Current Medical Research & Opinion 0300-7995 doi:10.1185/03007995.2014.981632
Vol. 31, No. 1, 2015, 107–114
Article ST-0245.R1/981632 All rights reserved: reproduction in whole or part not permitted
Curr Med Res Opin Downloaded from informahealthcare.com by University of Otago on 07/12/15 For personal use only.
Original article Predictors of 30 day hospital readmission in patients with type 2 diabetes: a retrospective, case–control, database study
Elizabeth Eby Former employee of Eli Lilly and Company, Indianapolis, IN, USA
Christine Hardwick Maria Yu Steve Gelwicks Ketra Deschamps Jin Xie Eli Lilly and Company, Indianapolis, IN, USA
Tom George Detroit Medical Center, Sinai-Grace Hospital, Detroit, MI, USA Address for correspondence: Christine Hardwick RN MPH, Eli Lilly Corporate Center, Indianapolis, IN 46285, USA. Tel: +1 (317) 433-6404; Fax: +1 (317) 277-3585; [email protected] Keywords: Diabetes mellitus – Hospitalization – Hospital readmission – Regression analysis – Type 2 diabetes mellitus Accepted: 22 October 2014; published online: 6 November 2014 Citation: Curr Med Res Opin 2015; 31:107–14
Abstract Objective: To assess factors predictive of all-cause, 30 day hospital readmission among patients with type 2 diabetes in the United States. Methods: A retrospective, case–control study using deidentified Humedica electronic health record data was conducted to identify patients 18 years old with 6 months of data prior to index hospitalization (preperiod) and 30 days of data after discharge (post-period). Combined methods of bootstrap resampling and stepwise logistic regression were used to identify factors associated with readmission. Results: Among 52,070 patients with type 2 diabetes and an initial hospitalization for any reason, 5201 (10.0%) patients were readmitted within 30 days and 46,869 (90.0%) patients showed no evidence of readmission. Diabetic treatment escalation; race; type 2 diabetes diagnosis prior to the index stay; pre-period heart failure; and number of pre-period, inpatient healthcare visits were among the strongest predictors of 30 day readmission. From a receiver-operating characteristic plot (mean area under curve of 0.693), the predictive accuracy of the final logistic regression model is considered modest. This result might be due to the unavailability of some variables or data. Conclusions: These results highlight the importance of the appropriate recognition of and treatment for type 2 diabetes, prior to and during hospitalization and following discharge, in order to impact a subsequent hospitalization. In our analysis, escalation of diabetic treatments (especially those escalated from having no records of antidiabetic medications to treatment with insulin) was the strongest predictor of 30 day readmission. Limitations of this study include the fact that hospitalizations and other encounters, outside the Humedica network, were not captured in this analysis.
Introduction Diabetes is associated with a substantial economic burden; total diabetes-related costs were estimated at $245 billion in the US in 2012, including $176 billion in direct medical costs and $69 billion in reduced productivity and other indirect costs1. The largest component of medical costs for diabetes is hospital inpatient care, which accounts for 43% of the total medical cost1. Patients with diabetes are more likely to be hospitalized than patients without diabetes – in 2012, more than 25% of US hospital inpatient stays were incurred by patients with diabetes, and more than 15% of those inpatients stays were ! 2015 Informa UK Ltd www.cmrojournal.com
Predictors of 30 day hospital readmission Eby et al.
107
Curr Med Res Opin Downloaded from informahealthcare.com by University of Otago on 07/12/15 For personal use only.
Current Medical Research & Opinion Volume 31, Number 1
January 2015
directly attributed to diabetes1. Patients with diabetes who are hospitalized incur nearly two-fold higher inpatient costs per capita2. Hospital stays for patients with diabetes are also more likely to follow an emergency room visit, compared to hospitalizations for patients without diabetes3. Additionally, multiple hospitalizations are common among patients with diabetes; these patients with multiple hospitalizations also experience longer hospital stays and considerably higher hospital costs2. The increased risk of hospitalization among patients with diabetes is due, in part, to increased comorbidities among these patients, in addition to health problems associated with diabetes and its complications2,3. Despite this increased hospitalization risk, healthcare providers are often unable to accurately identify patients at risk for readmission4. As part of the Patient Protections and Affordable Care Act (PPACA) passed in 2010, Medicare may reduce Medicare reimbursement payments to hospitals (by reducing Medicare reimbursement payments) with high preventable readmission rates5. These ‘preventable readmissions’ may be difficult to measure, but all-cause hospital readmissions within 30 days of discharge can be evaluated by using electronic health record (EHR) data6. The objective of this study was to assess (by using a large EHR database) the factors predictive of all-cause, 30 day hospital readmission among patients with type 2 diabetes in the US.
Patients and methods Study population The data for this retrospective study were obtained from Humedica’s deidentified EHR database. Humedica partners directly with large medical group practices, integrated delivery networks (IDNs), and hospitals to extract data from their EHRs and various health information technology systems. A subset of the data components include laboratory results, radiology and pathology reports, prescriptions written and dispensed, procedures, diagnoses, and other details of a patient’s office visit and hospital stay. At the time of this study, Humedica’s clinical data warehouse encompassed approximately 13 million patients. The data for this study spanned the period from 1 October 2009 through 30 September 2011, and the data are deidentified to comply with Health Insurance Portability and Accountability Act (HIPAA) regulations. Patients included in the analysis were aged 18 or older and had a diagnosis of type 2 diabetes (i.e., International Classification of Diseases, Ninth Revision [ICD-9] 250.x0 or 250.x2) or met criteria for undiagnosed type 2 diabetes (used medication indicative of type 2 diabetes and/or had laboratory values indicative of diabetes [HbA1c values of 108
Predictors of 30 day hospital readmission Eby et al.
6.5% {48 mmol/mol} or greater, fasting glucose 7.0 mmol/ L, or oral glucose tolerance test 11.1 mmol/L] in the absence of a diagnosis code for diabetes). Patients were also required to have records associated with an IDN; an index hospitalization not within 30 days of a previous inpatient discharge within the study period of 1 September 2009 – 31 August 2011; 6 months of data prior to the index hospitalization (i.e., the pre-period); 31 days of data postdischarge (to create the 30 day post-period); at least one valid blood glucose measurement in the pre-period or during the index hospitalization; and a discharge to their home. Patients were excluded if they had no valid pre- or post-period records; died during the index hospitalization or within 30 days of discharge; had a diagnosis of cancer (except skin cancer); had a diagnosis of type 1, secondary, or gestational diabetes; or were pregnant.
Statistical analyses Baseline variables of the study population were compared between the readmitted and non-readmitted cohorts and included pre-index demographics and disease characteristics (Table 1). These variables were selected based on a review of the literature2,7–14 and of the available variables in the Humedica dataset. We derived the diabetic treatment escalation variable by comparing the prescriptions written during the pre-period with those written during the post-period. Comorbidities present during the index hospitalization were identified by ICD-9 codes. Hypoglycemia episodes were calculated using the methods described by Ginde et al.15. Univariate test statistics (t-test for continuous variables and chi-squared test for categorical variables) were used, and univariate logistic regression was used to obtain odds ratio (OR) estimates of being readmitted. For the multivariate analyses, bootstrap sampling of the data was utilized. Bootstrap resampling is a common statistical method often used in estimating variability or standard error of estimates. In this study, it was used to generate a model training sample and a model validation sample. Bootstrapping is performed by generating a sample with replacement from the original sample. Because observations that are drawn are made available again for resampling, some can be selected repeatedly as part of the bootstrap sample and some may not be selected at all16. Other studies have shown the instability of automated variable selection methods, such as stepwise, in selecting the right predictors due to random noise17,18. In this paper, we extended the application of bootstrap methods to assess the variability of the outcome of automated stepwise logistic regression to identify predictors of hospital readmission. This method is similar to the methods proposed and utilized in other studies17,19. Five hundred bootstrap samples with replacement were selected from the original data cohort and used as training www.cmrojournal.com ! 2015 Informa UK Ltd
Current Medical Research & Opinion
Volume 31, Number 1
January 2015
Table 1. Variables included in the predictive model. Variable
Total Dataset, N ¼ 52,070
Curr Med Res Opin Downloaded from informahealthcare.com by University of Otago on 07/12/15 For personal use only.
