The original Charlson Comorbidity Index (CCI) chart review instrument designed by Charlson and colleagues produced a morbidity score that reflects mortality risk.1 The score is determined based on 19 medical conditions and adjusts for variable morbidity rates within a patient population. In clinical practice, risk assessment facilitates triage, prioritization, and proactive patient engagement. More recently, various administrative claims data versions of this public domain instrument have enabled healthcare organizations, payers, and researchers to adjust for mortality risk in claims-based studies of large patient populations. Although mortality risk assessment was the original intent of the CCI scoring instrument, the correlation of mortality risk with expected healthcare resource consumption expands the usefulness of the instrument. The use of the CCI facilitates the prioritization of care-management resources based on patient risk.
The CCI scoring system is useful for several situations, including clinical settings in large population healthcare organizations and for research purposes. This scoring tool is easily administered and should yield near-identical results, regardless of whether it is used in the context of patient examinations, chart review, administrative data, or autopsy. This flexibility is what distinguishes the CCI instrument from other risk-adjustment and risk-assessment tools.
The original CCI chart review instrument was based on 19 different medical conditions categories.1 Each category was assigned a score (weight) of 1, 2, 3, or 6.1 These condition-specific scores were summed for the overall CCI score. Implicit in the instrument design was a hierarchical structure in which a more severe condition trumped a less severe condition. For example, 1 point was assigned to mild liver disease and 3 points were assigned to moderate or severe liver disease1; only the more severe category’s score was active when both diagnoses were included in a medical record. A score for age was incorporated to account for mortality risk in the absence of clinical diagnoses: 1 point was added per decade, starting with the 50s. The final score was a sum of the scores for active condition categories plus the age adjustment.
An early CCI instrument based on diagnosis codes translated chart review condition categories into International Classification of Diseases, Ninth Revision (ICD-9) diagnosis codes with 3-digit specificity.2 The use of diagnostic coding for reimbursement purposes creates an incentive for providers to use more accurate and precise coding. Accordingly, newer claims-based versions of the CCI, such as the frequently used instrument designed by Deyo and colleagues, incorporated 4- and 5-digit codes.3
Why Propose a New Coding Scheme?
The Need for an ICD-10 Coding System
In October 2015, the US healthcare system transitioned from ICD-9 to International Classification of Diseases, Tenth Revision (ICD-10). This created an immediate need for a CCI coding algorithm that included ICD-10 codes. An increase in the number of codes from 14,025 to 69,823 also facilitated more nuanced and clinically updated categorization and severity assessment.4 Many countries use or have adapted the ICD-10 for their own healthcare systems.5,6 The ICD-10 CCI instrument created by Quan and colleagues for use in Canada includes the diabetes code families of E10, E11, E12, E13, and E14, whereas the current US version of ICD-10 uses E08, E09, E10, E11, and E13.7 An ICD-10 coding algorithm for a version of the CCI that is specific to the United States has not yet been published.
Reasons for Creating a New ICD-9 Scoring System
We identified 3 reasons to create a new ICD-9 scoring system for the CCI before designing a corresponding ICD-10 system. The first reason concerned the categorization of diabetes. Diabetes is largely represented by 2 ICD-9 parent codes: 249 (secondary diabetes) and 250 (types 1 and 2 diabetes). ICD-9 also allows diabetes to be coded as type 2 or unspecified (250.x0, 250.x2), which can result in misclassification. Older CCI coding schemes based on ICD-9 have not incorporated ICD-9 code 249, which prevents appropriate risk categorization for patients with secondary diabetes.
A second reason to update the ICD-9–based instrument was to reflect developments in the treatment of HIV infection, which in isolation is no longer associated with near-term (90-day) mortality. The category for AIDS created by Charlson and colleagues explicitly excluded HIV-positive status, whereas instruments based on administrative data either did not differentiate between the 2 conditions3,7,8 or did not include AIDS and HIV infection as clinical categories.9 Quan and colleagues assigned to HIV infection the same weight (6 points) that was assigned to metastatic cancer.7 Adding to the problem, the World Health Organization has discontinued using the code for AIDS in ICD-9 and ICD-10, leaving only HIV-positive category as a valid diagnosis.10
A final opportunity we found to improve the ICD-9 instrument involved the hierarchies implied by any risk-assessment instrument. Condition classification systems, such as diagnosis-related groups and commercial risk-adjustment tools, rely on the fundamental principle that the most severe diagnosis made within a time period for a particular body system is the only system activated when setting risk weights.
To encourage the proper observance of these hierarchies, more explicit guidance on how to apply the instrument was warranted, which created an opportunity to address the severity spectrum of renal disease. The differences in morbidity and risk for mortality between patients with stage 5 or end-stage kidney disease and less severe disease are considerable. Other condition categories, such as cancer, may also have been updated, but we chose to remain consistent with the original emphasis on chronic disease.
