An Evaluation of Recurrent Diabetic Ketoacidosis, Fragmentation of Care, and Mortality Across Chicago, Illinois

Diabetic ketoacidosis (DKA) is an acute, life-threatening complication of uncontrolled diabetes, classically characterized by the triad of uncontrolled hyperglycemia, metabolic acidosis, and ketosis, and can be seen in type 1 diabetes mellitus (T1DM) and type 2 diabetes mellitus (T2DM). DKA is also considered a severe complication of diabetes worldwide and has been shown to be an independent predictor of short-term mortality outside the U.S. (1). DKA hospitalizations have increased 40% in the last decade, and there were an estimated 140,000 DKA hospitalizations in 2009 in the U.S. (2). One estimate found that hospitalization for DKA results in medical expenditures of greater than $2.4 billion (3).

DKA is most commonly caused by nonadherence or insufficient insulin therapy but can also be caused by other medical events. One recent study examining two community hospitals identified 84% of cases were caused by omission of insulin, whereas medical illness only represented 31% of those admissions (4). A subset of patients has been found to have extremely high rates of hospitalization for DKA, and case reports have identified patients who are hospitalized as many as 11 times annually for DKA (5). This population represents a very high-risk diabetes population from a public health perspective. In addition, factors associated with recurrent DKA, including lack of adherence to therapy, low socioeconomic status, substance abuse, and low education (6,7), are also associated with patients who utilize care across multiple health systems, leading to added difficulties in long-term continuity and coordination of care. Health care fragmentation itself is also associated with an increase in hospital admissions and higher utilization of health care (8). However, there are few reports of regional data examining the incidence of recurrent DKA across multiple hospital or health care systems, fragmentation of care, and mortality. This study evaluated recurrence of DKA and fragmentation of care and its effect on all-cause mortality across multiple hospital systems for a large cohort of patients in Chicago.

The Chicago HealthLNK Data Repository (CHDR) is an electronic health record (EHR) linkage tool combining records from five large academic centers (University of Chicago Medical Center, Loyola University Medical Center, Northwestern Medicine, Rush University Medical Center, and University of Illinois at Chicago Medical Center), one large county health care system (Cook County Health and Hospital Systems), and a network of community health centers with multiple outpatient care sites (Alliance of Chicago). Data from the University of Chicago Medical Center and the Alliance of Chicago outpatient care sites were not included in the DKA analyses. Outpatient data were only included for the definition of fragmentation, as noted in a subsequent section. The CHDR uses a secure, Health Insurance Portability and Accountability Act (HIPAA)–compliant, Secure Hash Algorithm-512 hashing algorithm and a probability-weighted matching process to create a nonduplicated and merged set of EHR data spanning multiple sites in a large metropolitan area. The linkage process has been previously validated with a manually reconciled set of 11,292 patients (9). Appropriate institutional review board approval was obtained at each institution.

For this study, patients with diabetes were identified in the CHDR by querying all ICD-9 classification codes for diabetes (250.xx) for the years 2006 to 2012. Patients with DKA were defined as having one or more inpatient hospital encounters with an ICD-9 code of 250.1x (x = 0, 2 T1DM; x = 1, 3 T1DM). Inclusion criteria included age 18–89 and a 606XX or surrounding zip code (Greater Chicago area). Repository data included demographic information (birth year, sex, race, ethnicity, insurance status, zip code) and visit information (inpatient DKA encounters). Indication of death was obtained from the Social Security Death Index record, which was merged into the CHDR using a previously reported hashing and matching algorithm (9). Insurance status was defined as the last status on file. Presence of microvascular complications was defined as any ICD-9 code for microvascular complications (nephropathy: 250.4x, retinopathy: 250.5x, and neuropathy: 250.6x).

Patients with inpatient DKA encounters were separated into three groups of 1, 2–3, or ≥4 DKA encounters during the study period of 2006–2012. A recurrent patient was defined as a patient who had more than one hospitalization for DKA during the study period. Fragmentation was defined as hospitalization for DKA at more than one hospital during the study period. This measure of health care fragmentation was previously used by Galenter et al. (8). In addition, to further describe patterns of fragmentation, the primary site of care for each patient was determined, as defined by the patient’s institution with the highest volume of encounters of any type (inpatient, outpatient, or Emergency Department), for any diagnosis, among the institutions included in the study.