Not Readmitted n ¼ 46,869 (90.0%) Mean age, years (SD) Age, n (%) 18–35 36–50 51–64 65 Gender, n (%) Female Male Race, n (%) African American Caucasian Asian/Other/Unknown Region, n (%) Midwest Northeast South West Other/Unknown Mean Charlson Comorbidity Index (SD) Payer type, n (%) Commercial Dual eligible (Medicare/Medicaid) Medicaid Medicare Other Uninsured Diabetes diagnosis prior to index stay, n (%) Diagnosed Undiagnosed Diabetes treatment escalation Insulin-to-insulin None-to-insulin None-to-oral/NII Oral/NII-to-insulin Oral/NII-to-oral/NII None-to-none Hypoglycemia in the pre-period, n (%) Yes No Endocrinologist visit in the pre-period, n (%) Yes No Mean number of emergency room visits in the pre-period (SD) Mean number of inpatient visits in the pre-period (SD) Admitted from: Ambulatory care Emergency room Office or clinic None Other Discharged to: Home/self-care Home under care Mean length of index stay, days (SD)* HbA1c tested during hospitalization Yes No Mean number of HbA1c tests during hospitalization (SD) Random glucose test during hospitalization Yes No
P-value
Odds Ratio (95% CI)
50.0001 50.0001
1.01 (1.00–1.01)
Readmitted n ¼ 5201 (10.0%)
61.0 (12.3)
61.7 (12.4)
1853 (90.1) 7571 (90.4) 15,378 (91.0) 22,067 (89.2)
203 (9.9) 807 (9.6) 1523 (9.0) 2668 (10.8)
25,422 (90.3) 21,445 (89.7)
2738 (9.7) 2463 (10.3)
10,563 (89.3) 31,006 (90.2) 5300 (90.6)
1272 (10.7) 3381 (9.8) 548 (9.4)
18,926 (90.0) 4402 (90.4) 20,685 (90.0) 2773 (89.1) 83 (89.2) 1.0 (1.3)
2093 (10.0) 465 (9.6) 2295 (10.0) 338 (10.9) 10 (10.8) 1.4 (1.7)
12,450 (92.4) 1534 (86.2) 2029 (87.7) 19,555 (88.9) 10,184 (90.1) 1117 (93.0)
1023 (7.6) 245 (13.8) 284 (12.3) 2443 (11.1) 1122 (9.9) 84 (7.0)
25,765 (89.5) 21,104 (90.6)
3018 (10.5) 2183 (9.4)
3854 (86.7) 673 (67.7) 949 (82.0) 320 (79.4) 6924 (91.7) 34,149 (91.0)
590 (13.3) 321 (32.3) 208 (18.0) 83 (20.6) 628 (8.3) 3371 (9.0)
594 (87.1) 46,275 (90.1)
88 (12.9) 5113 (9.9)
1293 (2.8) 45,576 (97.2) 0.4 (1.1)
147 (2.8) 5054 (97.2) 0.7 (1.9)
0.91 (0.78–1.05) 0.88 (0.81–0.96) 0.82 (0.77–0.88) 0.0281
1.07 (1.01–1.13)
0.0041
1.10 (1.03–1.18)
0.4394
0.91 (0.80–1.02)
50.0001 50.0001
50.0001
0.0106
(OR vs. none-to-none:) 1.55 (1.41–1.70) 4.83 (4.21–5.54) 2.22 (1.90–2.59) 2.63 (2.06–3.35) 0.92 (0.84–1.00) NA 1.34 (1.07–1.68)
0.7778
1.03 (0.86–1.22)
50.0001
1.14 (1.12–1.16)
50.0001 50.0001
1.59 (1.51–1.68) (OR vs. other:) 0.94 (0.85–1.04) 1.21 (1.11–1.32) 0.96 (0.87–1.07) 1.05 (0.97–1.12) NA 0.71 (0.66–0.75)
50.0001
0.1 (0.3)
0.2 (0.6)
5838 (90.8) 6304 (88.5) 4665 (90.6) 12,998 (89.9) 17,064 (90.3)
592 (9.2) 822 (11.5) 484 (9.4) 1464 (10.1) 1839 (9.7)
37,122 (90.7) 9747 (87.4) 3.5 (2.5)
3792 (9.3) 1409 (12.6) 4.4 (2.9)
12,542 (90.2) 34,327 (89.9) 0.3 (0.5)
1365 (9.8) 3836 (10.1) 0.3 (0.5)
42,853 (89.8) 4016 (92.6)
4878 (10.2) 323 (7.4)
1.18 (1.16–1.20) (OR vs. uninsured:) 1.09 (0.87–1.38) 2.12 (1.64–2.75) 1.86 (1.44–2.40) 1.66 (1.33–2.08) 1.47 (1.16–1.84) NA 1.13 (1.07–1.20)
50.0001 50.0001
1.12 (1.11–1.13) 0.97 (0.91–1.04)
0.4260 0.5112 50.0001
0.98 (0.93–1.04) 1.41 (1.26–1.59)
(continued )
! 2015 Informa UK Ltd www.cmrojournal.com
Predictors of 30 day hospital readmission Eby et al.
109
Current Medical Research & Opinion Volume 31, Number 1
January 2015
Table 1. Continued. Variable
Total Dataset, N ¼ 52,070
Curr Med Res Opin Downloaded from informahealthcare.com by University of Otago on 07/12/15 For personal use only.
Not Readmitted n ¼ 46,869 (90.0%) Mean number of procedures/day during hospitalization (SD) Any encounter (excluding readmission) within 31 days post-discharge Yes No Surgery during index hospitalization Yes No Comorbidities, n (%)y Diabetic complications Yes No Hypoglycemia Yes No Any infection Yes No Sepsis Yes No Pneumonia Yes No Disease of lower extremity Yes No Nephropathy/renal complications Yes No End-stage renal disease Yes No Ischemic heart disease Yes No Cerebrovascular disease Yes No Depression Yes No Diabetic retinopathy Yes No Electrolyte imbalance Yes No COPD Yes No Hypertension Yes No Heart failure Yes No Other comorbidities Yes No
5.2 (7.5)
P-value
Odds Ratio (95% CI)
0.0002 50.0001
0.99 (0.99–1.00) 0.93 (0.92–0.94)
Readmitted n ¼ 5201 (10.0%) 4.8 (7.2)
38,313 (92.0) 8556 (81.9)
3311 (8.0) 1890 (18.1)
22,721 (90.4) 24,148 (89.6)
2413 (9.6) 2788 (10.4)
215 (95.1) 46,654 (90.0)
11 (4.9) 5190 (10.0)
338 (89.7) 46,531 (90.0)
39 (10.3) 5162 (10.0)
2485 (88.8) 44,384 (90.1)
315 (11.3) 4886 (9.9)
374 (85.4) 46,495 (90.1)
64 (14.6) 5137 (9.9)
1407 (88.7) 45,462 (90.1)
179 (11.3) 5022 (9.9)
977 (87.9) 45,892 (90.1)
135 (12.1) 5066 (9.9)
635 (88.7) 46,234 (90.0)
81 (11.3) 5120 (10.0)
1401 (82.6) 45,468 (90.3)
295 (17.4) 4906 (9.7)
4832 (90.2) 42,037 (90.0)
525 (9.8) 4676 (10.0)
2249 (92.3) 44,620 (89.9)
188 (7.7) 5013 (10.1)
855 (91.7) 46,014 (90.0)
77 (8.3) 5124 (10.0)
59 (89.4) 46,810 (90.0)
7 (10.6) 5194 (10.0)
2292 (88.8) 44,577 (90.1)
288 (11.2) 4913 (9.9)
1061 (89.3) 45,808 (90.0)
127 (10.7) 5074 (10.0)
10,835 (91.2) 36,034 (89.7)
1042 (8.8) 4159 (10.3)
2098 (84.7) 44,771 (90.3)
380 (15.3) 4821 (9.7)
11 (84.6) 46,858 (90.0)
2 (15.4) 5199 (10.0)
0.92 (0.87–0.97) 0.0043 0.0101
0.46 (0.25–0.84)
0.8169
1.04 (0.75–1.45)
0.0221
1.15 (1.02–1.30)
0.0012
1.55 (1.19–2.02)
0.08
1.15 (0.98–1.35)
0.0156
1.25 (1.04–1.50)
0.2340
1.15 (0.91–1.45)
50.0001
1.95 (1.72–2.22)
0.6276
0.98 (0.89–1.07)
0.0001
0.74 (0.64–0.87)
0.0761
0.81 (0.64–1.02)
0.8670
1.07 (0.49–2.34)
0.0413
1.14 (1.01–1.29)
0.4145
1.08 (0.90–1.30)
50.0001
0.83 (0.78–0.89)
50.0001
1.68 (1.50–1.88)
0.5164
1.64 (0.37–7.41)
CI ¼ confidence interval; COPD ¼ chronic obstructive pulmonary disease; NA ¼ not applicable; NII ¼ non-insulin injectable; SD ¼ standard deviation. *Mean length of index stay was truncated at 11 days for this analysis. yComorbidities assessed during index hospitalization.
110
Predictors of 30 day hospital readmission Eby et al.
www.cmrojournal.com ! 2015 Informa UK Ltd
Current Medical Research & Opinion
Volume 31, Number 1
January 2015
Table 2. Relevant variables not included in the predictive model. Total Dataset, N ¼ 52,070
Variable
Not Readmitted n ¼ 46,869 (90.0)%
Curr Med Res Opin Downloaded from informahealthcare.com by University of Otago on 07/12/15 For personal use only.
