was performed. Table 1 details the variables in the data set. Individuals with diabetes (type 1 and type 2) and their health care providers were tracked in 2012 (Year1) and 2013 (Year2) by using their unique patient and provider identifiers. To determine cohort population with diabetes, primary, secondary, and tertiary diagnoses for the ICD9 starting with 250 were used⁴² (250.00, 250.01, 250.02, 250.40, 250.41, 250.50, 250.51, 250.52, 250.60, 250, 61, 250.62, 250.80, 250.81, 250.82, 250.90, 250.91). All health services utilization and health care providers’ identifiers for the identified patients were extracted from the medical and medication files.

Table 1.

Description of variables in the data set

Health Administrative Data
Demographic Variables	Clinical Variables	Insurance Plans	Health Care Provider
Sex (male/female) Date of birth (MM/DD/YYYY) Person ID (unique identifier) Relationship to employee (employee, spouse, and child) Compensation type (hourly, salary).a	Principal diagnosis, secondary diagnosis, and tertiary diagnosis (ICD9) Claim ID (7-digit value)	Coverage indicator (dental, drug, vision, hearing, long-term disability, short-term disability, medical, Mental Health Services Act) Type of plan (no coverage, spouse opt out, Federal Health Plan, low deductible, high deductible)	Provider ID (unique identifier) Provider zip code (5-digit zip code of providers) Provider type (eg, primary care, dentist)

^aCompensation variable for spouse and child are based on the policy holder’s status.

Constructing the network

Network construction was based on the patient sharing relations between the health care providers, an approach that has been previously validated.¹² Presence of shared patients among the providers is interpreted as an information sharing relationship between the 2 providers.³⁴ Shared patients were determined from the health claims. The network nodes represented health care providers, and an edge between 2 nodes represented shared patients between the providers. The edge weights were defined by the number of shared patients, a proxy for the strength of the relationships between the providers. The network was limited to patients with diabetes and their corresponding health care providers. All types of health care providers were included in order to capture the health services needed to assess direct and indirect care and address potential comorbidities and complications of diabetes. Previous work demonstrated that information sharing relationships between health care providers were observed with link weights of 2 or more.¹² Thus, the network focused on providers who shared at least 2 patients. Excluding providers with only 1 shared patient has been recommended to remove relationships formed by chance that may not carry much information about the relationship.^12,17

Constructing communities

To further understand the structure of the network and identify the natural communities of the health care providers, local analysis of the network through community detection was performed.⁴³ Communities in this study were defined as groups of nodes that were more densely connected internally compared to their connection to the rest of the network.⁴⁴ The community detection algorithm used a multi-scale approach. Stability was the objective function optimized to find the best partitions of the network.⁴⁵ This method can identify smaller communities in the network, unlike the more common modularity-based community detection algorithms⁴⁶ which suffer from a resolution limit⁴⁷ and are unable to identify communities with edges fewer than $\sqrt{\raisebox{1ex}{$L$}\!\left/ \!\raisebox{-1ex}{$2$}\right.},\mathrm{}L$ being the number of edges in the entire network. Health care providers were assigned to single non-overlapping communities using this multi-scale method.

Metrics

SNA metrics were used to assess the global characteristics of the network and communities. Density measured the cohesions and frequency of collaboration among the health care providers.⁴⁸ Measures of centrality identified nodes with important roles in the network and greater access to other nodes.^49,50 Three centrality measures were calculated: 1) degree centrality, 2) betweenness centrality, and 3) closeness centrality. The centrality measures were applied on the largest component of the network to identify the influencer and providers with greater access and control over the flow of information in the network.

Degree centrality of a provider in the network showed the number of providers that were directly connected to that provider. The betweenness centrality measured the degree by which a node was “between” pairs of other nodes in the network. A node having higher betweenness centrality indicates its having more influence in the network for distributing information.³⁷ Higher closeness centrality of a node shows better access to the remaining nodes in the network.⁵¹ In addition to network density, we identified connected components in the network in which all nodes were directly and indirectly connected.⁵²Table 2 shows the measures of SNA used in the study and their definitions. The analysis was completed using SAS (v 9.4, SAS Inc., Cary, NC) and RStudio (version 0.99.903) with the igraph (version 1.1.2),⁵³ as well as devtools (version 1.12.0)⁵⁴ packages.

Table 2.