The general HDR population does not reflect the wider demographic distribution of Chicago; rather, it reflects the demographic distribution of the patient population served by the institutions in the data set. However, other analyses have shown that these demographics are comparable to the actual demographics of the geographical areas chosen and that the hospitals within CHDR serve most of the population within the given zip codes (9).

Although data on BMI, medications, and other clinical and laboratory values are available, this information in the CHDR was incomplete over time. In addition, for demographic data that was subject to change over time, such as insurance status, only the most recent value was stored and, therefore, may not represent a patient’s insurance status at the time of an earlier event. For mortality data, the presence of a death record was the only information available; thus, no information could be gathered on cause or date of death.

ICD-9 diagnosis codes for diabetes are separated by “T1DM” (250.x1 and 250.x3) or “T2DM or unspecified” diabetes (250.x0 and 250.x3).

The analytical statistics were calculated using SAS 9.4 software. Cross-tabulation χ² statistics were calculated to determine significant differences of patient counts among the groups. Logistic regression was performed to model the effect of health care fragmentation on total count of DKA visits where the visit count >1, and was adjusted for covariates of interest including age, sex, race, and insurance status. In addition, logistic regression was performed to analyze effect of DKA visit counts of 1, 2–3, or ≥4 (outcome) and fragmentation status (exposure) on mortality (outcome). To address any multicolinearity issues between DKA group and fragmentation, we ran two additional logistic regression models, one with DKA group and one with fragmentation as the main effect.

Of the 1,969,283 patients in the CHDR, 189,908 patients (9.6%) with diabetes were identified. Table 1 summarizes demographics of these patients and patients with DKA. Differences in demographic characteristics among the DKA groups are also reported.

Of the 189,908 patients with diabetes, 3,615 patients (1.9%) were admitted for DKA. Only 756 patients (24.7%) were labeled solely with T1DM–associated DKA codes, and 1,808 patients (59.1%) were labeled as DKA “T2DM or unspecified.” In addition, 493 patients (21.4%) were assigned both T1DM and T2DM ICD-9 acidosis codes at some point in their admission histories, with an increasing percentage of dual codes noted in those with a higher number of DKA encounters (data not shown). We found similar results for all other diabetes-related ICD-9 codes for patients with DKA: for all codes, only 5.4% of patients with DKA were associated with solely T1DM acidosis codes, 42.2% were labeled as “T2DM or unspecified,” and 52.4% were dual coded. Therefore, we did not attempt to analyze the data according to the type of diabetes.

The 3,615 patients had a total of 5,591 encounters for DKA. Patients with only 1 DKA visit were the largest group, at 2,835 patients (78.4%), 569 (15.7%) had 2–3 episodes, and 211 (5.8%) had ≥4 episodes of DKA. Although patients with ≥4 episodes represented only 5.8% of the total DKA patients, they accounted for 26.3% of encounters.

Patients with recurrent DKA more commonly self-identified as African American/black and were on Medicare or Medicaid or reported no insurance. See Table 1 for the demographics of those with diabetes and those with DKA. Of 3,615 patients with any DKA, 780 had recurrent DKA (>1), and 125 (16.0%) experienced fragmented care (hospitalized for DKA >1 institution). A higher proportion of patients in the ≥4 DKA group (59 [28.0%]) had fragmented care, compared with patients in the 2–3 DKA group (66 [11.6%], P < 0.0001), and the odds of having fragmented care were increased in the ≥4 group (odds ratio [OR] 2.96; 95% CI 1.99, 4.39; P < 0.0001). In addition, 35.1% of fragmented encounters were determined to be outside the patient’s “primary site” of health care. However, 99.1% of patients with recurrent DKA were hospitalized at two hospitals. Only 11 patients were hospitalized at more than two institutions, with all 11 in the ≥4 DKA group (8 at 3 institutions, 2 at 4 institutions, and 1 at 6 institutions). After adjustment for age, sex, race, and insurance status, fragmentation of care increased the DKA visit count 1.88-fold (P < 0.0001). Mean HbA_1c in those with recurrent DKA (with laboratories in CHDR at the time closest to last DKA encounter n = 2,498) was similar among those with any DKA, but was significantly increased in those who had fragmented care (11.10 ± 2.9) versus nonfragmented care (10.89 ± 5.2, P < 0.0001).

