Developing and validating a diabetes database in a large health system

doi:10.1016/j.diabres.2006.07.007

Diabetes Research and Clinical Practice

Volume 75, Issue 3, March 2007, Pages 313-319

https://doi.org/10.1016/j.diabres.2006.07.007 Get rights and content

Abstract

Background

One component of clinical information systems is a registry of patients. Registries allow providers to identify gaps in care at the population level. Registries also allow for rapid cycle continuous quality improvement, targeted practice change and improved outcomes. Most registries are built based on membership with an insurer or other selection criteria. Little, if any data exist on registries representing demographically heterogeneous populations.

Methods

Administrative and clinical data for the period 1/1/2000–12/30/03 were examined. In total, 46,082,941 lab reports, 233,292,544 medical records, and 9,351,415 medical record abstracts, representing approximately 2 million unique patients were searched. The diabetes source population was identified by presence of any one of the following criteria: ICD-9 code 250 (diabetes) for inpatient, emergency room or outpatient visits; any hemoglobin A1c result; blood glucose >200 mg/dl; or diabetes medication. A diagnosis of diabetes was verified by trained chart reviewers on a sample of patients. Single indicators and combinations were examined to determine optimal identification of these cases.

Results

In two separate validation studies, using two or more indicators or outpatient diagnosis maximized positive predictive value (PPV) (96 and 97%) and sensitivity (99 and 100%) and identified 55,807 individuals. When all patients with a single indicator of outpatient diagnosis (which had the highest single PPV of 94 and 95%) were included together with those having ≥2 indicators, the final sample size was 65,725.

Conclusion

Two or more indicators or an out-patient-diagnosis identifies a sizeable and unselective diabetes database which can be used to track processes and outcomes.

Introduction

There are currently 18 million people in the U.S. with diabetes. This number is expected to rise to 30.3 million people by the year 2030 [1]. Diabetes thus represents a major health burden, especially as an increase in the prevalence of diabetes complications is also anticipated, given the current level of suboptimal diabetes care in this country [2]. Diabetes therefore poses a major challenge to healthcare systems.

The Chronic Care Model [3] outlines the components of a health care delivery system that promotes quality disease management and provides a potential useful framework with which to respond to this challenge. The use of clinical information systems is a key element of this model [2], [4], [5], [6]. Information systems that contain data on the level of care, including both processes and outcomes, offer a valuable tool for health systems. These information systems allow for continuous quality improvement, practice change, and improved outcomes. However traditional systems used to track quality of care, such as chart audit or patient and provider self-report, lack internal validity, are expensive, and inefficient.

A number of diabetes registries are currently available in the United States (Table 1). They have examined key issues, including quality of care [7], sociodemographic differences in health care delivery [7], [8], [9], outcomes [10], [11], risk stratification [12], and costs [13], [14]. In general, these databases are derived from homogenous populations, defined by a health plan, insurer or membership in another group, limiting the translation of methodologies to other populations. Further, specific criteria employed to identify people with diabetes vary within these registries, making comparisons between populations and registries difficult. It would thus be helpful to develop a registry that reflects a more general population that receives diabetes care in a variety of provider settings, and includes a broad spectrum of patient demographics and insurers. It would also facilitate further study to develop a standardized method of identifying patients with diabetes, in such settings, in order to optimize estimation of cases, and comparison of findings.

With these considerations in mind, our objectives were to: (1) develop a diabetes registry using administrative and clinical data in a major health care system serving a large heterogeneous population; (2) develop a set of criteria that most reproducibly and accurately identifies people with diabetes.

Section snippets

The setting

The University of Pittsburgh Medical Center (UPMC) is a large health care system and academic medical center in western Pennsylvania. UPMC is one of the largest non-profit integrated health care systems in the United States comprised of 19 tertiary care, specialty and community hospitals, physician practices, imaging and surgery facilities, post-acute rehabilitation and in-home services, long-term care facilities, pharmacy services, a managed care health plan, and a Community Health Division.