Measures and properties of social network used in the analysis

Network Measures and Properties	Definition
Density	Proportion of edges in a graph to the maximum possible number of edges with value of 0 to 1.52
Degree Centrality	The number of ties a node has in a graph.52
Betweenness Centrality	A measure of centrality based on shortest path, showing the extent to which a node is between other pairs of nodes.52
Closeness Centrality	A measure of closeness of each node to other nodes in the network.52
Component	A subgroup of a graph in which all nodes are connected.52
Community	Subgroup of a network with denser internal connections compared to its connections to the rest of the network.52

RESULTS

Out of a total of 15 183 individuals with at least 1 medical claim in Year1, 827 patients were identified with diabetes (5.4%). The patient cohort consisted of 416 (50%) women with an average age of 54 ± 9.9 years and 411 (50%) men with an average age of 57.6 ± 10.0 years. In addition to diabetes, other common health conditions (ie, comorbidities) identified in the cohort were: 1) hypertension (ICD9 = 401.1 & 401.9) with 271 (32%) patients in Year1 and 296 (35%) patients in Year2; and 2) hyperlipidemia (ICD9 = 272.4) with 152 (18%) patients in Year1 and 159 (19%) patients in Year2.

The cohort with diabetes received health care services from 2567 health care providers and 2541 health care providers in Year1 and Year2, respectively. A total of 1523 of the providers (59%) present in Year1 were also present in Year2. Number of providers after the exclusion criteria of removing providers with <2 patients in common for Year1 was 896 and for Year2, 836. The provider type with the highest utilization for both years was a medical laboratory, with 418 patients in both years. The second most utilized provider was a mail-order pharmacy with 306 patients in Year1 and 337 patients in Year2.

Network characteristics

The network of health care providers in Year1 consisted of 896 nodes and 10 200 edges with density 0.0261. The network had 7 components (Table 2) with all nodes connected directly or indirectly.⁵² The largest component had 884 nodes with 10 194 edges, while the other 6 components each had 2 nodes. These 6 smaller component nodes (health care providers) had only 1 shared patient with the rest of the network, but since edges with weights <2 were removed (see methods section) they resulted in separate components. The edge weights (numbers of patients shared between the providers) ranged from 2 to 271 (average of 4.03 and standard deviation of 7.3).

Focusing on the largest component, the median degree centrality for all the network nodes was 8, the median betweenness centrality of the networks nodes was 108.8, and the median closeness centrality was 0.0002. Table 3 presents the top 10 health care providers with highest degree, betweenness, and closeness centrality. Although medical laboratory was the provider type identified as having the highest degree and betweenness centrality, the actual medical laboratory provider differed in the 2 metrics. Specifically, Provider ID 1667 had the highest degree centrality, while Provider ID 1439 had the highest betweenness centrality. Although degree and betweenness centrality are different SNA metrics, nodes present in the top 10 for degree and betweenness centrality were similar. Specifically, 80% of providers (Provider IDs 397, 292, 2114, 107, 171, 489, 1667, and 382) ranked in the top 10 for degree centrality were also in the top 10 for betweenness centrality metrics (Table 3). Top nodes identified by the closeness centrality were also identified by degree and betweenness centrality as top providers (Provider IDs 397, 292, 2114, 107, 489, 382).

Table 3.

Year1 providers with the highest degree, betweenness, and closeness centrality

Provider IDa	Provider Type	Degree Centrality	Provider’s Unique Patients
1667	Medical laboratory	482	418
397	Mail-order pharmacy	468	306
292	Mail-order pharmacy	450	306
2114	Medical laboratory	377	221
107	Home health organization	359	156
1670	Hospital	340	175
171	Hospital	338	127
489	Mail-order pharmacy	321	183
382	Mail-order pharmacy	308	172
126	Hospital	247	92
		Betweenness Centrality (Normalized Betweenness Centrality)	Provider’s Unique Patients
1439	Medical laboratory	47 938 (0.123)	131
2114	Medical laboratory	38 786(0.099)	221
397	Mail-order pharmacy	36 389 (0.093)	306
107	Home health organization	35 793 (0.091)	156
292	Mail-order pharmacy	28 776 (0.073)	306
489	Mail-order pharmacy	23 323 (0.0598)	183
1667	Medical laboratory	23 142 (0.0594)	418
159	Hospital	19 470 (0.050)	30
382	Mail-order pharmacy	19 343 (0.049)	172
171	Hospital	14 971 (0.038)	127
		Closeness Centrality	Provider’s Unique Patients
397	Mail-order pharmacy	0.000262	306
292	Mail-order pharmacy	0.000254	306
107	Home health organization	0.000248	156
1859	Podiatry	0.000246	8
489	Mail-order pharmacy	0.000245	183
1439	Medical laboratory	0.000245	131
382	Mail-order pharmacy	0.000242	172
928	Family practice	0.000242	4
2114	Medical laboratory	0.000241	221
2386	Hospital	0.000241	4

^aProvider IDs are random numbers generated for each provider and do not represent their real IDs.

