Abstract
Objectives.
To evaluate if the networks of diabetic patients sharing physicians are associated with emergency department (ED) visits and hospitalizations
Study Design:
Retrospective cohort study
Methods:
We used administrative data from a large insurer in Hawaii in 2010. Three types of networks were defined based on patient visits: (1) the total number of links from one patient to other patients sharing a physician; (2) the number of other patients connected by sharing the physician seen the most often; and (3) the number of other patients connected by seeing all the same physicians during the year. The networks were characterized into thirds based on their complexity and analyzed using zero-inflated negative binomial regression models on ED visits and hospitalizations.
Results:
The study included 38,767 diabetes patients with a mean age of 64 years. Patients sharing the most physicians had double the risks of ED visits and hospitalizations. Patients linked by belonging to the largest primary care practices had a 28% reduced odds of ED visits. Patients linked by seeing all of the same physicians during the year had the fewest primary care providers and specialists visits and 25% to 50% reductions in ED visits and hospitalizations.
Conclusions:
Networks of diabetic patients sharing all the same physicians were associated with decreased ED visits and hospitalizations. Encouraging diabetic patients to find a provider they like and trust and to stay in the provider’s care may help reduce the risks of adverse events. Physicians building loyalty among their patients may reduce their patients’ risks.
Keywords: diabetes, network, health care, emergency department, hospitalizations
Diabetes often has an insidious progression over the lifetime of the patient (1,2). Careful and continuing monitoring and control of diabetes can reduce the risks of acute or long-term complications. Physicians can direct and support diabetic patients in their care but it is essential to include the patients in the decision making for successful diabetes management (3,4). The size of the diabetes panel that a physician treats, the consistency of patient-physician interactions, and the trust of patients in their primary physicians may all play roles in the identification and development of optimal care.
Diabetes healthcare networks can be constructed from diabetic patients that share the same physicians (5). Such networks offer opportunities to better understand population health and clinical practice (6–8). Networks of patients sharing physicians contain patients who most often are unaware of one another. The networks are defined by different patients making similar choices in their care; the networks bridge across patients and create opportunities for population health management (5–8).
In this article we hypothesized that the results will show significant associations between the patient networks and the risks of the adverse health outcomes.
Methods
Data and study population.
We utilized the administrative claims data from a large health insurer in Hawaii in 2010. The diagnosis of diabetes was based on criteria from the Health Effectiveness Data and Information Set (HEDIS) (9). The study population was patients with diabetes mellitus enrolled in fee-for service plans (10). Only diabetes patients with continuous enrollment throughout the year were included. All the patients were free to choose their physicians. Patients were included if they had complete data on the study variables (i.e., no missing data on the analysis variables). About 95% of the patients without missing data in the study variables were included in the study. The University of Hawaii Institutional Review Board (IRB) exempted the study from IRB review.
Study variables.
Patient characteristics included demographic variables, morbidity, major chronic diseases, and island of residence. Morbidity was classified as high or low by the insurer based on the Adjusted Clinical Groups developed by John Hopkins University (11). Major chronic diseases were coronary artery disease, congestive heart failure, and chronic kidney disease. Residence was categorized as living on the most populous island of Oahu or on the more rural neighboring islands (Kauai, Maui, and Hawaii). Physician visits were defined to primary care providers (e.g., internal medicine, general practice, or family practice physicians) or specialists (e.g., cardiologists and endocrinologists). The primary objective was to understand how patient’s physician networks may affect their adverse outcomes. Healthcare networks in general are based on interconnections between people such as people that are friends or work together; (12, 13), and in this study, the networks are formed by patients connected to physicians by having had an office visit.
The patient-physician networks were categorized by
the number of links to other patients seeing the same physicians (type 1)
the number of links to patients having the same primary physician (type 2)
the number of patients linked by seeing all the same physicians during the year (type3). (8)
Figure 1 illustrates the networks conceptually. All patients belong to the type1, type2, and type3 networks. The network of patients sharing one or more physicians represents minimal similarity in care since patient sharing one physician may see many others. A patient that changes doctors often may share physicians with the most other patients. Patients having the same primary physician would be managed more similarly than other patients, but might also receive care from other physicians. We defined the primary physician as the physician seen most often (or the most recently in case of ties). Patients seeing all the same physicians would experience the most similar pattern of care. The networks were created from relational administrative files. One file had data on office visits and two others had data on emergency department (ED) visits and hospitalizations.
Figure 1.