Table 2 summarizes DKA fragmentation of care and mortality data. Overall, 490 DKA patients (13.6%) died during the study period. Number of DKA encounters (OR 1.28; 95% CI 1.04, 1.58; P = 0.02) and age (OR 1.06; 95% CI 1.05, 1.07; P < 0.0001) were associated with death in the fully adjusted model, which included fragmentation, DKA visit count, age, sex, race, and insurance status. Care fragmentation was not a significant factor in the model. To address any multicolinearity issues between DKA group and fragmentation, we ran two additional logistic regression models, one with DKA group and one with fragmentation as the main effect. Point estimates and significance did not change significantly from the full model, suggesting that any effects we see in the final model are not driven by a multicolinearity issue. Odds of death were also decreased for African American and Hispanic patients (OR 0.76; 95% CI 0.58, 0.99; P = 0.0418) compared with white patients (OR 0.42; 95% CI 0.24, 0.73; P = 0.002).

Significant attention has been paid to the presence and need for reduction of DKA in the pediatric population and other international populations (10,11) but has been limited in the adult U.S. population. Although DKA as a complication was only identified in a small number of patients (1.9% of entire CHDR patients with diabetes), this population had a large number of primary high-acuity encounters for a potentially avoidable complication; in addition, it represents a population with multiple diabetes-related complications and significant mortality. Recurrence was common, with recurrent patients representing a disproportionate percentage of admissions for DKA. Notably, patients with ≥4 DKAs were only 5.8% of the total DKA group but represented more than a quarter of all admissions. Unique to this study but not unexpected, we found that 28.0% of these patients with ≥4 DKAs had fragmented hospital care. Conversely, patients with fragmented care were more likely to have higher frequencies of DKA recurrence. We acknowledge that socioeconomic confounders may play a significant role. In addition, patients with fragmented care often were hospitalized at a center that was not their primary site of health care, as determined by location of the highest volume of encounters. More than one-third of their hospitalizations met this description, indicating health care fragmentation was a clinically relevant aspect of their health care rather than an occasional event. Also worth noting is that any measure of recurrence and fragmentation in this data set is likely an underestimate, because only five institutions were represented.

A substantial portion (13.6%) of patients with DKA died of any cause during this 6-year period. Interestingly, patients with more DKA episodes were younger on average, yet mortality was associated with increasing age and number of admissions for DKA. Mortality appeared to be greater in Caucasian than African American or Hispanic patients with DKA, but the reasons for this are not clear. Owing to the lack of patient-level data, we do not have information regarding other comorbidities that might have contributed to the mortality. There are no databases with sufficient numbers of patients dying in DKA in the literature to be able to determine whether our findings are unique. Fragmentation of care in those with DKA was not associated with mortality. However, the relationship between recurrence and fragmentation is clearly complex. Most notably, this population is relatively small in the context of the larger Chicago population, and recurrence is generally isolated to movement across two institutions. This cohort of patients with diabetes are clearly a very high-risk group, in potential cost and complications, and early identification and tracking may be of great benefit from the public health perspective.

The strengths of this study include its substantial sample size and the novel methodology by which patients are identified (while maintaining HIPAA compliance) and then deidentified across multiple institutions in the same region in the U.S. Use of the EHR, across multiple institutions, allowed us to have a timely and accurate representation of fragmentation and recurrent DKA, and subsequent short-term mortality, across a large U.S. metropolitan city. In addition, it identified a high-risk population of fragmented patients with recurrent DKA, with increased risk of death with additional insurance data, revealing a largely publically funded insurance/missing insurance cohort.

Among the limitations of this study, the clearest is the lack of patient-level medical record data. All patient information was extracted into a structured, anonymized database, and individual clinical histories and notes were not accessible. Therefore detailed descriptions about patient histories, including details about concomitant psychosocial stressors, adherence measures, and other comorbidities, were unavailable. Gathering such data at each individual institution may be possible in future studies; however, this would likely be very resource intensive and require significant cross-institution collaboration. Such an initiative could greatly improve such data sources.

Most of the limitations of this study are reflective of the limitations of the EHR data itself. Medical billing and ICD-9 coding may not accurately reflect clinical care; for example, the type of diabetes was difficult to ascertain in this large subgroup of patients, as has been previously reported (12). As we noted, 21.4% of patients were assigned both T1DM and T2DM ICD-9 acidosis codes at different points in their admissions. In a recent study at our institution, we found there was substantial disagreement with the designation of type of diabetes between an endocrinologist’s review of the admission and what was written in the patient record by the attending physician and the ICD-9 code assigned (13). There was also substantial disagreement even with the diagnosis of DKA (13). These types of diagnostic discrepancies need to be evaluated when analyzing all large databases extracted from EHRs. To improve consistency in reporting disease states and events between institutions, cross-institution collaborations and working groups would need to be set up in a prospective manner.