The network of Year2 was connected and composed of 836 nodes (health care providers) and 9722 edges (patients shared among health care providers) with a density of 0.0278. The edge weight range was from 2 to 181 (with the average of 3.97 and standard deviation of 6.04). Compared to Year1, number of nodes and of edges were reduced, but density and average edge weights were similar.

Similar to Year1, the median degree centrality was 8 for all the nodes in the network, the median betweenness centrality was 107.67, and the median closeness centrality was 0.0002. In contrast to Year1, mail-order pharmacy was the provider type with the highest degree, betweenness, and closeness centrality (Table 4). Table 4 presents the top 10 health care providers identified as having the highest centrality for Year2. As with Year1, 80% of the providers ranked in the top 10 for both degree and betweenness centrality metrics (Provider IDs 301, 1628, 2064, 106, 403, 170, 1630, and 1525). Mail-order pharmacy (ID 292) ranked highest based on degree, betweenness, and closeness centrality. Comparing provider IDs from Tables 3 and 4, 90% of the providers from Year1 (Table 3) also appeared as central providers based on the betweenness and degree centrality measures in Year2. Three of the top providers identified by closeness centrality (Provider IDs 292, 382, and 107) were also identified as central by degree and betweenness centrality.

Table 4.

Year2 providers with the highest degree, betweenness, and closeness centrality

Provider IDs	Provider Type	Degree Centrality	Provider’s Unique Patients
292	Mail-order pharmacy	499	337
1667	Medical laboratory	461	418
2114	Medical laboratory	351	218
107	Home health organization	333	148
382	Mail-order pharmacy	325	184
171	Hospital	309	124
1670	Hospital	287	125
1525	Worksite clinic	261	155
288	Retail pharmacy	260	155
2292	Pathology	251	80
		Betweenness Centrality (Normalized Betweenness Centrality)	Provider’s Unique Patients
292	Mail-order pharmacy	67 063 (0.192)	337
382	Mail-order pharmacy	30 997 (0.089)	184
107	Health home organization	30 314 (0.087)	148
2114	Medical laboratory	30 032 (0.086)	218
1439	Medical laboratory	29 060 (0.083)	129
1667	Medical laboratory	25 109 (0.072)	418
170	Pathology	11 878 (0.034)	124
397	Mail-order pharmacy	10 980 (0.031)	57
1525	Worksite clinic	8553 (0.024)	155
1670	Hospital	7969 (0.022)	125
		Closeness Centrality	Provider’s Unique Patients
292	Mail-order pharmacy	0.000286	337
382	Mail-order pharmacy	0.000265	184
107	Health home organization	0.000263	148
1929	Anesthesiologist	0.000261	4
1570	Neurologist	0.000259	8
1435	Emergency Medicine	0.000256	5
2482	Gynecologist	0.000256	9
1979	Supply Center	0.000255	4
2298	Dermatologist	0.000255	11
2131	Preventive Supply Center	0.000255	8

^aProvider IDs are random numbers generated for each provider and do not represent their real IDs.

The following graphs in Figure 1, provide the distribution of nodes degree (a), betweenness (b), and closeness (c) centrality for Year1 and Year2.

Figure 1.

Distribution of the degree (a), betweenness (b), and closeness centrality (c) of the nodes for Year1 and Year2.

Network communities

Although network analysis with shared patients identified central providers and global network characteristics, it lacks the granularity needed to address the study’s objective of identifying working relationships within and between care teams. As mentioned in the methods section, edges between providers with less than 2 patients shared were removed. The main component of the network in Year1 had 884 nodes and 10 194 edges. A multi-scale community detection algorithm⁴⁵ was applied on this component to identify groups of health care providers more tightly linked together through patients sharing relationships. Forty-six communities were detected for Year1 with the sizes of 3 to 155 members. Communities of providers served an average of 115 patients.