Illustration of three types of patient-physician networks
Patients are presented as blue rectangles and physicians as red or yellow circles. The yellow circles represent the primary physician, the one seen the most often or the most recently in case of ties. The left panel shows two networks of patients that share one or more physicians. The top illustration shows two patients that share two physicians; one of the pair sees two physicians not shared by the other patient. The lower network shows two patients that share one physician; both patients see one other physician. The middle panel illustrates patients that see all of the same physicians. The top illustration shows two patients that share the only physician they see. The bottom illustration shows three patients that share two physicians. The right panel shows two networks of patients that share the same primary physician. In the top network the patients only see their primary physician and in the bottom network one patient visits a second physician as well.
We analyzed the networks as predictors of ED visits and hospitalizations, a proxy measure for the occurrence of adverse outcomes.
Statistical analysis.
Descriptive results included percentages for categorical variables and means with standard deviations for continuous variables. Regression models used two-part zero-inflated negative binomial models (ZNIBs) (14). The first part used logistic regression to analyze the risk of having one or more ED visit (or hospitalizations); the second part used negative binomial regression with a log link to analyze the rate of ED visits (or hospitalizations) from based on any diagnosis codes among patients that had one or more. Results are presented as odds ratios and rate ratios, respectively.
For the analyses, we divided the three network structures into lower, middle, and upper tertiles. The tertiles do not assume a shape of the relationships and retain a sufficient number of patients in the tertiles to assess the significance of possible differences. The lower category was selected as the reference in the regression models. As an example, groups formed based on the total number of links to other patients included a third with the lowest total, a middle third, and an upper third that included patients sharing the most physicians. Initial models analyzed network structures separately. In the final models, the lower and middle categories were combined to avoid creating numerous smaller groups. Being in the upper was coded as one and in the two combined categories as zero. The reference for these analyses was the combined categories.
Statistical analyses were performed with SAS version 9.4 (Cary, IN) and R version 3.5.
Results
The study included 38,767 diabetes patients (Table 1). The average patient shared one or more physicians with 270 other diabetic patients; they shared the same primary physician with 169 other patients; and they shared all their physicians seen with 110 other patients. The number (%) of ED visits by tertiles of the three exposure variables are for sharing all physicians 1,597 (58%), 507 (18%) and 656 (24%) for sharing primary physician 1,010 (37%), 793 (28%), and 957 (35%,) and for all physicians seen 474 (17%), 534 (19%) and 1,752 (63%). The number (%) of hospitalizations by tertiles of the three exposure variables are for sharing all physicians 3,740 (52%), 1,572 (22%) and 1,850 (26%) for sharing primary physician 2,313 (32%), 2,245 (31%), and 2,604 (36%,) and for all physicians seen 3,934 (55%), 1,745 (24%) and 1,483 (21%).
Table 1.
Characteristics of the study population
| Characteristic | Percentage or Mean ± Standard Deviation |
|---|---|
| Age groups (years) | |
| 18–44 | 3,328 (8.6%) |
| 45–54 | 5,555 (14.3%) |
| 55–64 | 10,241 (26.4%) |
| 65–74 | 9,839 (25.4%) |
| 75 and above | 9,804 (25.3%) |
| Female | 18,972 (48.9%) |
| Oahu | 28,498 (73.5%) |
| High morbidity | 17,075 (44.0%) |
| Coronary artery disease | 8,598 (22.2%) |
| Congestive heart failure | 3,535 (9.1%) |
| Chronic kidney disease | 2,192 (5.7%) |
| Primary care physicians seen | 1.26 ± 0.86 |
| Endocrinologists and cardiologists seen | 0.27 ± 0.77 |
| Total number of physicians shared | 269.7 ± 235.4 |
| Number of other patients seeing the same primary care physician | 168.1 ± 140.10 |
| Number of other patients seeing all of the same physicians | 110.4 ± 101.7 |
We extended the analyses using two-part ZINB models that separate the risks of first events and the rates of repeated events (Table 2 and Figure2). Diabetes patients with the highest total number of physicians shared had triple the risk of having one or more ED visits, but no increase in the rate of repeated events. Such patients had more than double the risks of hospitalizations (OR=2.22) and 35% increased rates of repeat hospitalizations. Patients in the upper third based on the number of diabetic patients that shared the same primary physician had no increase in the risk of emergency department visits; they had about a 28% reduced rate of additional ED visits if one occurred. Patients in the upper third based on sharing all the same physicians during the year fared the best; they had 70% decreased odds of one more ED visits and a 30% lower rate of repeated events (Table 2). The odds of a hospitalization for patients in the upper third were about a fourth of patients in the lowest third.