Validation study

Results of the validation studies are displayed in Table 2. In the first validation the six screening criteria were applied to patients from three local outpatient clinics (endocrine, internal medicine, and family practice). A total of 254 charts were reviewed. In this test-population, with at least one potential indicator of diabetes, 54.6% of patients were confirmed to have diabetes. When indicators were combined, as the number of indicators increased, the PPV and specificity increased, while

Discussion

As part of our efforts to improve diabetes care in our healthcare system we developed a diabetes registry as a key first step. This project involved the unique opportunity to use currently existing administrative and clinical data representing a heterogeneous diabetic population. This registry is applicable to a broad range of ages, races, and insurers, reflective of the general diabetes population of western Pennsylvania. To the best of our knowledge, this diabetes registry is the first

References (20)

K.M. Newton et al.
The use of automated data to identify complications and comorbidities of diabetes: a validation study
J. Clin. Epidmiol.
(1999)
S. Wild et al.
Global prevalence of diabetes. estimates for the year 2000 and projections for 2030
Diabetes Care
(2004)
J.B. Saaddine et al.
A diabetes report card for the United States: quality of care in the 1990s
Ann. Intern. Med.
(2002)
E.G. Wagner
Meeting the needs of chronically ill people
BMJ
(2001)
Institute for Health Care Improvement, vol. 2005, Institute for Health Care Improvement,...
E.H. Wagner et al.
Improving outcomes in chronic illness
Managed Care Quarterly
(1996)
E.H. Wagner et al.
Organizing care for patients with chronic illness
The Mill Bank Quarterly
(1996)
D.R. Arday et al.
Variation in diabetes care among states. Do patient characteristics matter?
Diabetes Care
(2002)
A.M. McBean et al.
Racial variation in the control of diabetes among elderly medicare managed care beneficiaries
Diabetes Care
(2003)
T.L. Gary et al.
Racial comparisons of health care and glycemic control for African American and white diabetic adults in an urban managed care organization
Dis. Manag.
(2004)

There are more references available in the full text version of this article.

Cited by (62)