One of the smaller communities is shown in Figure 2 to illustrate community characteristics detected using this approach. This community of care providers was composed of 6 providers: a general surgeon (ID 501), a pharmacist (ID 556), an ophthalmologist (ID 1174), 1 family practice (ID 2151), an anesthesiologist (ID 2438), and a gynecologist (ID 2513). Based on the degree, betweenness, and closeness centrality applied to the community, the central node (ID 1174) was an ophthalmologist with degree of 4, betweenness of 8 and closeness of 0.077.

Figure 2.

One of the communities detected in the network of health care providers of Year1. The node 1174 is an ophthalmologist, the node 2151 is a family practice, the node 2513 is a gynecologist, the node 556 is a pharmacist, the node 2438 is an anesthesiologist, and the node 501 is a general surgeon. Degree, betweenness, and closeness centrality of each node is depicted next to it. The number on each edge represents the edge weight.

Thirty-five out of 46 (76%) detected communities included a primary care practitioner (internal medicine, family practice). Forty-five out of 46 (97%) communities included a specialist (endocrinologist, nephrologist, ophthalmologist, oncologist, gynecologist, cardiovascular specialist, podiatrist, and dermatologist). To determine the most central provider in each community, centrality metrics were used. Among the health care providers, pharmacists were detected as the central nodes in 24% of the communities. Radiologists and hospitals were the next most commonly central providers, observed in 13% and 11% of the communities, respectively. Specific services provided by these central providers (ie, radiologists and hospitals) were identified using Current Procedural Terminology codes for the patients. Chest x-rays (71020) were the most common procedure performed by central radiology providers. Procedures provided at hospitals varied widely and included screening mammogram digital (G022), fluoroscopic guidance (77003), chemistry procedure-creatinine (82565), chemistry procedure-calcium (82310), and blood count (52025).

Comparable to Year1, 45 communities were detected from the network in Year2 with the multi-scale community detection algorithm.⁴⁵ Similarly, resulting communities had a range of 3 to 145 health care providers (nodes) and served an average of 119 patients. Similar to Year1, detected communities commonly had a primary care physician (detected in 87% of the communities) and a specialist (detected in 98% of the communities). Pharmacists were the central providers in 22% of the communities based on SNA centrality metrics. Next most common central providers were radiologists and hospitals, observed in 18% and 11% of communities, respectively. For the central radiologists, screening mammogram digital (G0202) was the most commonly performed procedure. Procedure performed by the hospital providers were diverse and included glucose monitoring (82948), screening mammogram digital (G0202), blood test (36415) and injection fentanyl citrate (J3010).

DISCUSSION

The study demonstrated a framework integrating SNA and state-of-the-art community detection techniques to provide insights and metrics for assessing care teams and collaborations for complex, chronic disease management. This approach leveraged big health data sets to analytically identify key players and influencers in the care teams for patients with diabetes. Previous studies that used SNA application on large-scale data to assess coordination and working relationships between health care providers only focused on physicians who provided direct care to the patients or provided care to patients with a variety of health conditions.^34,35,37 Extending SNA to all provider types within the care system is needed to identify gaps in the services in comparison to guidelines and to quantify team relationships for complex chronic conditions such as diabetes.⁵⁵ In this study, we included all health care providers involved in the care of patients with diabetes and used the patient sharing approach and SNA on health administrative data to identify key stakeholders and providers in the care teams.

Potentially contrasting results were presented regarding the stability and continuity of the care provider population over the 2 years. Specifically, over 40% of the providers were different between the 2 years. Numerous reasons may cause the change in providers, including change of the insurance plans by patient or physician’s office change of policy.^56,57 However, application of the community detection and other SNA approaches showed strong consistency between the years. Specifically, the number of nodes and edge weights were similar indicating that key providers of services and services needed were relatively stable. Moreover, our results showed that 9 central providers based on degree and betweenness centrality from Year1 (Table 3) also appeared as central in Year2. Distribution of centrality measures in Figure 1, shows similar patterns for the node centralities between the 2 years as well. These results suggest a consistency of usage of these central providers and continuation of their working relationships with other providers. Continuous provider relationships with patients is an important aspect of the care continuity⁵⁸ which can positively impact health care costs, outcomes,⁵⁹ and care coordination.⁶⁰ Thus, although a global analysis of the provider turnover raises alarms regarding the patient care, our approach showed that providers with central roles in the network of providers remained consistent over the period of the study.