Table 2.
Associations between the characteristics of patient networks and the relative odds of emergency department (ED) visits and hospitalizations for having one or more outcome and for having repeat outcomes among patients having one or more
| Characteristics of Patient Networks | Outcome | Tertile of Number Shared | Relative Odds of | Relative Rates | ||
|---|---|---|---|---|---|---|
| One or More Outcomes | of Repeat Outcomes | |||||
|
|
|
|||||
| OR (95% CI) | P value | RR (95% CI) | P value | |||
| Patients sharing one or more | ED visits | Highest | 3.61 (2.67, 4.87) | < 0.001 | 0.82 (0.65, 1.04) | 0.10 |
| physicians | Middle | 1.70 (1.22, 2.36) | < 0.001 | 0.71 (0.53, 0.95) | 0.02 | |
| Lowest | 1 | -- | 1 | -- | ||
| Hospitalizations | Highest | 2.22 (1.67, 2.95) | < 0.001 | 1.35 (1.19, 1.54) | 0.00 | |
| Middle | 1.04 (0.80, 1.36) | 0.76 | 1.06 (0.92, 1.22) | 0.39 | ||
| Lowest | 1 | -- | 1 | -- | ||
| Patients seeing the same | ED Visits | Highest | 1.12 (0.86, 1.46) | 0.40 | 0.72 (0.59, 0.88) | < 0.001 |
| primary physician | Middle | 1.12 (0.86, 1.46) | 0.40 | 0.97 (0.87, 1.07) | 0.53 | |
| Lowest | 1 | -- | 1 | -- | ||
| Hospitalizations | Highest | 0.97 (0.75, 1.26) | 0.84 | 0.95 (0.86, 1.06) | 0.40 | |
| Middle | 0.97 (0.75, 1.26) | 0.83 | 0.97 (0.87, 1.07) | 0.53 | ||
| Lowest | 1 | -- | 1 | -- | ||
| Patients seeing all of the | ED Visits | Highest | 0.30 (0.22, 0.41) | < 0.001 | 0.75 (0.60, 0.95) | 0.02 |
| same physicians | Middle | 0.42 (0.30, 0.58) | < 0.001 | 0.96 (0.76, 1.21) | 0.74 | |
| Lowest | 1 | -- | 1 | -- | ||
| Hospitalizations | Highest | 0.29 (0.22, 0.40) | < 0.001 | 0.88 (0.78, 1.00 | 0.06 | |
| Middle | 0.31 (0.23, 0.42) | < 0.001 | 0.83 (0.73, 0.94) | < 0.001 | ||
| Lowest | 1 | -- | 1 | -- | ||
Figure 2.
Associations with being in the upper third for the number other patients sharing one or more physicians (All Physicians), the number of other patients seeing the same primary physician (Primary Physicians), and the number of other patients seeing all of the same physicians (Total Number).
A final regression model tested the independent effects of the predictors. Analyses were limited to indicators of being in the highest vs. the lowest or middle thirds (Table 3, Figure 2). For ED visits, the biggest differences were a more than doubling of the odds among patients in the upper third based on the number of shared physicians (OR=2.89), and about a 30% reduction in odds among patients most often seeing all the same physicians. For hospitalizations a more than doubling in odds occurred among in the upper third based on the number of shared physicians. Patients most often sharing all the same physicians had a 30% decrease in odds.
Table 3.
Associations with being in the upper third for emergency department (ED) visits and hospitalizations by the number other patients sharing physicians, the number of other patients seeing the same primary physician, and the number of other patients seeing all of the same physicians
| Relative Odds of one or more outcome | Relative Rates of Repeat outcomes | ||||
|---|---|---|---|---|---|
|
|
|||||
| Predictor | Outcome | OR (95% CI) | P value | RR (95% CI) | P value |
| Number other patients sharing physicians | ED visits | 2.89 (2.21, 3.77) | < 0.001 | 1.01 (0.85, 1.21) | 0.89 |
| Number of other patients seeing the same primary physician | 1.04 (0.80, 1.36) | 0.78 | 0.80 (0.68, 0.95) | 0.01 | |
| Number of other patients seeing all of the same physicians | 0.70 (0.53, 0.94) | 0.02 | 0.80 (0.64, 0.99) | 0.044 | |
| Number other patients sharing physicians | Hospitalizations | 2.16 (1.69, 2.77) | < 0.001 | 1.32 (1.19, 1.47) | < 0.001 |
| Number of other patients seeing the same primary physician | 0.91 (0.73, 1.15) | 0.43 | 0.95 (0.86, 1.04) | 0.24 | |
| Number of other patients seeing all of the same physicians | 0.61 (0.48, 0.78) | < 0.001 | 0.95 (0.85, 1.06) | 0.36 | |
Patients sharing the most physicians averaged 1.37 ± 0.76 primary care physicians and 0.71 ± 1.14 specialists during the year. By contrast patients in the largest networks based on sharing the same primary physician averaged 1.11 ± 0.59 primary care providers and 0.39 ± 0.86 specialists, and patients seeing all the same providers during the year averaged 0.95 ± 0.38 primary care providers and 0.15 ± 0.49 specialists.