Our study was intended to meet 4 main objectives: to create an expanded and refined ICD-9 scoring system for the CCI, taking into account the 3 clinical conditions mentioned earlier (ie, diabetes, HIV/AIDS, renal disease); to create an ICD-10 CCI scoring system with features that correspond to the refined ICD-9 system; to assess the comparability of the ICD-9–based and ICD-10–based scores; and to validate the new scoring systems by testing their association with same-year hospital admissions and near-term mortality.
The essentials of this new set of scoring system are:
- Disease-specific refinements
- Matching translation to ICD-10
- Flexibility to allow use as a chart review instrument, as originally conceived by Mary Charlson.
We refer to this new set of scoring scheme as CDMF CCI to reflect these characteristics.
Our new instrument was designed to yield a similar score whether the diagnoses were identified during face-to-face interaction with a patient or based on administrative data. Thus, the structure of the new instrument was kept consistent with the original chart review instrument published by Charlson and colleagues.1 We also kept our scheme congruous with the administrative data versions created by Deyo and others.3,7-9 The tool proposed here followed Deyo’s decision to collapse 3 nonmetastatic cancer categories into 1 category,3 but did create 2 additional categories by splitting renal disease into mild-moderate and severe categories, and by splitting the HIV/AIDs category into HIV-positive without AIDs and HIV-positive with AIDS. (Although liver disease was categorized as mild and moderate to severe in the original instrument, in this instrument renal disease has been categorized as mild to moderate and severe to reflect the dramatic increase in risk when a patient transitions to stage 5 kidney disease.)
The AIDS category was created by combining codes for HIV infection with those for opportunistic infections or AIDS-related cancers. In addition, the ICD-9 code family for secondary diabetes (249.x) was incorporated into the CDMF coding scheme, corresponding to E08, E09, and E13 in ICD-10. The primary tools we used to identify diagnosis codes were www.ICD9Data.com and www.ICD10Data.com, which are matched resources with suggested crosswalks between ICD-9 and ICD-10.
The point values we assigned to the new instrument were the same as those suggested by the original CCI instrument, with a few exceptions. For renal disease, we changed the 2 points for a single category to 1 for mild-to-moderate disease and to 3 for severe renal disease. We assigned 3 points to HIV without AIDS, based on our authors’ consensus, and ascribed 6 points to AIDS as is the case in previous scoring versions.
We tested the ICD-9 and ICD-10 scoring systems in the Medicare Advantage population of Humana (a national health and wellness organization), including only individuals who were continuously enrolled in Humana Medicare Advantage in at least 1 of 3 consecutive 12-month periods. Individuals were excluded if they died during the 12-month continuous enrollment period or if they were aged <65 years or >110 years. Those with dual eligibility for Medicaid and Medicare were also excluded.
The initial two 12-month time windows were the last 1-year periods in which ICD-9 codes were used exclusively in the United States (ie, October 2013-September 2014 and October 2014-September 2015). The third time window was the earliest 1-year window where ICD-10 codes were used exclusively (ie, October 2015-September 2016). The CCI scoring was based on diagnosis codes in claims for services received during these time periods.
New Diagnosis Codes and Scoring Systems
The primary products of our initiative were scoring schemes and ICD-9 and ICD-10 code tables for the 19 CCI medical conditions. Appendix I Table SI-1 shows the points (1, 2, 3, or 6) we assigned to each medical condition. Of these 19 conditions, 11 received 1 point. Especially serious conditions or severe levels of a condition received more points (eg, 1 point for diabetes without chronic complications and 2 points for diabetes with chronic complications).
Appendix I Table SI-2 displays the 6 hierarchy categories. In each category, only the more severe condition should contribute to the CCI score when codes for both conditions are listed on an individual’s claims record; for example, if an individual has cerebrovascular disease (1 point) and hemiplegia or paraplegia (2 points), only the 2 points for hemiplegia or paraplegia are counted. Table 1A and Table 1B provide an illustration of how the Deyo instrument, the new ICD-9 system, and the new ICD-10 system compare for HIV/AIDS. The full set of condition-specific tabular comparisons is shown in Appendix I Tables SI-3a to SI-3s and is designed to allow the replication of the new scheme.
We used 4 sets of analyses to assess the performance of the scoring systems. First, we computed the prevalence of each of the 19 CCI condition categories for all 3 periods and assessed for consistency. Second, as a preliminary validation of the updated CCI instrument, we assessed the association between the CCI score and the current-year hospital admissions (marked by discharge dates) and the association with near-term (90-day) mortality. The relationship between CCI score and admissions was evaluated by using a linear regression model to predict the mean admissions per 1000, which was adjusted for sex (reference, female) and race (reference, white), using 13 binary variables for CCI scoring (reference value, 2). This approach allowed for an assessment of whether the relationship between the CCI score and utilization was linear.