When expanding SNA to capture all providers, those that provided direct care were not the central nodes in the network. Based on the degree, betweenness, and closeness centrality, pharmacy providers were among the central providers in Year1 (Table 3) and Year2 (Table 4) in the overall network. These findings may reflect a bias of this approach toward pharmacists as the network was designed based on the patient sharing¹² and providers with higher number of patients would be identified as central by the degree centrality.⁵² Despite the possibility of this bias and because of the crucial role of the pharmacist in the diabetes management,^61–63 considering the pharmacists as central providers is not unrealistic. The central role of pharmacy captured by this analysis reflects both utilization trends and intervention opportunities shown in other studies. For example, usage of mail-order pharmacies is increasing in the United States, and studies have shown that patients with diabetes who use these pharmacies have better medication adherence.^64,65 Although they may not have the face-to-face consultation of retail pharmacists, the phone consultation option provides a convenient way for patients to connect with the pharmacists.⁶⁶ Moreover, usage of mail-order pharmacists in the health care delivery level improves access to medication for chronic illnesses.⁶⁷ Our data showed a similar trend for usage of the mail-order pharmacists for the study cohort. Intervention programs for diabetes can be leveraged by the findings from this research, which identifies potential central providers other than primary care that could be effective in increasing medication adherence for the patient.⁶⁵

Eighty-seven percent of the communities detected in Year1 included primary care physicians vs 98% that included specialists. Most communities detected had a combination of primary care, specialists, and pharmacists, but not all providers outlined as the care team by the Americans Diabetes Association⁹ were identified in each community. Patients might have different types of needs based on their characteristics (eg, age, complication); therefore, the relationships of providers and their strength may vary based on patients. The more frequent appearance of the specialists in the communities compared to the primary care providers may be due to various reasons. First, some patients may self-refer rather than go through the primary care channel to see a specialist.⁶⁸ Other patients might have more complications and require specialists such as endocrinologists to manage their care.⁶⁹ The multi-scale community detection⁴⁵ is a useful tool to understand the structures of the relationships among the providers and appearance patterns for different types of physicians. SNA measures amplify this technique by identifying the central providers in each community.

This study has several limitations. As the results show, social network centrality measures might be biased toward providers with larger numbers of patients. Although we used 3 different measures of centrality (degree, betweenness, and closeness) and community detection to address this limitation to some extent, additional refinement in the approach may be needed to better recognize central providers. Another limitation was the population cohort, which was drawn from 1 large employer in the Midwest. Expanding the data set is needed to determine whether findings are generalizable to population of patients. This study is limited by provider definitions of the data set. For example, medical labs may have pathologist consultants; however, pathologist consultants may also be labeled separately in other cases. Although the study could be impacted by the limitations of the claims data, the methodology is still effective for identifying the care team interactions and central providers. A multi-scale community detection algorithm was used to identify providers with denser connections. This algorithm assigned providers to non-overlapping communities. This might be a limitation as health care providers may not belong to a single team (community). Despite the limitation, the algorithm is useful for detecting providers with a higher number of interactions. Finally, studies are needed to determine the associations of SNA metrics and community composition with patient health outcomes.

CONCLUSION

This study demonstrated the proposed approach for identifying key stakeholders, working relationships, and composition of care teams in the communities. A multiyear analysis was performed to understand the consistency and changes of the study provider networks. Although a necessary first step is to measure provider relationships, further work is needed to link these metrics to health outcomes and costs. The long-term goal of this research is to translate the SNA and community detection framework for designing strategies for improving provider collaboration and assessing how these relationships impact patients’ health outcomes and health care services costs. Although a few studies have looked at the associations between patient outcomes and the providers’ network characteristics, there is a lack of research for assessing relations of these networks to patient safety outcomes, including measures of adverse events.⁷⁰

FUNDING

This research received no specific grant from any funding agency in the public, commercial or not-for-profit sectors.

AUTHOR CONTRIBUTIONS

All authors (Ostovari, Steele-Morris, Griffin, Yu) contributed to the design of the work, analysis and interpretation of the data, as well as drafting, critically revising, and final approval of the manuscript. All authors agreed to be accountable for all aspects of the work to ensure accuracy and integrity.

CONFLICT OF INTEREST STATEMENT

None declared.