We present an International Classification of Diseases, Tenth Revision (ICD-10) translation of the adapted Diabetes Complications Severity Index and show its performance in predicting hospitalizations, mortality, and healthcare-associated costs.
Objectives: To assess the performance of the adapted Diabetes Complications Severity Index (aDCSI) translated to International Classification of Diseases, Tenth Revision (ICD-10) in predicting hospitalizations, mortality, and healthcare-associated costs.
Study Design: Retrospective closed cohort study based on secondary data analysis.
Methods: We translated the aDCSI to ICD-10 and calculated aDCSI scores based on health insurance claims data. To assess predictive performance, we used multivariate regression models to calculate risk ratios (RRs) of hospitalizations and mortality and linear predictors of cost.
Results: We analyzed a sample of 157,115 patients with diabetes mellitus. RRs of hospitalizations (total and cause specific) rose with increasing aDCSI scores. Predicting total hospitalizations over a 4-year period, unadjusted RRs were 1.22 for an aDCSI score of 1 (compared with a score of 0), 1.55 for a score of 2, 1.77 for a score of 3, 2.11 for a score of 4, and 2.72 for scores of 5 and higher. Cause-specific hospitalizations and mortality showed similar results. Costs clearly increased in each successive score category.
Conclusions: Our study supports the validity of the aDCSI as a severity measure for complications of diabetes, as it correlates to and predicts total and cause-specific hospitalizations, mortality, and costs. The aDCSI’s performance in ICD-10—coded data is comparable with that in International Classification of Diseases, Ninth Revision—coded data.
Am J Manag Care. 2019;25(2):e45-e49Takeaway Points
Diabetes is among the top 10 causes of death worldwide, and its global prevalence is increasing.1-3 Healthcare expenditures for a patient with diabetes are more than twice as high as they are for an average patient, with costs mainly driven by inpatient care and medications used to treat diabetes-related complications.4,5 Risks of hospitalization, mortality, and healthcare costs are associated with the number and severity of diabetes-related complications.6-8 There is an urgent need for a validated diabetes-specific severity score for comorbidities to adjust for differences in diabetes-specific morbidity in a large array of studies. To enable epidemiological studies of large claims databases, such a score would ideally not include clinical data.
To systematically quantify diabetes complications, Young and colleagues developed the Diabetes Complications Severity Index (DCSI).6 The DCSI uses 7 categories of diabetes complications (ophthalmic, renal, neurologic, cerebrovascular, cardiovascular, peripheral vascular, and metabolic). Each category is scored with either 0 (no complication), 1 (nonsevere complication), or 2 (severe complication), except for neurologic complications, which score a maximum of 1. The highest possible DCSI score is therefore 13. As the DCSI needs laboratory values to calculate scores, Chang et al developed the adapted DCSI (aDCSI) specifically for use with claims data, which rarely contain laboratory results.9 The DCSI and aDCSI were validated with regard to hospitalization risk, mortality risk, and healthcare costs in patients with diabetes.6-9 Different approaches exist for the validation and use of the aDCSI, including correlative use to reflect current severity of diabetes complications and predictive use to reflect future risk of hospitalization, mortality, or costs. For example, Chang and colleagues9 correlated aDCSI scores over a 4-year period with hospitalizations during the same period, whereas Chen and Hsiao7 additionally reported on predictive performance. The development of both DCSI and aDCSI was based on International Classification of Diseases, Ninth Revision (ICD-9) codes. Although Glasheen et al10 translated the DCSI to International Classification of Diseases, Tenth Revision (ICD-10) and Wilke et al8 used the aDCSI with ICD-10 data, no systematic validation of the aDCSI with ICD-10 data has been performed with long-term follow-up. Because most healthcare systems use ICD-10—based diagnostic coding, studies on the performance of the aDCSI with ICD-10 data are urgently needed, especially after the switch from ICD-9 to ICD-10 in the United States in 2015.
Based on an analysis of health insurance claims data, the aims of this study are to translate and adapt the aDCSI to ICD-10 and show its performance for correlation with and prediction of hospitalizations, mortality, and healthcare costs over 4 years.
Study Design and Setting
We conducted a retrospective closed cohort study based on secondary data analysis. The study is based on claims data from a large statutory health insurance fund in southern Germany (AOK Baden-Württemberg). This insurer covers 4 million inhabitants of the state of Baden-Württemberg (almost half the state’s population). Our data set contains data on all continuously insured persons 18 years and older living in Baden-Württemberg (see eAppendix Table 1 [eAppendix available at ajmc.com] for the complete list of inclusion criteria). The data include comprehensive information on ambulatory care, drug prescriptions, and hospital care. The analysis was carried out as part of an evaluation of general practitioner—centered healthcare and was fully approved by the Ethics Committee of Frankfurt University Hospital.