Mean number of days to first encounter post-discharge (SD) Type of first encounter post-discharge Ambulatory patient services Emergency patient Inpatient None Office or clinic patient Other patient type Mean HbA1c during hospitalization* (SD) Mean number of random glucose tests during hospitalization (SD)
7.6 (6.8) 5096 (91.1%) 1968 (84.4%) 0 (0.0%) 8556 (100.0%) 19,641 (93.7%) 11,608 (91.1%) 7.3% (2.2%) (56 mmol/mol [24 mmol/mol]) 12.2 (21.2)
P-value
Odds Ratio (95% CI)
50.0001 50.0001
0.93 (0.92–0.94)
0.1787
0.98 (0.96–1.01)
50.0001
1.01 (1.01–1.01)
Readmitted n ¼ 5201 (10.0)% 5.1 (4.9) 499 (8.9%) 364 (15.6%) 1890 (100.0%) 0 (0.0%) 1318 (6.3%) 1130 (8.9%) 7.2% (2.1%) (55 mmol/mol [23 mmol/mol]) 16.9 (26.5)
CI ¼ confidence interval; SD ¼ standard deviation. *Mean HbA1c values were available for 12,542 patients in the non-readmitted cohort and for 1365 patients in the readmitted cohort.
datasets. Each sample was selected to be the same size as the original data cohort (N ¼ 52,070). Since samples were obtained with replacement, there were observations that were selected repeatedly and there were observations that were not selected at all, as part of the training dataset. For each bootstrap sample, the observations not selected were included in a validation dataset. A logistic regression model was fit on each training dataset to identify significant predictors of 30 day hospital readmission. A stepwise variable selection procedure was used to select model variables, and 0.05 was the significance level used for variables to enter and stay in the model. Variables were included in the model only as main effects; higher level terms (e.g., quadratic) or interaction terms between variables were not under consideration in the stepwise algorithm. A receiveroperating characteristic (ROC) curve was constructed for the model. The ROC is a plot of sensitivity versus specificity, where sensitivity is the proportion of correctly identified readmitted patients and specificity is the proportion of correctly identified non-readmitted patients. Area under the curve (AUC) of the ROC plot measures the accuracy of the model in predicting readmission, where AUC of 1.0 corresponds to a perfect prediction and AUC of 0.5 corresponds to a random guess. The regression model was also applied to the validation dataset, and the AUC was calculated to assess the predictive power of the model when applied to data not used to develop the model.
Results Among 52,070 patients with type 2 diabetes and an initial hospitalization for any reason, 5201 (10.0%) patients were readmitted within 30 days and 46,869 (90.0%) patients showed no evidence of readmission. Variables included ! 2015 Informa UK Ltd www.cmrojournal.com
in the predictive model are presented in Table 1. Several patient characteristics, pre- and post-period resource utilization, and index hospitalization characteristics were significantly different between those who were and were not readmitted. Males; African Americans; patients with a diabetes diagnosis; and patients covered by Medicaid, Medicare, or both (dual-eligible) had higher readmission rates. Readmitted patients had a higher mean Charlson Comorbidity Index20 and a greater number of endocrinologist, emergency room, and inpatient visits in the pre-period. Patients who experienced escalation of their diabetic treatments (especially those escalating from having no records of anti-diabetic medications to treatment with insulin) and those who had at least one random blood glucose test during index hospitalization were readmitted more frequently, whereas those who had surgery during their index hospitalization and those who were discharged to home under self-care were readmitted less frequently. Additionally, the mean length of stay for the index hospitalization among readmitted patients was greater than that of non-readmitted patients. Table 2 presents relevant variables that could not be used in the modeling, due to the amount of missing data or because the variable was directly related to readmission. Patients whose first encounter was an emergency room visit were readmitted more frequently than those who visited an ambulatory clinic, outpatient office, or other setting. The number of days to the first encounter was lower for the readmitted cohort. All variables in Table 1 were included in the stepwise model. The number of times (out of 500 iterations) that each variable was included in the overall logistic regression model (based on a 0.05 significance level) is presented in Table 3. All variables entered into the model were significant (P50.05) in at least some of the iterations. Predictors of 30 day hospital readmission Eby et al.
111
Current Medical Research & Opinion Volume 31, Number 1
January 2015
Table 3. Odds ratio estimates and percentage (out of the 500 iterations) that each variable significantly predicts 30 day readmissions (475% cutoff).
Curr Med Res Opin Downloaded from informahealthcare.com by University of Otago on 07/12/15 For personal use only.
Variable
Surgery during index hospitalization HbA1c tested during hospitalization Source of admission (vs. other): Ambulatory care Emergency room Office or clinic None End-stage renal disease Heart failure Discharge disposition Any encounter within 30 days post-discharge (excluding readmission) Charlson Comorbidity Index Diabetes diagnosis prior to index stay Hypertension Payer type (vs. uninsured): Commercial Dual eligible (Medicare/Medicaid) Medicaid Medicare Other Length of index stay* Number of emergency room visits in the pre-period Number of inpatient visits in the pre-period Diabetic treatment escalation (vs. none–none): Insulin-to-insulin None-to-insulin None-to-oral Oral-to-insulin Oral-to-oral
Odds ratio %
Mean
Minimum
Maximum
75.4 98.4 99
0.88 0.85
0.76 0.76
0.94 0.93
0.95 1.21 1.18 0.97 1.38 1.33 0.84 0.31 1.14 0.68 0.80
0.78 0.97 0.98 0.81 1.16 1.15 0.75 0.29 1.10 0.61 0.7
1.15 1.43 1.41 1.12 1.70 1.60 0.92 0.35 1.20 0.76 0.88
1.3.0 1.97 1.69 1.72 1.50 1.11 1.10 1.22
0.94 1.31 1.14 1.17 0.98 1.09 1.06 1.10
1.77 2.89 2.37 2.43 2.04 1.13 1.14 1.38
1.59 6.56 3.33 3.56 1.20
1.32 5.15 2.54 2.47 1.03
1.94 8.17 4.32 4.96 1.37
99.0 99.2 99.2 100.0 100.0 100.0 100.0 100.0
100.0 100.0 100.0 100.0
*Mean length of index stay was truncated at 11 days for this analysis.
There were 14 variables that were included in over 98% of all 500 iterations, which indicates strength in predicting readmissions. These variables include those that significantly increased the risk of readmission, such as treatment escalation from no diabetes medications in the pre-period to insulin in the post-period, as well as variables that were protective from readmission, such as having an outpatient encounter following index hospitalization discharge. The resulting AUC statistics for the 500 model iterations in the training and validation datasets are summarized in Table 3. The predictive accuracy of the final logistic regression model, for each iteration, had a mean AUC of 0.699, ranging from 0.687 to 0.710. Similar accuracy in prediction was seen when the model was applied to the validation data, where the mean AUC was 0.693 (range 0.679–0.709).
Discussion This analysis identified factors predictive of 30 day readmission in patients with type 2 diabetes. In our analysis, escalation of diabetic treatments (especially those escalated from having no records of anti-diabetic medications 112
Predictors of 30 day hospital readmission Eby et al.
to treatment with insulin) was the strongest predictor of 30 day readmission with a mean OR estimate of 6.557 (vs. no treatment). Since treatment escalation was such a strong predictive variable, this study underlines the importance of early and appropriate interventions for patients with diabetes. Although treatment escalation could have been related to other variables (e.g., time to first encounter, hypoglycemia episodes, adverse effects of treatment), these findings reflect the importance of close follow-up of patients who are undergoing treatment intensification. Additional factors associated with greater likelihood of 30 day readmission include longer index hospital stays and more pre-period emergency room and inpatient visits. Because being diagnosed with diabetes prior to the index stay was protective of 30 day readmission (mean OR: 0.675), this study supports the importance of accurate diagnosis and appropriate follow-up for patients who may be at risk for diabetes (e.g., patients with one or more glucose tests that may indicate diabetes and for whom followup diagnostics would be encouraged). The association of a higher readmission risk in those patients without an established diagnosis might reflect inaccurate diagnosis or could reflect a more rapid course with less access to care and www.cmrojournal.com ! 2015 Informa UK Ltd
Curr Med Res Opin Downloaded from informahealthcare.com by University of Otago on 07/12/15 For personal use only.