Discussion
The study results demonstrate that networks with similarities in care experience similar risks of adverse health outcomes. Patients in the largest networks formed by sharing all the same physicians had reduced risks of hospitalizations and ED visits. They incurred lower rates of repeated ED visits. Such patients saw the fewest primary care providers and specialist physicians; the patients may share a strong loyalty to their primary care providers, providers who may refer their patients to the same specialists, if needed. An international study reported continuous primary care may help reduce rates of ED use (15).
Patients in the upper third based on the number of shared physicians had more than double the odds of ED visits and hospitalizations. They had worse outcomes despite seeing the broadest range of physicians. Sharing more physicians did not ensure better care. Patients in practices in the upper third based on the number of diabetes patients managed had a lower rate of repeat of ED visits. Patients in the largest practices did not incur lower rates of hospitalizations.
Several previous studies have looked at the networks of physicians treating hospitalized patients. Physicians agree that they share patients identified in Medicare data, and their recognition increases with the number of patients shared (16). Patient sharing is the greatest when physicians are based at the same hospital, in geographic proximity, or have similar patient panels such as racial composition or comorbidity (17). Within hospitals greater connectivity was associated with greater spending, longer stays, and more physician visits (18,19). At the same time, hospitals whose primary care providers had higher centrality incurred lower spending on imaging and tests; and having more primary care providers in the networks was associated with fewer specialist and emergency department visits. These studies show the promise and challenge of working with networks in healthcare (20).
Metrics based on providers sharing patients are associated with enhanced teamwork (21), better response to guidelines (22), and lower inpatient and outpatient costs and rates of hospitalizations (23). Statistical models based on physician collaboration networks of patient sharing were used to predict hospitalization cost and length of stay (24, 25). One study investigated the possibility of configuring care teams from physician-physician networks using administrative data (26). For outpatient settings however physician teamwork is more limited and may challenge applications of provider networks (27).
Limitations and strengths.
First, the results are based on associations across a year and cannot be interpreted as causal. The associations between network structures and adverse outcomes occur cross-sectionally in a single year. The network structures cannot be assumed to cause (or prevent) adverse events over time. The study is exploratory and the patients are from a single large insurer in a single state, and one with a distinct geographical location and ethnic population. Our population is younger than the Medicare populations examined in other studies. The results could be biased by the patient’s health or other factors not considered. The free choice of patients in the study provides a strength in revealing patterns based on the natural variability of patient choices. The definition of diabetes met the quality standards of the HEDIS program.
Conclusion.
Diabetes patients should be encouraged to find a provider they like and trust and to stay in the provider’s care. This practice may help reduce the patients’ risk of adverse outcomes. At the same time, physicians should make efforts to build loyalty among their patients.
Acknowledgements
The research was supported by grants U54MD007584 and U54MD007601 from NIH/NIMHD and U54GM104944 from NIH/NIGMS.
Footnotes
The authors have no conflicts of interest
References
- 1.Weir GC, Bonner-Weir S. Five stages of evolving beta-cell dysfunction during progression to diabetes. Diabetes. 2004;53 Suppl 3:S16–21. [DOI] [PubMed] [Google Scholar]
- 2.Fonseca VA. Defining and characterizing the progression of type 2 diabetes. Diabetes care. 2009;32 Suppl 2:S151–6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.American Diabetes Association. 4. Lifestyle Management: Standards of Medical Care in Diabetes-2018. Diabetes Care. 2018;41(Suppl 1):S38–s50. [DOI] [PubMed] [Google Scholar]
- 4.Mitri J, Gabbay R. Understanding Population Health Through Diabetes Population Management. Endocrinology and Metabolism Clinics of North America. 2016;45(4):933–42. [DOI] [PubMed] [Google Scholar]
- 5.Davis J LE, Taira DA, Chen J. Healthcare Network Analysis of Patients With Diabetes and Their Physicians. Am J Managed Care. In press. [PMC free article] [PubMed] [Google Scholar]
- 6.Luke DA, Harris JK. Network analysis in public health: history, methods, and applications. Annual Review of Public Health. 2007;28:69–93. [DOI] [PubMed] [Google Scholar]
- 7.O’Malley AJ. The analysis of social network data: an exciting frontier for statisticians. Statistics in Medicine. 2013;32(4):539–55. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.O’Malley AJ, Marsden PV. The Analysis of Social Networks. Health services & Outcomes Research Methodology. 2008;8(4):222–69. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.HEDIS and Performance Measurement [NCQA web site]. Available at: https://www.ncqa.org/hedis/. Accessed July 20, 2019.