Individuals who had a diagnosis of diabetes mellitus in 2010 (ICD-10 codes E10-E14) and were receiving antidiabetic medication were included in the study cohort. To further increase diagnostic specificity, we required a diagnosis of diabetes to be coded in at least 3 calendar quarters per year. This should exclude cases of gestational diabetes.
On the basis of expert consensus, the ICD-9 version of the aDCSI9 was translated to ICD-10 by the Institute of General Practice (University of Frankfurt) and the PMV Research Group at the Department of Child and Adolescent Psychiatry, Psychosomatics, and Psychotherapy (University of Cologne). The suggestions made by Wilke and colleagues8 were taken into consideration in the translation. Because our translation was undertaken prior to the translation of the DCSI by Glasheen and colleagues,10 their work was not considered. Our primary translation was to ICD-10-GM (German Modification, 201311), and our analysis is based on this translation. We also present a translation to the CDC’s ICD-10, Clinical Modification (ICD-10-CM)12 for comparison. The detailed translation can be found in eAppendix Table 2.
Based on our translation, we calculated aDCSI scores for all patients in our study cohort depending on their status in 2010. We assessed outcomes during the observation period from the beginning of 2011 to the end of 2014. Clinical outcomes included overall hospitalizations; hospitalizations due to stroke, coronary heart disease (CHD), and acute renal failure; and mortality. To predict cost, we calculated total costs of inpatient care and costs of prescription medication per patient from 2011 to 2014. Costs are those that insurance reimbursed and do not include private costs (eg, over-the-counter medications).
Performance in Correlative Use
To facilitate comparison with previous results, we also calculated aDCSI scores based on all coded diagnoses in the entire observation period (2011-2014). We calculated models based on this “cumulative” aDCSI for overall hospitalizations during the observation period.
We used univariate and multivariate regression models to calculate risk ratios (RRs) for overall hospitalizations and specific hospitalizations due to stroke, CHD, and acute renal failure. The aDCSI was included as a categorical variable (0, 1, 2, 3, 4, and ≥5) in our regression models. Because of overdispersion in Poisson models (as previously reported9), we used negative binomial models to calculate RRs. Negative binomial models did not show overdispersion, and results were very similar to those in the Poisson model. We therefore show them in comparison with the results of Young and colleagues6 and Chang and colleagues.9 For direct medical cost, a linear regression model was used to estimate incremental increase in cost per unit of the aDCSI score. We used Cox’s proportional hazards models to estimate hazard ratios for all-cause mortality. Adjusted and unadjusted models were calculated for age, sex, and insulin use. For a comparison of the aDCSI’s performance in predictive and correlative use, we calculated the area under the curve (AUC) of the receiver operating characteristic (ROC) curve. Descriptive analyses and modeling were performed in SAS version 9.3 (SAS Institute Inc; Cary, North Carolina) and ROC analysis in SPSS version 22 (IBM Corp; Armonk, New York).
Our study cohort included 157,115 patients with diabetes. The mean age was 69.5 years and 52.7% of patients were female. A total of 134,163 (85.4%) patients were alive at the end of our observation period. Further characteristics are summarized in Table 1.
Patients with higher aDCSI scores tended to be older and have higher general comorbidity, as assessed using the Charlson Comorbidity Index. Insulin use and mortality rose with increasing aDCSI score.
Adjusted and unadjusted multiple regression models showed increasing RRs of hospitalization with increasing aDCSI scores compared with those without complications (Table 26,9 and Table 36). Correlative use showed larger increases in RRs with increasing aDCSI score. Adjusted and unadjusted RRs showed an approximately linear increase for aDCSI scores of 1 to 5 and above. Hospitalization for ischemic stroke, CHD, and acute renal failure showed similar increases in RRs (Table 4). RRs were lower in the adjusted models, suggesting that part of the increased risk in higher aDCSI score categories is related to covariates, and mainly to older age in these groups. Relative risk of hospitalization was higher in men than women. Assessed using the AUC, the performance of the aDCSI in predictive use (to predict overall hospitalizations) was 0.642 (95% CI, 0.639-0.644). The AUC for correlative use (correlation with overall hospitalizations) was 0.735 (95% CI, 0.732-0.737). The ROC curve is shown in the Figure.