Current Medical Research & Opinion
fewer opportunities for evaluation by endocrinologists or other clinicians. This variable (absence of an established diagnosis) might also be a surrogate marker for lower quality of care. Accurate diagnosis and appropriate follow-up are required for both patients with diabetes and for those who might be at risk for diabetes (i.e., those who meet some diagnostic criteria but are not yet diagnosed). From the ROC analysis (mean AUC of 0.693), the predictive accuracy of the final logistic regression model is considered modest. This result might be due to the unavailability of some variables or data that influence readmission from the EHR database and, thus, from inclusion into the predictive model, or because readmission itself is not amenable to more precise prediction from any other observed or unobserved variables. Many of the variables identified as predictive of 30 day readmission in this analysis are consistent with those found in previous studies of short-term hospital readmission among patients with diabetes. Patients with diabetes and greater comorbidity burden9, particularly those with congestive heart failure7,9,21, chronic obstructive pulmonary disease7, or end-stage organ disease9, have been shown to experience more frequent hospital readmission; this result is consistent with our findings. The absence of a diabetes diagnosis prior to the index stay has also been previously shown to be predictive of 30 day readmission8, which is an intriguing finding also corroborated by our analysis. Interestingly, our descriptive results showed no significant differences in HbA1c values for the readmitted versus the non-readmitted patients. Race and insurance status have been identified as important contributors to readmission risk among patients with diabetes, and these findings are also consistent with our analysis: non-Caucasian patients with diabetes, particularly African Americans, are more likely to be readmitted than are Caucasian patients2,13,22, and patients with public insurance (e.g., Medicare and/or Medicaid) are more likely to be readmitted than those with private insurance2,8,9,22. The variables related to race and insurance status may reflect, at least in part, socioeconomic factors (e.g., lower income, reduced access to healthcare coverage and resources) associated with these populations, as opposed to characteristics of these populations’ diabetes severity or treatments. Other studies have measured the effectiveness of various interventions at reducing hospital readmission risk in patients with diabetes. Although different diabetic treatment regimens did not affect post-discharge outcomes in one study23, another study found that team interventions (i.e., a nurse educator and an endocrinologist) significantly reduced the rate of recurrent hospitalizations24. In patients with diabetes who were hospitalized for myocardial infarction, the type of glucose-lowering treatment received as monotherapy was not associated with differences in prognosis or readmission risk25. ! 2015 Informa UK Ltd www.cmrojournal.com
Volume 31, Number 1
January 2015
Limitations of this study include that hospitalizations and other encounters outside the Humedica network were not captured in this database. Some variables related to diabetes (e.g., duration of diabetes) were not available and were not included in the analysis. HbA1c values were not available for all patients and, thus, were only reported descriptively for those who did have values; therefore, HbA1c values were not included in the predictive model. Additionally, data on medication compliance were not available, and severity of disease and reasons for hospitalization were not provided for every patient. Furthermore, as in any retrospective study using EHR data, these results rely on the accuracy of reporting and coding within the database (e.g., comorbidities during the index hospitalization may be underreported, as these were identified by ICD-9 codes in the EHR data). An additional potential limitation is that the statistical methods used stepwise logistic regression and were only limited to main effects of the variable, rather than interactions. Strengths of the current analysis include the consistency with previous studies on this topic, the inclusion of both medical and surgical patients in the study population, and the application of the novel variable of ‘diabetic treatment escalation’. To our knowledge, this is the first such analysis to consider the effect of treatment escalation on short-term hospital readmission among patients with type 2 diabetes.
Conclusion The identification of factors predictive of 30 day hospital readmission among patients with type 2 diabetes underlines the importance of appropriate recognition of and treatment for diabetes prior to and during hospitalization. Approaches to managing diabetes that take these factors and patient characteristics into account may lead to reduced hospital readmission rates and, thus, to reduced cost burdens to patients, payers, and healthcare providers.
Transparency Declaration of funding This study was funded by Eli Lilly and Company. All authors were involved in study design and data analysis; all authors developed, reviewed, and edited the manuscript; and all authors contributed to the discussion. Declaration of financial/other relationships C.H., M.Y., S.G., K.D., and J.X. have disclosed that they are employees of Eli Lilly and Company. E.E. has disclosed that she was an employee of Eli Lilly and Company at the time this study was completed. T.G. has disclosed that he has no significant relationships with or financial interests in any commercial companies related to this study or article.
Predictors of 30 day hospital readmission Eby et al.
113
Current Medical Research & Opinion Volume 31, Number 1
January 2015
CMRO peer reviewers on this manuscript have no relevant financial or other relationships to disclose.
Curr Med Res Opin Downloaded from informahealthcare.com by University of Otago on 07/12/15 For personal use only.
Acknowledgments The authors wish to thank Karla C. Villines of Eli Lilly and Company for critical review of the manuscript, and Jeffrey Walter and Maria Rovere of _ospitali Health Clinical for assistance writing and editing the manuscript, respectively. Previous presentation: Portions of this analysis were presented at the 73rd Scientific Sessions of the American Diabetes Association, 23 June 2013, Chicago, IL, USA.
References 1. American Diabetes Association. Economic costs of diabetes in the US in 2012. Diabetes Care 2013;36:1033-46 2. Jiang HJ, Stryer D, Friedman B, et al. Multiple hospitalizations for patients with diabetes. Diabetes Care 2003;26:1421-6 3. Fraze T, Jiang J, Burgess J. Hospital Stays for Patients with Diabetes, 2008. Healthcare Cost and Utilization Project (HCUP) Statistical Brief #93, Rockville, MD: Agency for Healthcare Research and Quality, 2010 4. Allaudeen N, Schnipper JL, Orav EJ, et al. Inability of providers to predict unplanned readmissions. J Gen Intern Med 2011;26:771-6 5. Stone J, Hoffman GJ. Medicare Hospital Readmissions: Issues, Policy Options and PPACA. Congressional Research Service, 2010. Available at: http:// www.ncsl.org/documents/health/Medicare_Hospital_Readmissions_and_ PPACA.pdf [Last accessed 26 February 2014] 6. Kansagara D, Englander H, Salanitro A, et al. Risk prediction models for hospital readmission a systematic review. JAMA 2011;306:1688-98 7. Smith DM, Giobbie-Hurder A, Weinberger M, et al. Predicting nonelective hospital readmissions: a multi-site study. J Clin Epidemiol 2000;53:1113-18 8. Robbins JM, Webb DA. Diagnosing diabetes and preventing rehospitalizations: the urban diabetes study. Med Care 2006;44:292-6 9. Kim H, Ross JS, Melkus GD, et al. Scheduled and unscheduled hospital readmissions among patients with diabetes. Am J Manag Care 2010;16:760-7 10. Kanel K, Elster S, Vrbin C. PHRI Readmission Briefs. Brief I: Overview of Six Target Chronic Diseases. Pittsburgh, PA: Pittsburgh Regional Health Initiative, 2010
114
Predictors of 30 day hospital readmission Eby et al.
11. McHugh MD, Carthon JM, Kang XL. Medicare readmissions policies and racial and ethnic health disparities: a cautionary tale. Policy Polit Nurs Pract 2010;11:309-16 12. Hernandez AF, Greiner MA, Fonarow GC, et al. Relationship between early physician follow-up and 30-day readmission among Medicare beneficiaries hospitalized for heart failure. JAMA 2010;303:1716-22 13. Jiang HJ, Andrews R, Stryer D, et al. Racial/ethnic disparities in potentially preventable readmissions: the case of diabetes. Am J Public Health 2005;95:1561-7 14. van Walraven C, Dhalla IA, Bell C, et al. Derivation and validation of an index to predict early death or unplanned readmission after discharge from hospital to the community. CMAJ 2010;182:551-7 15. Ginde AA, Blanc PG, Lieberman RM, et al. Validation of ICD-9-CM coding algorithm for improved identification of hypoglycemia visits. BMC Endocr Disord 2008;8:4 16. Chernick MR. Bootstrap Methods: A Guide for Practitioners and Researchers, 2nd Edn. Hoboken, NJ: John Wiley & Sons Inc., 2008 17. Austin PC, Tu JV. Automated variable selection methods for logistic regression produced unstable models for predicting acute myocardial infarction mortality. J Clin Epidemiol 2004;57:1138-46 18. Derksen S, Keselman HJ. Backward, forward, and stepwise automated subset selection algorithms: frequency of obtaining authentic and noise variables. Br J Math Stat Psychol 1992;45:265-82 19. Efron B, Gong G. A leisurely look at the bootstrap, the jackknife, and crossvalidation. Am Stat 1983;37:36-48 20. Deyo R, Cherkin DC, Ciol MA. Adapting a clinical comorbidity index for use with ICD-9-CM administrative databases. J Clin Epidemiol 1992;45:613-19 21. Cramer S, Chapa G, Kotsos T, et al. Assessing multiple hospitalizations for health-plan-managed Medicaid diabetic members. J Healthc Qual 2010;32:7-14 22. Estrada CL, Danielson KK, Drum ML, et al. Hospitalization subsequent to diagnosis in young patients with diabetes in Chicago, Illinois. Pediatrics 2009;124:926-34 23. Sarma S, Mentz RJ, Kwasny MJ, et al. Association between diabetes mellitus and post-discharge outcomes in patients hospitalized with heart failure: findings from the EVEREST trial. Eur J Heart Fail 2013;15:194-202 24. Koproski J, Pretto Z, Poretsky L. Effects of an intervention by a diabetes team in hospitalized patients with diabetes. Diabetes Care 1997;20:1553-5 25. Horsdal HT, Johnsen SP, Søndergaard F, et al. Type of preadmission glucose-lowering treatment and prognosis among patients hospitalized with myocardial infarction: a nationwide follow-up study. Diabetologia 2008;51:567-74
www.cmrojournal.com ! 2015 Informa UK Ltd
Vol. 31, No. 1, 2015, 107–114
Article ST-0245.R1/981632 All rights reserved: reproduction in whole or part not permitted
Curr Med Res Opin Downloaded from informahealthcare.com by University of Otago on 07/12/15 For personal use only.