- 10.Taira DA, Seto BK, Davis JW, et al. Examining Factors Associated With Nonadherence And Identifying Providers Caring For Nonadherent Subgroups. Journal of Pharmaceutical Health Services Research: an official journal of the Royal Pharmaceutical Society of Great Britain. 2017;8(4):247–53. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Clark DO, Von Korff M, Saunders K, et al. A chronic disease score with empirically derived weights. Medical Care. 1995;33(8):783–95. [DOI] [PubMed] [Google Scholar]
- 12.Perry BL PB, Borgatti SP. Egocentric Network Analysis Foundations, Methods, and M0dels2018. [Google Scholar]
- 13.Borgatti SP EM, Johnson JC. Analyzing Social Networks2013. [Google Scholar]
- 14.Farewell VT, Long DL, Tom BDM, Yiu S, Su L. Two-Part and Related Regression Models for Longitudinal Data. Annual Review of Statistics and Its Application. 2017;4:283–315. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.van den Berg MJ, van Loenen T, Westert GP. Accessible and continuous primary care may help reduce rates of emergency department use. An international survey in 34 countries. Family Practice. 2016;33(1):42–50. [DOI] [PubMed] [Google Scholar]
- 16.Barnett ML, Landon BE, O’Malley AJ, et al. Mapping physician networks with self-reported and administrative data. Health Services Research. 2011;46(5):1592–609. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Landon BE, Keating NL, Barnett ML, et al. Variation in patient-sharing networks of physicians across the United States. JAMA. 2012;308(3):265–73. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Barnett ML, Christakis NA, O’Malley J, et al. Physician patient-sharing networks and the cost and intensity of care in US hospitals. Medical Care. 2012;50(2):152–60. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Landon BE, Keating NL, Onnela JP, et al. Patient-Sharing Networks of Physicians and Health Care Utilization and Spending Among Medicare Beneficiaries. JAMA internal medicine. 2018;178(1):66–73. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Lewis VA, Fisher ES. Social networks in health care: so much to learn. JAMA. 2012;308(3):294–6. [DOI] [PubMed] [Google Scholar]
- 21.Carson MB, Scholtens DM, Frailey CN, et al. Characterizing Teamwork in Cardiovascular Care Outcomes: A Network Analytics Approach. Circulation Cardiovascular Quality and Outcomes. 2016;9(6):670–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Moen EL, Austin AM, Bynum JP, et al. An analysis of patient-sharing physician networks and implantable cardioverter defibrillator therapy. Health Services & Outcomes Research Methodology. 2016;16(3):132–53. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Pollack CE, Weissman GE, Lemke KW, et al. Patient sharing among physicians and costs of care: a network analytic approach to care coordination using claims data. Journal of General Internal Medicine. 2013;28(3):459–65. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Uddin S, Kelaher M, Piraveenan M. Impact of Physician Community Structure on Healthcare Outcomes. Studies in Health Technology and Informatics. 2015;214:152–8. [PubMed] [Google Scholar]
- 25.Uddin S. Exploring the impact of different multi-level measures of physician communities in patient-centric care networks on healthcare outcomes: A multi-level regression approach. Scientific Reports. 2016;6:20222. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Ito M, Appel AP, de Santana VF, Moyano LG. Analysis of the Existence of Patient Care Team Using Social Network Methods in Physician Communities from Healthcare Insurance Companies. Studies in Health Technology and Informatics. 2017;245:412–6. [PubMed] [Google Scholar]
- 27.Mandl KD, Olson KL, Mines D, et al. Provider collaboration: cohesion, constellations, and shared patients. Journal of General Internal Medicine. 2014;29(11):1499–505. [DOI] [PMC free article] [PubMed] [Google Scholar]