A total of 22,952 deaths occurred during the 4-year observation period from 2011 to 2014. The hazard ratios for mortality increased with a rise in aDCSI score compared with those without complications (eAppendix Table 3). The adjusted model showed the same trend, but adjustment led to a decrease in hazard ratios compared with the unadjusted model. The risk of dying was higher in men than women.
Higher aDCSI scores were related to increased costs of inpatient care and prescription medications. Compared with a score of 0, medication costs roughly doubled and inpatient costs tripled for patients with aDCSI scores of 5 or greater (eAppendix Table 4).
Based on a large and representative cohort, our results show that scores on the ICD-10 translation of the aDCSI reliably stratify risk of hospitalization and mortality and can be used for healthcare cost prediction. In addition, we show for the first time that the aDCSI predicts specific patient-relevant outcomes related to diabetes (hospitalization for ischemic stroke, CHD, and acute renal failure). This is the first validation study of the aDCSI based on a patient cohort from Germany, adding to the external validity of the score.
Correlative use of the aDCSI reveals higher RRs and AUCs than predictive use. This is to be expected, because patients with acute complications, reflected in a higher aDCSI, are more likely to be hospitalized. This is also true when the aDCSI increases following a hospitalization during which complications were diagnosed. Depending on the intended use of the aDCSI, the distinction between correlative use and predictive use should be taken into consideration.
We observed somewhat lower RRs in higher score categories than were described in previous studies, even in correlative analysis.6,7,9 One possible explanation is the relatively high hospitalization rate in Germany (a mean of 2.3 hospitalizations per patient over 4 years in our sample compared with 0.94 hospitalizations over 4 years in the study by Chang et al9).
Disease severity scores based on claims data (and sometimes additional laboratory data) rely on the documentation and validity of coded diagnoses, which in turn depend on national systems of patient care and reimbursement. Claims data often include a certain degree of simultaneous under- and overcoding. Undercoding can occur when clinicians consider diagnoses to be clinically irrelevant (which should barely affect diseases used to calculate the aDCSI), when diseases are not detected (this may be relevant, such as for chronic renal failure, which is asymptomatic in early stages), and when diseases are not reimbursable (this largely depends on the specificities of the respective healthcare system). Overcoding can result when codes of acute diseases are not removed from electronic management systems when the acute disease has subsided. In Germany, previous studies have shown that the average number of diagnoses per patient is high compared with other countries.13 Country-specific modalities like these are likely to contribute to observed differences in the performance of severity scores.
Most diagnostic codes are quite specific for diabetes-related complications, but some are not (this limitation applies to the aDCSI and similar scores in general). Stroke or renal failure, for instance, are common complications of diabetes, but they can have other causes (eg, atrial fibrillation for stroke). Although not specific to diabetes, these factors can still be of prognostic relevance. The accuracy of the documentation of a patient’s diagnosis is important in all applications of the aDCSI.
Besides this general limitation, although our data set covers a broad population, some selection effects are possible, as we were able to analyze only the data from the AOK insurance fund, and this fund has a slightly larger proportion of insured persons without employment and with chronic diseases than other insurers in the state. Nevertheless, AOK Baden-Württemberg covers almost half the state’s population. The population analyzed should therefore be fairly representative for insurance-covered individuals of industrialized societies. However, for the United States, the uninsured constitute a major challenge facing healthcare today, with many experiencing more severe diabetes and related complications.14,15 A limitation in the methods used in this study is that we did not take person-time at risk into account (ie, those that die early on during the observation period are no longer at risk of hospitalization) when calculating risk of hospitalization and costs. To ensure comparability with previous studies, we decided against the use of more sensitive statistical methods.
Besides prediction, we believe that the main use of the aDCSI is to adjust multivariate models in health service research to reflect the severity of diabetes complications. In this regard, a standardized translation to ICD-10 facilitates the comparability of studies. There are minor differences between the ICD-10-GM used in this study and the CDC’s ICD-10-CM that may limit direct comparison, but we believe that the effects are small. However, when using the aDCSI, coding conventions for the population under review should be taken into account.
Our study supports the overall validity of the aDCSI and its translation to ICD-10 as a severity measure for diabetes complications, as it correlates to and predicts overall and cause-specific hospitalizations, mortality, and cost.Author Affiliations: Institute of General Practice, Goethe University (FSW, AG, RL, MH, MB, KK), Frankfurt am Main, Germany; PMV Research Group, University of Cologne (IS, IK), Cologne, Germany.
Source of Funding: The statutory health insurer AOK Baden-Württemberg supported our study and provided the administrative data sources/claims data. The funder had no role in the analysis, interpretation, or publication of the data.
Author Disclosures: The authors report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article.