Original article Predictors of 30 day hospital readmission in patients with type 2 diabetes: a retrospective, case–control, database study
Elizabeth Eby Former employee of Eli Lilly and Company, Indianapolis, IN, USA
Christine Hardwick Maria Yu Steve Gelwicks Ketra Deschamps Jin Xie Eli Lilly and Company, Indianapolis, IN, USA
Tom George Detroit Medical Center, Sinai-Grace Hospital, Detroit, MI, USA Address for correspondence: Christine Hardwick RN MPH, Eli Lilly Corporate Center, Indianapolis, IN 46285, USA. Tel: +1 (317) 433-6404; Fax: +1 (317) 277-3585; [email protected] Keywords: Diabetes mellitus – Hospitalization – Hospital readmission – Regression analysis – Type 2 diabetes mellitus Accepted: 22 October 2014; published online: 6 November 2014 Citation: Curr Med Res Opin 2015; 31:107–14
Abstract Objective: To assess factors predictive of all-cause, 30 day hospital readmission among patients with type 2 diabetes in the United States. Methods: A retrospective, case–control study using deidentified Humedica electronic health record data was conducted to identify patients 18 years old with 6 months of data prior to index hospitalization (preperiod) and 30 days of data after discharge (post-period). Combined methods of bootstrap resampling and stepwise logistic regression were used to identify factors associated with readmission. Results: Among 52,070 patients with type 2 diabetes and an initial hospitalization for any reason, 5201 (10.0%) patients were readmitted within 30 days and 46,869 (90.0%) patients showed no evidence of readmission. Diabetic treatment escalation; race; type 2 diabetes diagnosis prior to the index stay; pre-period heart failure; and number of pre-period, inpatient healthcare visits were among the strongest predictors of 30 day readmission. From a receiver-operating characteristic plot (mean area under curve of 0.693), the predictive accuracy of the final logistic regression model is considered modest. This result might be due to the unavailability of some variables or data. Conclusions: These results highlight the importance of the appropriate recognition of and treatment for type 2 diabetes, prior to and during hospitalization and following discharge, in order to impact a subsequent hospitalization. In our analysis, escalation of diabetic treatments (especially those escalated from having no records of antidiabetic medications to treatment with insulin) was the strongest predictor of 30 day readmission. Limitations of this study include the fact that hospitalizations and other encounters, outside the Humedica network, were not captured in this analysis.
Introduction Diabetes is associated with a substantial economic burden; total diabetes-related costs were estimated at $245 billion in the US in 2012, including $176 billion in direct medical costs and $69 billion in reduced productivity and other indirect costs1. The largest component of medical costs for diabetes is hospital inpatient care, which accounts for 43% of the total medical cost1. Patients with diabetes are more likely to be hospitalized than patients without diabetes – in 2012, more than 25% of US hospital inpatient stays were incurred by patients with diabetes, and more than 15% of those inpatients stays were ! 2015 Informa UK Ltd www.cmrojournal.com
Predictors of 30 day hospital readmission Eby et al.
107
Curr Med Res Opin Downloaded from informahealthcare.com by University of Otago on 07/12/15 For personal use only.
Current Medical Research & Opinion Volume 31, Number 1
January 2015
directly attributed to diabetes1. Patients with diabetes who are hospitalized incur nearly two-fold higher inpatient costs per capita2. Hospital stays for patients with diabetes are also more likely to follow an emergency room visit, compared to hospitalizations for patients without diabetes3. Additionally, multiple hospitalizations are common among patients with diabetes; these patients with multiple hospitalizations also experience longer hospital stays and considerably higher hospital costs2. The increased risk of hospitalization among patients with diabetes is due, in part, to increased comorbidities among these patients, in addition to health problems associated with diabetes and its complications2,3. Despite this increased hospitalization risk, healthcare providers are often unable to accurately identify patients at risk for readmission4. As part of the Patient Protections and Affordable Care Act (PPACA) passed in 2010, Medicare may reduce Medicare reimbursement payments to hospitals (by reducing Medicare reimbursement payments) with high preventable readmission rates5. These ‘preventable readmissions’ may be difficult to measure, but all-cause hospital readmissions within 30 days of discharge can be evaluated by using electronic health record (EHR) data6. The objective of this study was to assess (by using a large EHR database) the factors predictive of all-cause, 30 day hospital readmission among patients with type 2 diabetes in the US.
Patients and methods Study population The data for this retrospective study were obtained from Humedica’s deidentified EHR database. Humedica partners directly with large medical group practices, integrated delivery networks (IDNs), and hospitals to extract data from their EHRs and various health information technology systems. A subset of the data components include laboratory results, radiology and pathology reports, prescriptions written and dispensed, procedures, diagnoses, and other details of a patient’s office visit and hospital stay. At the time of this study, Humedica’s clinical data warehouse encompassed approximately 13 million patients. The data for this study spanned the period from 1 October 2009 through 30 September 2011, and the data are deidentified to comply with Health Insurance Portability and Accountability Act (HIPAA) regulations. Patients included in the analysis were aged 18 or older and had a diagnosis of type 2 diabetes (i.e., International Classification of Diseases, Ninth Revision [ICD-9] 250.x0 or 250.x2) or met criteria for undiagnosed type 2 diabetes (used medication indicative of type 2 diabetes and/or had laboratory values indicative of diabetes [HbA1c values of 108
Predictors of 30 day hospital readmission Eby et al.
6.5% {48 mmol/mol} or greater, fasting glucose 7.0 mmol/ L, or oral glucose tolerance test 11.1 mmol/L] in the absence of a diagnosis code for diabetes). Patients were also required to have records associated with an IDN; an index hospitalization not within 30 days of a previous inpatient discharge within the study period of 1 September 2009 – 31 August 2011; 6 months of data prior to the index hospitalization (i.e., the pre-period); 31 days of data postdischarge (to create the 30 day post-period); at least one valid blood glucose measurement in the pre-period or during the index hospitalization; and a discharge to their home. Patients were excluded if they had no valid pre- or post-period records; died during the index hospitalization or within 30 days of discharge; had a diagnosis of cancer (except skin cancer); had a diagnosis of type 1, secondary, or gestational diabetes; or were pregnant.
Statistical analyses Baseline variables of the study population were compared between the readmitted and non-readmitted cohorts and included pre-index demographics and disease characteristics (Table 1). These variables were selected based on a review of the literature2,7–14 and of the available variables in the Humedica dataset. We derived the diabetic treatment escalation variable by comparing the prescriptions written during the pre-period with those written during the post-period. Comorbidities present during the index hospitalization were identified by ICD-9 codes. Hypoglycemia episodes were calculated using the methods described by Ginde et al.15. Univariate test statistics (t-test for continuous variables and chi-squared test for categorical variables) were used, and univariate logistic regression was used to obtain odds ratio (OR) estimates of being readmitted. For the multivariate analyses, bootstrap sampling of the data was utilized. Bootstrap resampling is a common statistical method often used in estimating variability or standard error of estimates. In this study, it was used to generate a model training sample and a model validation sample. Bootstrapping is performed by generating a sample with replacement from the original sample. Because observations that are drawn are made available again for resampling, some can be selected repeatedly as part of the bootstrap sample and some may not be selected at all16. Other studies have shown the instability of automated variable selection methods, such as stepwise, in selecting the right predictors due to random noise17,18. In this paper, we extended the application of bootstrap methods to assess the variability of the outcome of automated stepwise logistic regression to identify predictors of hospital readmission. This method is similar to the methods proposed and utilized in other studies17,19. Five hundred bootstrap samples with replacement were selected from the original data cohort and used as training www.cmrojournal.com ! 2015 Informa UK Ltd
Current Medical Research & Opinion
Volume 31, Number 1
January 2015
Table 1. Variables included in the predictive model. Variable
Total Dataset, N ¼ 52,070
Curr Med Res Opin Downloaded from informahealthcare.com by University of Otago on 07/12/15 For personal use only.