The data set is also limited in that it is solely the experience of medical centers in Chicago. The data described here may therefore not reflect the experiences of other regions or community hospitals although it may be reflective of other large urban populations with multiple health systems. However, the potential use of the EHR linkage tool used in this data set (9) could allow for other registries, institutions, or related systems to safely share HIPAA-compliant data in this or other disease states. Lastly, we acknowledge, there are other potential confounders that this study was unable to address, including diagnosis of T1DM versus T2DM diabetes, psychiatric diagnoses, and substance abuse.

This study should prompt further investigation into this particularly high-risk group of patients with diabetes (those with recurrent DKA), with particular attention to patients with fragmented health care who represent a unique challenge for continuity of care. Recurrent DKA represents a challenge in diabetes care and a potentially avoidable one. Concentration on this small but particularly high-risk group may help to curtail acute care utilization and diabetes costs overall. However, this study does open up new questions about acute care utilization in this cohort and the possible effects of geography and location of patients and institutions on fragmentation and recurrence. As a start, individual institutions should try to identify such high-risk patients with diabetes early and use resources to intervene early on in their disease course. For example, a diabetes consultation or diabetes educator could attempt to determine the cause of the DKA in every DKA admission to provide education to prevent recurrence, or a hospitalist team could use clinical decision support tools to adequately diagnose patients. For those with recurrent DKA, a more intensive intervention, such as the “Multisystemic Therapy” outlined by Ellis et al. (14,15) or the Novel Interventions in Children’s Healthcare (16,17), could be undertaken. Although these are pediatric interventions, they could be adapted for care in adults as well. These interventions involve patient and family education, coordination of care, cognitive-behavioral intervention, and family systems approaches (18). Newer approaches using newer technologies involving video conferencing and text messaging have also been advocated (18). In addition, members of our group have previously reported use of regional informatics/health information exchanges for improved care coordination in the area of antibiotic-resistant infections (19,20).

From an economic perspective, the cost for a hospitalization for DKA was U.S. $6,444 in 1994 (21); when adjusted for health care inflation, this increases to $14,180 in 2016 U.S. dollars (22). Therefore, for our 211 patients with ≥4 admissions for DKA totaling 1,470 episodes, this would calculate to $20,844,600 or $98,790 for each of the 211 patients. If even 50% of this figure were used for interventions, a considerable cost savings would accrue.

Patients with fragmented care may utilize health care differently from patients who limit their care to one center. Therefore, unique approaches requiring collaboration between institutions may also be necessary to improve the health of this particular group of patients. This study, in addition to identifying the scope and toll of recurrent DKA on patients in Chicago, identifies just such a population of patients for whom efforts at prevention and follow-up may require a more tailored approach.

Acknowledgments. The authors acknowledge all participating centers that contribute data to CHDR so such studies can be possible. This includes University of Chicago Medical Center (David Meltzer), Loyola University Medical Center (Frances Weaver), Northwestern Medicine (Abel N. Kho), Rush University Medical Center (Bala Hota), University of Illinois at Chicago Medical Center (Bill Galanter), Cook County Health and Hospital Systems (William Trick), and Alliance of Chicago (Fred Rachman).

Duality of Interest. M.E.M. currently receives research and/or grant support from Bayer, Novo Nordisk, Novartis, Ipsen, and Johnson & Johnson and serves as a consultant for Novo Nordisk, Novartis, Ipsen, Janssen, Merck, Pfizer, and Takeda. A.W. currently receives grant support from Merck, Johnson & Johnson, and Bayer. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. J.A.M. helped with the design, researched the data, contributed to the analysis, and wrote the manuscript. K.L.J. contributed to the analytic plan, completed data analysis, and edited the manuscript. T.A.D. and J.J.B. and helped complete the analysis and reviewed the manuscript. S.G., M.E.M., and A.N.K. contributed to the overall research design, contributed to the discussion, and reviewed and edited the manuscript. A.W. researched the data, designed the study and analytic plan, oversaw the analysis, and wrote and edited the manuscript. A.W. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Prior Presentation. Parts of this study were presented orally at the 75th Scientific Sessions of the American Diabetes Association, Boston, MA, 5–9 June 2015.