Authorship Information: Concept and design (FSW, AG, IS, IK, RL, MH, MB, KK); acquisition of data (RL); analysis and interpretation of data (FSW, AG, MH, MB, KK); drafting of the manuscript (FSW, KK); critical revision of the manuscript for important intellectual content (AG, IS, IK, MB, KK); statistical analysis (AG); administrative, technical, or logistic support (AG, RL, MH, MB, KK); supervision (IS, IK, MB, KK); and correcting translation of adapted Diabetes Complications Severity Index (IS).
Address Correspondence to: Felix Sebastian Wicke, Dr Med, Institute of General Practice, Goethe University, Theodor-Stern-Kai 7, 60590 Frankfurt am Main, Germany. Email: email@example.com.REFERENCES
1. The top 10 causes of death. World Health Organization website. who.int/news-room/fact-sheets/detail/the-top-10-causes-of-death. Published May 24, 2018. Accessed January 9, 2019.
2. NCD Risk Factor Collaboration (NCD-RisC). Worldwide trends in diabetes since 1980: a pooled analysis of 751 population-based studies with 4.4 million participants. Lancet. 2016;387(10027):1513-1530. doi: 10.1016/S0140-6736(16)00618-8.
3. Cho NH, Shaw JE, Karuranga S, et al. IDF Diabetes Atlas: global estimates of diabetes prevalence for 2017 and projections for 2045. Diabetes Res Clin Pract. 2018;138:271-281. doi: 10.1016/j.diabres.2018.02.023.
4. American Diabetes Association. Economic costs of diabetes in the U.S. in 2012. Diabetes Care. 2013;36(4):1033-1046. doi: 10.2337/dc12-2625.
5. Köster I, Huppertz E, Hauner H, Schubert I. Direct costs of diabetes mellitus in Germany - CoDiM 2000-2007. Exp Clin Endocrinol Diabetes. 2011;119(6):377-385. doi: 10.1055/s-0030-1269847.
6. Young BA, Lin E, Von Korff M, et al. Diabetes Complications Severity Index and risk of mortality, hospitalization, and healthcare utilization. Am J Manag Care. 2008;14(1):15-23.
7. Chen HL, Hsiao FY. Risk of hospitalization and healthcare cost associated with Diabetes Complication Severity Index in Taiwan’s National Health Insurance Research Database. J Diabetes Complications. 2014;28(5):612-616. doi: 10.1016/j.jdiacomp.2014.05.011.
8. Wilke T, Mueller S, Groth A, et al. Treatment-dependent and treatment-independent risk factors associated with the risk of diabetes-related events: a retrospective analysis based on 229,042 patients with type 2 diabetes mellitus. Cardiovasc Diabetol. 2015;14:14. doi: 10.1186/s12933-015-0179-2.
9. Chang HY, Weiner JP, Richards TM, Bleich SN, Segal JB. Validating the adapted Diabetes Complications Severity Index in claims data. Am J Manag Care. 2012;18(11):721-726.
10. Glasheen WP, Renda A, Dong Y. Diabetes Complications Severity Index (DCSI)—update and ICD-10 translation. J Diabetes Complications. 2017;31(6):1007-1013. doi: 10.1016/j.jdiacomp.2017.02.018.
11. Internationale statistische Klassifikation der Krankheiten und verwandter Gesundheitsprobleme (10. Revision): German Modification. Deutsche Institut für Medizinische Dokumentation und Information website. https://www.dimdi.de/static/de/klassifikationen/icd/icd-10-gm/kode-suche/htmlgm2013/. Published September 21, 2012. Accessed January 9, 2019.
12. International Classification of Diseases, Tenth Revision: Clinical Modification (ICD-10-CM). CDC website. cdc.gov/nchs/icd/icd10cm.htm. Updated July 26, 2016. Accessed May 1, 2018.
13. Erler A, Beyer M, Muth C, Gerlach FM, Brennecke R. Garbage in — garbage out? Validität von Abrechnungsdiagnosen in hausärztlichen Praxen. Dtsch Gesundheitsw. 2009;71(12):823-831.
14. Hu R, Shi L, Rane S, Zhu J, Chen CC. Insurance, racial/ethnic, SES-related disparities in quality of care among US adults with diabetes. J Immigr Minor Health. 2014;16(4):565-575. doi: 10.1007/s10903-013-9966-6.
15. Ridic G, Gleason S, Ridic O. Comparisons of health care systems in the United States, Germany and Canada. Mater Sociomed. 2012;24(2):112-120. doi: 10.5455/msm.2012.24.112-120.