Not Readmitted n ¼ 46,869 (90.0%) Mean age, years (SD) Age, n (%) 18–35 36–50 51–64 65 Gender, n (%) Female Male Race, n (%) African American Caucasian Asian/Other/Unknown Region, n (%) Midwest Northeast South West Other/Unknown Mean Charlson Comorbidity Index (SD) Payer type, n (%) Commercial Dual eligible (Medicare/Medicaid) Medicaid Medicare Other Uninsured Diabetes diagnosis prior to index stay, n (%) Diagnosed Undiagnosed Diabetes treatment escalation Insulin-to-insulin None-to-insulin None-to-oral/NII Oral/NII-to-insulin Oral/NII-to-oral/NII None-to-none Hypoglycemia in the pre-period, n (%) Yes No Endocrinologist visit in the pre-period, n (%) Yes No Mean number of emergency room visits in the pre-period (SD) Mean number of inpatient visits in the pre-period (SD) Admitted from: Ambulatory care Emergency room Office or clinic None Other Discharged to: Home/self-care Home under care Mean length of index stay, days (SD)* HbA1c tested during hospitalization Yes No Mean number of HbA1c tests during hospitalization (SD) Random glucose test during hospitalization Yes No
P-value
Odds Ratio (95% CI)
50.0001 50.0001
1.01 (1.00–1.01)
Readmitted n ¼ 5201 (10.0%)
61.0 (12.3)
61.7 (12.4)
1853 (90.1) 7571 (90.4) 15,378 (91.0) 22,067 (89.2)
203 (9.9) 807 (9.6) 1523 (9.0) 2668 (10.8)
25,422 (90.3) 21,445 (89.7)
2738 (9.7) 2463 (10.3)
10,563 (89.3) 31,006 (90.2) 5300 (90.6)
1272 (10.7) 3381 (9.8) 548 (9.4)
18,926 (90.0) 4402 (90.4) 20,685 (90.0) 2773 (89.1) 83 (89.2) 1.0 (1.3)
2093 (10.0) 465 (9.6) 2295 (10.0) 338 (10.9) 10 (10.8) 1.4 (1.7)
12,450 (92.4) 1534 (86.2) 2029 (87.7) 19,555 (88.9) 10,184 (90.1) 1117 (93.0)
1023 (7.6) 245 (13.8) 284 (12.3) 2443 (11.1) 1122 (9.9) 84 (7.0)
25,765 (89.5) 21,104 (90.6)
3018 (10.5) 2183 (9.4)
3854 (86.7) 673 (67.7) 949 (82.0) 320 (79.4) 6924 (91.7) 34,149 (91.0)
590 (13.3) 321 (32.3) 208 (18.0) 83 (20.6) 628 (8.3) 3371 (9.0)
594 (87.1) 46,275 (90.1)
88 (12.9) 5113 (9.9)
1293 (2.8) 45,576 (97.2) 0.4 (1.1)
147 (2.8) 5054 (97.2) 0.7 (1.9)
0.91 (0.78–1.05) 0.88 (0.81–0.96) 0.82 (0.77–0.88) 0.0281
1.07 (1.01–1.13)
0.0041
1.10 (1.03–1.18)
0.4394
0.91 (0.80–1.02)
50.0001 50.0001
50.0001
0.0106
(OR vs. none-to-none:) 1.55 (1.41–1.70) 4.83 (4.21–5.54) 2.22 (1.90–2.59) 2.63 (2.06–3.35) 0.92 (0.84–1.00) NA 1.34 (1.07–1.68)
0.7778
1.03 (0.86–1.22)
50.0001
1.14 (1.12–1.16)
50.0001 50.0001
1.59 (1.51–1.68) (OR vs. other:) 0.94 (0.85–1.04) 1.21 (1.11–1.32) 0.96 (0.87–1.07) 1.05 (0.97–1.12) NA 0.71 (0.66–0.75)
50.0001
0.1 (0.3)
0.2 (0.6)
5838 (90.8) 6304 (88.5) 4665 (90.6) 12,998 (89.9) 17,064 (90.3)
592 (9.2) 822 (11.5) 484 (9.4) 1464 (10.1) 1839 (9.7)
37,122 (90.7) 9747 (87.4) 3.5 (2.5)
3792 (9.3) 1409 (12.6) 4.4 (2.9)
12,542 (90.2) 34,327 (89.9) 0.3 (0.5)
1365 (9.8) 3836 (10.1) 0.3 (0.5)
42,853 (89.8) 4016 (92.6)
4878 (10.2) 323 (7.4)
1.18 (1.16–1.20) (OR vs. uninsured:) 1.09 (0.87–1.38) 2.12 (1.64–2.75) 1.86 (1.44–2.40) 1.66 (1.33–2.08) 1.47 (1.16–1.84) NA 1.13 (1.07–1.20)
50.0001 50.0001
1.12 (1.11–1.13) 0.97 (0.91–1.04)
0.4260 0.5112 50.0001
0.98 (0.93–1.04) 1.41 (1.26–1.59)
(continued )
! 2015 Informa UK Ltd www.cmrojournal.com
Predictors of 30 day hospital readmission Eby et al.
109
Current Medical Research & Opinion Volume 31, Number 1
January 2015
Table 1. Continued. Variable
Total Dataset, N ¼ 52,070
Curr Med Res Opin Downloaded from informahealthcare.com by University of Otago on 07/12/15 For personal use only.
Not Readmitted n ¼ 46,869 (90.0%) Mean number of procedures/day during hospitalization (SD) Any encounter (excluding readmission) within 31 days post-discharge Yes No Surgery during index hospitalization Yes No Comorbidities, n (%)y Diabetic complications Yes No Hypoglycemia Yes No Any infection Yes No Sepsis Yes No Pneumonia Yes No Disease of lower extremity Yes No Nephropathy/renal complications Yes No End-stage renal disease Yes No Ischemic heart disease Yes No Cerebrovascular disease Yes No Depression Yes No Diabetic retinopathy Yes No Electrolyte imbalance Yes No COPD Yes No Hypertension Yes No Heart failure Yes No Other comorbidities Yes No
5.2 (7.5)
P-value
Odds Ratio (95% CI)
0.0002 50.0001
0.99 (0.99–1.00) 0.93 (0.92–0.94)
Readmitted n ¼ 5201 (10.0%) 4.8 (7.2)
38,313 (92.0) 8556 (81.9)
3311 (8.0) 1890 (18.1)
22,721 (90.4) 24,148 (89.6)
2413 (9.6) 2788 (10.4)
215 (95.1) 46,654 (90.0)
11 (4.9) 5190 (10.0)
338 (89.7) 46,531 (90.0)
39 (10.3) 5162 (10.0)
2485 (88.8) 44,384 (90.1)
315 (11.3) 4886 (9.9)
374 (85.4) 46,495 (90.1)
64 (14.6) 5137 (9.9)
1407 (88.7) 45,462 (90.1)
179 (11.3) 5022 (9.9)
977 (87.9) 45,892 (90.1)
135 (12.1) 5066 (9.9)
635 (88.7) 46,234 (90.0)
81 (11.3) 5120 (10.0)
1401 (82.6) 45,468 (90.3)
295 (17.4) 4906 (9.7)
4832 (90.2) 42,037 (90.0)
525 (9.8) 4676 (10.0)
2249 (92.3) 44,620 (89.9)
188 (7.7) 5013 (10.1)
855 (91.7) 46,014 (90.0)
77 (8.3) 5124 (10.0)
59 (89.4) 46,810 (90.0)
7 (10.6) 5194 (10.0)
2292 (88.8) 44,577 (90.1)
288 (11.2) 4913 (9.9)
1061 (89.3) 45,808 (90.0)
127 (10.7) 5074 (10.0)
10,835 (91.2) 36,034 (89.7)
1042 (8.8) 4159 (10.3)
2098 (84.7) 44,771 (90.3)
380 (15.3) 4821 (9.7)
11 (84.6) 46,858 (90.0)
2 (15.4) 5199 (10.0)
0.92 (0.87–0.97) 0.0043 0.0101
0.46 (0.25–0.84)
0.8169
1.04 (0.75–1.45)
0.0221
1.15 (1.02–1.30)
0.0012
1.55 (1.19–2.02)
0.08
1.15 (0.98–1.35)
0.0156
1.25 (1.04–1.50)
0.2340
1.15 (0.91–1.45)
50.0001
1.95 (1.72–2.22)
0.6276
0.98 (0.89–1.07)
0.0001
0.74 (0.64–0.87)
0.0761
0.81 (0.64–1.02)
0.8670
1.07 (0.49–2.34)
0.0413
1.14 (1.01–1.29)
0.4145
1.08 (0.90–1.30)
50.0001
0.83 (0.78–0.89)
50.0001
1.68 (1.50–1.88)
0.5164
1.64 (0.37–7.41)
CI ¼ confidence interval; COPD ¼ chronic obstructive pulmonary disease; NA ¼ not applicable; NII ¼ non-insulin injectable; SD ¼ standard deviation. *Mean length of index stay was truncated at 11 days for this analysis. yComorbidities assessed during index hospitalization.
110
Predictors of 30 day hospital readmission Eby et al.
www.cmrojournal.com ! 2015 Informa UK Ltd
Current Medical Research & Opinion
Volume 31, Number 1
January 2015
Table 2. Relevant variables not included in the predictive model. Total Dataset, N ¼ 52,070
Variable
Not Readmitted n ¼ 46,869 (90.0)%
Curr Med Res Opin Downloaded from informahealthcare.com by University of Otago on 07/12/15 For personal use only.