Validation of Diagnostic Coding for Diabetes Mellitus in Hospitalized Patients
2022, Endocrine Practice
Some studies have shown that there is an undercoding of diabetes mellitus among hospitalized patients, which can have adverse clinical and financial implications for health systems. We aimed to validate the discharge diagnostic coding of diabetes mellitus in hospitalized patients using clinical and laboratory-based diagnostic indicators as the reference.
This was a retrospective cohort study of 83 690 discharges of 48 615 unique adult patients who were hospitalized in an academic medical center over 4.5 years and had at least 4 blood glucose measurements during admission. A missing diabetes code (MDC) was defined using 2 criteria. MDC₁ was defined as the presence of any of the following: blood glucose ≥200 (x2), A1C ≥6.5%, home antihyperglycemic medication, or preadmission code for diabetes, whereas MDC₂ was defined as preadmission diabetes or at least 2 other criteria in MDC₁. Multivariable logistic regression was used to identify factors associated with MDC compared to the present diabetes code.
MDC₁ and MDC₂ were present in 12 186 (14.6%) and 3542 (4.7%) discharges, respectively. Factors associated with both MDC₁ and MDC₂ were medium-dose steroid use [adjusted odds ratio (aOR) 2.11, 2.01], high-dose steroid use (aOR 4.70, 2.50), intermediate medical care service (aOR 1.65, 1.55), infection (aOR 1.21, 1.34), and hepatic disease (aOR 1.93, 1.92).
In this retrospective study, MDC ranged from 5% to 15% and was associated with various clinical factors. Further prospective studies are needed to validate these findings, explore the mechanisms behind these associations, and understand the clinical and financial implications.
Identifying Local Hot Spots of Pediatric Chronic Diseases Using Emergency Department Surveillance
2017, Academic Pediatrics
To use novel geographic methods and large-scale claims data to identify the local distribution of pediatric chronic diseases in New York City.
Using a 2009 all-payer emergency claims database, we identified the proportion of unique children aged 0 to 17 with diagnosis codes for specific medical and psychiatric conditions. As a proof of concept, we compared these prevalence estimates to traditional health surveys and registry data using the most geographically granular data available. In addition, we used home addresses to map local variation in pediatric disease burden.
We identified 549,547 New York City children who visited an emergency department at least once in 2009. Though our sample included more publicly insured and uninsured children, we found moderate to strong correlations of prevalence estimates when compared to health surveys and registry data at prespecified geographic levels. Strongest correlations were found for asthma and mental health conditions by county among younger children (0.88, P = .05 and 0.99, P < .01, respectively). Moderate correlations by neighborhood were identified for obesity and cancer (0.53 and 0.54, P < .01). Among adolescents, correlations by health districts were strong for obesity (0.95, P = .05), and depression estimates had a nonsignificant, but strong negative correlation with suicide attempts (−0.88, P = .12). Using SaTScan, we also identified local hot spots of pediatric chronic disease.
For conditions easily identified in claims data, emergency department surveillance may help estimate pediatric chronic disease prevalence with higher geographic resolution. More studies are needed to investigate limitations of these methods and assess reliability of local disease estimates.
The local geographic distribution of diabetic complications in New York City: Associated population characteristics and differences by type of complication
2016, Diabetes Research and Clinical Practice
Citation Excerpt :
We used unique identifiers within SPARCS to identify ED visits by the same individual throughout the study period. As previously published [4], we identified any individual who had ever received a primary or secondary ICD-9 diagnosis code with the prefix 250 during any ED visit including ED admissions [14]. We used the ED patient population as our prior study demonstrated that the demographics of unique ED users is similar to Census estimates after accounting for repeat ED users.
To identify population characteristics associated with local variation in the prevalence of diabetic complications and compare the geographic distribution of different types of complications in New York City.
Using an all-payer database of emergency visits, we identified the proportion of unique adults with diabetes who also had cardiac, neurologic, renal and lower extremity complications. We performed multivariable regression to identify associations of demographic and socioeconomic factors, and diabetes-specific emergency department use with the prevalence of diabetic complications by Census tract. We also used geospatial analysis to compare local hotspots of diabetic complications.
We identified 4.6 million unique New York City adults, of which 10.5% had diabetes. Adjusting for demographic and socioeconomic factors, diabetes-specific emergency department use was associated with severe microvascular renal and lower extremity complications (p-values < 0.001), but not with severe macrovascular cardiac or neurologic complications (p-values of 0.39 and 0.29). Our hotspot analysis demonstrated significant geographic heterogeneity in the prevalence of diabetic complications depending on the type of complication. Notably, the geographic distribution of hotspots of myocardial infarction were inversely correlated with hotspots of end-stage renal disease and lower extremity amputations (coefficients: −0.28 and −0.28).