Mean number of days to first encounter post-discharge (SD) Type of first encounter post-discharge Ambulatory patient services Emergency patient Inpatient None Office or clinic patient Other patient type Mean HbA1c during hospitalization* (SD) Mean number of random glucose tests during hospitalization (SD)
7.6 (6.8) 5096 (91.1%) 1968 (84.4%) 0 (0.0%) 8556 (100.0%) 19,641 (93.7%) 11,608 (91.1%) 7.3% (2.2%) (56 mmol/mol [24 mmol/mol]) 12.2 (21.2)
P-value
Odds Ratio (95% CI)
50.0001 50.0001
0.93 (0.92–0.94)
0.1787
0.98 (0.96–1.01)
50.0001
1.01 (1.01–1.01)
Readmitted n ¼ 5201 (10.0)% 5.1 (4.9) 499 (8.9%) 364 (15.6%) 1890 (100.0%) 0 (0.0%) 1318 (6.3%) 1130 (8.9%) 7.2% (2.1%) (55 mmol/mol [23 mmol/mol]) 16.9 (26.5)
CI ¼ confidence interval; SD ¼ standard deviation. *Mean HbA1c values were available for 12,542 patients in the non-readmitted cohort and for 1365 patients in the readmitted cohort.
datasets. Each sample was selected to be the same size as the original data cohort (N ¼ 52,070). Since samples were obtained with replacement, there were observations that were selected repeatedly and there were observations that were not selected at all, as part of the training dataset. For each bootstrap sample, the observations not selected were included in a validation dataset. A logistic regression model was fit on each training dataset to identify significant predictors of 30 day hospital readmission. A stepwise variable selection procedure was used to select model variables, and 0.05 was the significance level used for variables to enter and stay in the model. Variables were included in the model only as main effects; higher level terms (e.g., quadratic) or interaction terms between variables were not under consideration in the stepwise algorithm. A receiveroperating characteristic (ROC) curve was constructed for the model. The ROC is a plot of sensitivity versus specificity, where sensitivity is the proportion of correctly identified readmitted patients and specificity is the proportion of correctly identified non-readmitted patients. Area under the curve (AUC) of the ROC plot measures the accuracy of the model in predicting readmission, where AUC of 1.0 corresponds to a perfect prediction and AUC of 0.5 corresponds to a random guess. The regression model was also applied to the validation dataset, and the AUC was calculated to assess the predictive power of the model when applied to data not used to develop the model.
Results Among 52,070 patients with type 2 diabetes and an initial hospitalization for any reason, 5201 (10.0%) patients were readmitted within 30 days and 46,869 (90.0%) patients showed no evidence of readmission. Variables included ! 2015 Informa UK Ltd www.cmrojournal.com
in the predictive model are presented in Table 1. Several patient characteristics, pre- and post-period resource utilization, and index hospitalization characteristics were significantly different between those who were and were not readmitted. Males; African Americans; patients with a diabetes diagnosis; and patients covered by Medicaid, Medicare, or both (dual-eligible) had higher readmission rates. Readmitted patients had a higher mean Charlson Comorbidity Index20 and a greater number of endocrinologist, emergency room, and inpatient visits in the pre-period. Patients who experienced escalation of their diabetic treatments (especially those escalating from having no records of anti-diabetic medications to treatment with insulin) and those who had at least one random blood glucose test during index hospitalization were readmitted more frequently, whereas those who had surgery during their index hospitalization and those who were discharged to home under self-care were readmitted less frequently. Additionally, the mean length of stay for the index hospitalization among readmitted patients was greater than that of non-readmitted patients. Table 2 presents relevant variables that could not be used in the modeling, due to the amount of missing data or because the variable was directly related to readmission. Patients whose first encounter was an emergency room visit were readmitted more frequently than those who visited an ambulatory clinic, outpatient office, or other setting. The number of days to the first encounter was lower for the readmitted cohort. All variables in Table 1 were included in the stepwise model. The number of times (out of 500 iterations) that each variable was included in the overall logistic regression model (based on a 0.05 significance level) is presented in Table 3. All variables entered into the model were significant (P50.05) in at least some of the iterations. Predictors of 30 day hospital readmission Eby et al.
111
Current Medical Research & Opinion Volume 31, Number 1
January 2015
Table 3. Odds ratio estimates and percentage (out of the 500 iterations) that each variable significantly predicts 30 day readmissions (475% cutoff).
Curr Med Res Opin Downloaded from informahealthcare.com by University of Otago on 07/12/15 For personal use only.
Variable
Surgery during index hospitalization HbA1c tested during hospitalization Source of admission (vs. other): Ambulatory care Emergency room Office or clinic None End-stage renal disease Heart failure Discharge disposition Any encounter within 30 days post-discharge (excluding readmission) Charlson Comorbidity Index Diabetes diagnosis prior to index stay Hypertension Payer type (vs. uninsured): Commercial Dual eligible (Medicare/Medicaid) Medicaid Medicare Other Length of index stay* Number of emergency room visits in the pre-period Number of inpatient visits in the pre-period Diabetic treatment escalation (vs. none–none): Insulin-to-insulin None-to-insulin None-to-oral Oral-to-insulin Oral-to-oral
Odds ratio %
Mean
Minimum
Maximum
75.4 98.4 99
0.88 0.85
0.76 0.76
0.94 0.93
0.95 1.21 1.18 0.97 1.38 1.33 0.84 0.31 1.14 0.68 0.80
0.78 0.97 0.98 0.81 1.16 1.15 0.75 0.29 1.10 0.61 0.7
1.15 1.43 1.41 1.12 1.70 1.60 0.92 0.35 1.20 0.76 0.88
1.3.0 1.97 1.69 1.72 1.50 1.11 1.10 1.22
0.94 1.31 1.14 1.17 0.98 1.09 1.06 1.10
1.77 2.89 2.37 2.43 2.04 1.13 1.14 1.38
1.59 6.56 3.33 3.56 1.20
1.32 5.15 2.54 2.47 1.03
1.94 8.17 4.32 4.96 1.37
99.0 99.2 99.2 100.0 100.0 100.0 100.0 100.0
100.0 100.0 100.0 100.0
*Mean length of index stay was truncated at 11 days for this analysis.
There were 14 variables that were included in over 98% of all 500 iterations, which indicates strength in predicting readmissions. These variables include those that significantly increased the risk of readmission, such as treatment escalation from no diabetes medications in the pre-period to insulin in the post-period, as well as variables that were protective from readmission, such as having an outpatient encounter following index hospitalization discharge. The resulting AUC statistics for the 500 model iterations in the training and validation datasets are summarized in Table 3. The predictive accuracy of the final logistic regression model, for each iteration, had a mean AUC of 0.699, ranging from 0.687 to 0.710. Similar accuracy in prediction was seen when the model was applied to the validation data, where the mean AUC was 0.693 (range 0.679–0.709).
Discussion This analysis identified factors predictive of 30 day readmission in patients with type 2 diabetes. In our analysis, escalation of diabetic treatments (especially those escalated from having no records of anti-diabetic medications 112
Predictors of 30 day hospital readmission Eby et al.
to treatment with insulin) was the strongest predictor of 30 day readmission with a mean OR estimate of 6.557 (vs. no treatment). Since treatment escalation was such a strong predictive variable, this study underlines the importance of early and appropriate interventions for patients with diabetes. Although treatment escalation could have been related to other variables (e.g., time to first encounter, hypoglycemia episodes, adverse effects of treatment), these findings reflect the importance of close follow-up of patients who are undergoing treatment intensification. Additional factors associated with greater likelihood of 30 day readmission include longer index hospital stays and more pre-period emergency room and inpatient visits. Because being diagnosed with diabetes prior to the index stay was protective of 30 day readmission (mean OR: 0.675), this study supports the importance of accurate diagnosis and appropriate follow-up for patients who may be at risk for diabetes (e.g., patients with one or more glucose tests that may indicate diabetes and for whom followup diagnostics would be encouraged). The association of a higher readmission risk in those patients without an established diagnosis might reflect inaccurate diagnosis or could reflect a more rapid course with less access to care and www.cmrojournal.com ! 2015 Informa UK Ltd
Curr Med Res Opin Downloaded from informahealthcare.com by University of Otago on 07/12/15 For personal use only.
Current Medical Research & Opinion
fewer opportunities for evaluation by endocrinologists or other clinicians. This variable (absence of an established diagnosis) might also be a surrogate marker for lower quality of care. Accurate diagnosis and appropriate follow-up are required for both patients with diabetes and for those who might be at risk for diabetes (i.e., those who meet some diagnostic criteria but are not yet diagnosed). From the ROC analysis (mean AUC of 0.693), the predictive accuracy of the final logistic regression model is considered modest. This result might be due to the unavailability of some variables or data that influence readmission from the EHR database and, thus, from inclusion into the predictive model, or because readmission itself is not amenable to more precise prediction from any other observed or unobserved variables. Many of the variables identified as predictive of 30 day readmission in this analysis are consistent with those found in previous studies of short-term hospital readmission among patients with diabetes. Patients with diabetes and greater comorbidity burden9, particularly those with congestive heart failure7,9,21, chronic obstructive pulmonary disease7, or end-stage organ disease9, have been shown to experience more frequent hospital readmission; this result is consistent with our findings. The absence of a diabetes diagnosis prior to the index stay has also been previously shown to be predictive of 30 day readmission8, which is an intriguing finding also corroborated by our analysis. Interestingly, our descriptive results showed no significant differences in HbA1c values for the readmitted versus the non-readmitted patients. Race and insurance status have been identified as important contributors to readmission risk among patients with diabetes, and these findings are also consistent with our analysis: non-Caucasian patients with diabetes, particularly African Americans, are more likely to be readmitted than are Caucasian patients2,13,22, and patients with public insurance (e.g., Medicare and/or Medicaid) are more likely to be readmitted than those with private insurance2,8,9,22. The variables related to race and insurance status may reflect, at least in part, socioeconomic factors (e.g., lower income, reduced access to healthcare coverage and resources) associated with these populations, as opposed to characteristics of these populations’ diabetes severity or treatments. Other studies have measured the effectiveness of various interventions at reducing hospital readmission risk in patients with diabetes. Although different diabetic treatment regimens did not affect post-discharge outcomes in one study23, another study found that team interventions (i.e., a nurse educator and an endocrinologist) significantly reduced the rate of recurrent hospitalizations24. In patients with diabetes who were hospitalized for myocardial infarction, the type of glucose-lowering treatment received as monotherapy was not associated with differences in prognosis or readmission risk25. ! 2015 Informa UK Ltd www.cmrojournal.com
Volume 31, Number 1
January 2015
Limitations of this study include that hospitalizations and other encounters outside the Humedica network were not captured in this database. Some variables related to diabetes (e.g., duration of diabetes) were not available and were not included in the analysis. HbA1c values were not available for all patients and, thus, were only reported descriptively for those who did have values; therefore, HbA1c values were not included in the predictive model. Additionally, data on medication compliance were not available, and severity of disease and reasons for hospitalization were not provided for every patient. Furthermore, as in any retrospective study using EHR data, these results rely on the accuracy of reporting and coding within the database (e.g., comorbidities during the index hospitalization may be underreported, as these were identified by ICD-9 codes in the EHR data). An additional potential limitation is that the statistical methods used stepwise logistic regression and were only limited to main effects of the variable, rather than interactions. Strengths of the current analysis include the consistency with previous studies on this topic, the inclusion of both medical and surgical patients in the study population, and the application of the novel variable of ‘diabetic treatment escalation’. To our knowledge, this is the first such analysis to consider the effect of treatment escalation on short-term hospital readmission among patients with type 2 diabetes.
Conclusion The identification of factors predictive of 30 day hospital readmission among patients with type 2 diabetes underlines the importance of appropriate recognition of and treatment for diabetes prior to and during hospitalization. Approaches to managing diabetes that take these factors and patient characteristics into account may lead to reduced hospital readmission rates and, thus, to reduced cost burdens to patients, payers, and healthcare providers.
Transparency Declaration of funding This study was funded by Eli Lilly and Company. All authors were involved in study design and data analysis; all authors developed, reviewed, and edited the manuscript; and all authors contributed to the discussion. Declaration of financial/other relationships C.H., M.Y., S.G., K.D., and J.X. have disclosed that they are employees of Eli Lilly and Company. E.E. has disclosed that she was an employee of Eli Lilly and Company at the time this study was completed. T.G. has disclosed that he has no significant relationships with or financial interests in any commercial companies related to this study or article.
Predictors of 30 day hospital readmission Eby et al.
113
Current Medical Research & Opinion Volume 31, Number 1
January 2015
CMRO peer reviewers on this manuscript have no relevant financial or other relationships to disclose.
Curr Med Res Opin Downloaded from informahealthcare.com by University of Otago on 07/12/15 For personal use only.
Acknowledgments The authors wish to thank Karla C. Villines of Eli Lilly and Company for critical review of the manuscript, and Jeffrey Walter and Maria Rovere of _ospitali Health Clinical for assistance writing and editing the manuscript, respectively. Previous presentation: Portions of this analysis were presented at the 73rd Scientific Sessions of the American Diabetes Association, 23 June 2013, Chicago, IL, USA.
References 1. American Diabetes Association. Economic costs of diabetes in the US in 2012. Diabetes Care 2013;36:1033-46 2. Jiang HJ, Stryer D, Friedman B, et al. Multiple hospitalizations for patients with diabetes. Diabetes Care 2003;26:1421-6 3. Fraze T, Jiang J, Burgess J. Hospital Stays for Patients with Diabetes, 2008. Healthcare Cost and Utilization Project (HCUP) Statistical Brief #93, Rockville, MD: Agency for Healthcare Research and Quality, 2010 4. Allaudeen N, Schnipper JL, Orav EJ, et al. Inability of providers to predict unplanned readmissions. J Gen Intern Med 2011;26:771-6 5. Stone J, Hoffman GJ. Medicare Hospital Readmissions: Issues, Policy Options and PPACA. Congressional Research Service, 2010. Available at: http:// www.ncsl.org/documents/health/Medicare_Hospital_Readmissions_and_ PPACA.pdf [Last accessed 26 February 2014] 6. Kansagara D, Englander H, Salanitro A, et al. Risk prediction models for hospital readmission a systematic review. JAMA 2011;306:1688-98 7. Smith DM, Giobbie-Hurder A, Weinberger M, et al. Predicting nonelective hospital readmissions: a multi-site study. J Clin Epidemiol 2000;53:1113-18 8. Robbins JM, Webb DA. Diagnosing diabetes and preventing rehospitalizations: the urban diabetes study. Med Care 2006;44:292-6 9. Kim H, Ross JS, Melkus GD, et al. Scheduled and unscheduled hospital readmissions among patients with diabetes. Am J Manag Care 2010;16:760-7 10. Kanel K, Elster S, Vrbin C. PHRI Readmission Briefs. Brief I: Overview of Six Target Chronic Diseases. Pittsburgh, PA: Pittsburgh Regional Health Initiative, 2010
114
Predictors of 30 day hospital readmission Eby et al.
11. McHugh MD, Carthon JM, Kang XL. Medicare readmissions policies and racial and ethnic health disparities: a cautionary tale. Policy Polit Nurs Pract 2010;11:309-16 12. Hernandez AF, Greiner MA, Fonarow GC, et al. Relationship between early physician follow-up and 30-day readmission among Medicare beneficiaries hospitalized for heart failure. JAMA 2010;303:1716-22 13. Jiang HJ, Andrews R, Stryer D, et al. Racial/ethnic disparities in potentially preventable readmissions: the case of diabetes. Am J Public Health 2005;95:1561-7 14. van Walraven C, Dhalla IA, Bell C, et al. Derivation and validation of an index to predict early death or unplanned readmission after discharge from hospital to the community. CMAJ 2010;182:551-7 15. Ginde AA, Blanc PG, Lieberman RM, et al. Validation of ICD-9-CM coding algorithm for improved identification of hypoglycemia visits. BMC Endocr Disord 2008;8:4 16. Chernick MR. Bootstrap Methods: A Guide for Practitioners and Researchers, 2nd Edn. Hoboken, NJ: John Wiley & Sons Inc., 2008 17. Austin PC, Tu JV. Automated variable selection methods for logistic regression produced unstable models for predicting acute myocardial infarction mortality. J Clin Epidemiol 2004;57:1138-46 18. Derksen S, Keselman HJ. Backward, forward, and stepwise automated subset selection algorithms: frequency of obtaining authentic and noise variables. Br J Math Stat Psychol 1992;45:265-82 19. Efron B, Gong G. A leisurely look at the bootstrap, the jackknife, and crossvalidation. Am Stat 1983;37:36-48 20. Deyo R, Cherkin DC, Ciol MA. Adapting a clinical comorbidity index for use with ICD-9-CM administrative databases. J Clin Epidemiol 1992;45:613-19 21. Cramer S, Chapa G, Kotsos T, et al. Assessing multiple hospitalizations for health-plan-managed Medicaid diabetic members. J Healthc Qual 2010;32:7-14 22. Estrada CL, Danielson KK, Drum ML, et al. Hospitalization subsequent to diagnosis in young patients with diabetes in Chicago, Illinois. Pediatrics 2009;124:926-34 23. Sarma S, Mentz RJ, Kwasny MJ, et al. Association between diabetes mellitus and post-discharge outcomes in patients hospitalized with heart failure: findings from the EVEREST trial. Eur J Heart Fail 2013;15:194-202 24. Koproski J, Pretto Z, Poretsky L. Effects of an intervention by a diabetes team in hospitalized patients with diabetes. Diabetes Care 1997;20:1553-5 25. Horsdal HT, Johnsen SP, Søndergaard F, et al. Type of preadmission glucose-lowering treatment and prognosis among patients hospitalized with myocardial infarction: a nationwide follow-up study. Diabetologia 2008;51:567-74
www.cmrojournal.com ! 2015 Informa UK Ltd