We found differences in the local geographic distribution of diabetic complications, which highlight the contrasting risk factors for developing macrovascular versus microvascular diabetic complications. Based on our analysis, we also found that high diabetes-specific emergency department use was correlated with poor diabetic outcomes. Emergency department utilization data can help identify the location of specific populations with poor glycemic control.
Association between blood alcohol concentration and mortality in critical illness
2015, Journal of Critical Care
Citation Excerpt :
Diabetes mellitus is defined by ICD-9-CM code 250. xx in the 2-years before hospital discharge [67,68]. Early ICU admission is defined as ICU admission within 48 hours of hospital admission.
In animal models of renal, intestinal, liver, cardiac, and cerebral ischemia, alcohol exposure is shown to reduce ischemia-reperfusion injury. Inpatient mortality of trauma patients is shown to be decreased in a dose-dependent fashion relative to blood alcohol concentration (BAC) at hospital admission. In this study, we examined the association between BAC at hospital admission and risk of 30-day mortality in critically ill patients.
We performed a 2-center observational study of patients treated in medical and surgical intensive care units in Boston, Massachusetts.
Medical and surgical intensive care units in 2 teaching hospitals in Boston, Massachusetts.
We studied 11 850 patients, 18 years or older, who received critical care between 1997 and 2007. The exposure of interest was the BAC determined in the first 24 hours of hospital admission and categorized a priori as BAC less than 10 mg/dL (below level of detection), 10 to 80 mg/dL, 80 to 160 mg/dL, and greater than 160 mg/dL. The primary outcome was all-cause mortality in the 30 days after critical care initiation. Secondary outcomes included 90- and 365-day mortality after critical care initiation. Mortality was determined using the US Social Security Administration Death Master File, and 365-day follow-up was present in all cohort patients. Adjusted odds ratios (ORs) were estimated by multivariable logistic regression models with inclusion of covariate terms thought to plausibly interact with both BAC and mortality. Adjustment included age, sex, race (white or nonwhite), type (surgical vs medical), Deyo-Charlson index, sepsis, acute organ failure, trauma, and chronic liver disease.
Thirty-day mortality of the cohort was 13.7%. Compared to patients with BAC levels less than 10 mg/dL, patients with levels greater than or equal to 10 mg/dL had lower odds of 30-day mortality; for BAC levels 10 to 79.9 mg/dL, the OR was 0.53 (95% confidence interval [CI], 0.40-0.70); for BAC levels 80 to 159.9 mg/dL, it was 0.36 (95% CI, 0.26-0.49); and for BAC levels greater than or equal to 160 mg/dL, it was 0.35 (95% CI, 0.27-0.44). After multivariable adjustment, the OR of 30-day mortality was 0.97 (0.72-1.31), 0.79 (0.57-1.10), and 0.69 (0.54-0.90), respectively. When the cohort was analyzed with sepsis as the outcome of interest, the multivariable adjusted odds of sepsis in patients with BAC 80 to 160 mg/dL or greater than 160 mg/dL were 0.72 (0.50-1.04) or 0.68 (0.51-0.90), respectively, compared to those with BAC less than 10 mg/dL. In a subset of patients with blood cultures drawn (n = 4065), the multivariable adjusted odds of bloodstream infection in patients with BAC 80 to 160 mg/dL or greater than 160 mg/dL were 0.53 (0.27-1.01) or 0.49 (0.29-0.83), respectively, compared to those with BAC less than 10 mg/dL.
Analysis of 11 850 adult patients showed that having a detectable BAC at hospitalization was associated with significantly decreased odds of 30-day mortality after critical care. Furthermore, BAC greater than 160 mg/dL is associated with significantly decreased odds of developing sepsis and bloodstream infection.
Vulnerability of Older Adults in Disasters: Emergency Department Utilization by Geriatric Patients after Hurricane Sandy
2018, Disaster Medicine and Public Health Preparedness
Type 2 Diabetes Mellitus and Periodontitis: Bidirectional Association in Population-based 15-year Retrospective Cohorts
2023, Journal of Clinical Endocrinology and Metabolism

View all citing articles on Scopus

View full text

Developing and validating a diabetes database in a large health system

Abstract

Background

Methods

Results

Conclusion

Introduction

Section snippets

The setting

Validation study

Discussion

J. Clin. Epidmiol.

Global prevalence of diabetes. estimates for the year 2000 and projections for 2030

Diabetes Care

A diabetes report card for the United States: quality of care in the 1990s

Ann. Intern. Med.

Meeting the needs of chronically ill people

BMJ

Improving outcomes in chronic illness

Managed Care Quarterly

Organizing care for patients with chronic illness

The Mill Bank Quarterly

Variation in diabetes care among states. Do patient characteristics matter?

Diabetes Care

Racial variation in the control of diabetes among elderly medicare managed care beneficiaries

Diabetes Care

Racial comparisons of health care and glycemic control for African American and white diabetic adults in an urban managed care organization

Dis. Manag.