We constructed a robust Poisson regression model11 to assess the relationship between a CCI score and mortality in the 3-month period after the end of the analytic time window (October-September) and before the first enrollment month of a new plan year (January). We chose to use a robust Poisson regression model to predict near-term mortality, which allows for the direct modeling of relative risks. As in the utilization model, we included sex and race as covariates in addition to the CCI score (represented as 13 binary variables for the CCI score; reference value, 2).
Third, we applied the Deyo system and the new ICD-9 system to the later of the 2 ICD-9 era study panels. The mean CCI score or disease prevalence within the subpopulations with diagnoses were related to key changes in the new instruments. Using the Deyo scoring system, these subpopulations were identified within the October 2014 to September 2015 panel: HIV/AIDS, diabetes, and renal disease. As noted earlier, the Deyo system ascribes the same score to all individuals with HIV infection regardless of whether the patient has AIDS; it does not include the codes for secondary diabetes; and it does not take the severity of renal disease into account. Each subpopulation was then further divided according to the more granular scoring allowed by the new ICD-9 system, and the 2 sets of mean CCI scores (HIV/AIDS, renal disease) or 2 sets of prevalence values (diabetes) were computed for each resulting group.
Finally, we constructed 3 logistic regression models for the prediction of near-term mortality, using CCI score as the independent variable, and the area under the receiver operating characteristics (AUC-ROC) curve was calculated for each model. These models computed CCI score according to the Deyo system, the new ICD-9 system (for the October 2014-September 2015 panel), and the new ICD-10 system; the same hierarchical scheme shown in Appendix I Table SI-2 was applied in all 3 calculations.
The research supporting our new instruments was performed by a biomedical engineer (WPG), who had previously developed an SAS algorithm using claims data that was based on the CCI scoring scheme of the Deyo version of the CCI. Our new code sets were reviewed by a senior physician (AR), who is familiar with administrative data and with Humana’s enhancement of the Diabetes Complications Severity Index.12 A second physician (JD) then performed a check on this work and created code-scoring tables for each of the CCI dimensions, showing separate codes for the original Deyo ICD-9 CCI, the CDMF CCI with ICD-9 codes, and the CDMF CCI applied to ICD-10.
These 3 researchers then resolved any remaining inconsistencies. An SAS program was created that reflected the ICD-9 scoring modifications and the new ICD-10 scoring system. In an iterative process, the prevalence of the 19 conditions was compared between ICD-9 and ICD-10 populations to confirm the performance of the code, and modifications were made in response to any discrepancies.
In several of the analyses, the new ICD-9 and ICD-10 scoring systems yielded comparable findings across similar populations. Appendix II Table SII-1 shows the composition of the three 1-year population panels. The populations grew from 1,791,171 to 2,163,082 from the first (October 2013-September 2014) to the third (October 2015-September 2016) panel. The demographics were very similar between the 3 panels, with a near-identical distribution of the age-sex subgroups. From the first to the third panel, the populations were on average 0.16 years younger, included 0.69% more men, and included 2.59% more individuals of nonwhite race. Although the differences were small, they were significant (nonoverlapping 95% confidence intervals [CIs]) as a result of the size of the populations.
Appendix II Table SII-1 also presents the mean CCI scores. Although the mean CCI scores were not significantly different between the first and second panels, there was a significant decrease, as shown by nonoverlapping 95% CIs, from the second panel (mean CCI score, 4.915) to the third panel (mean CCI score, 4.883) with the shift from ICD-9 to ICD-10. The number and proportion of individuals with each of the 19 conditions, after application of the 6 condition hierarchies, is shown in Appendix II Table SII-2; these data reveal a similar morbidity profile across the 3 populations. There was some shifting of classification within hierarchies with the transition from ICD-9 to ICD-10, notably with a greater proportion of diabetes complications and metastatic cancer present in the ICD-10 era. Other changes occurred steadily from the earliest to the latest panels, indicating changes in enrollment composition from year to year and/or possible population morbidity drift, rather than an artifact of changes in the ICD diagnosis classification system.
Appendix II Tables SII-3a and SII-3b show the distribution of CCI scores. In the absence of age adjustment, more than 66% of the population had at least 1 CCI condition category (score of ≥1), with the prevalence steadily decreasing from no condition to the maximum of 20 conditions for a few individuals in the first and third panels (Appendix II Table SII-3a). With the age adjustment, the minimum CCI score was 2 and the mode was 3. The distribution was right skewed and nearly identical for all panels (Appendix II Table SII-3b).
The 2 approaches to validation yielded expected results. Figure 1 shows that mean inpatient admissions per 1000 per year increased monotonically as CCI score values increased from 2 to a capped value of ≥15. Appendix II Table SII-4 reveals nonoverlapping CIs for the means within all 3 panels. The same pattern persisted, again with nonoverlapping CIs, after adjustment for race and sex (Table 2). These data suggest a near-linear relationship with a slightly increasing slope from low to high CCI score level. The 3 utilization models explained 16.8%, 16.7%, and 16.6% of the variability within the 3 panels.
Figure 2 shows the CCI score prevalence (line graph) and the corresponding mortality risk (bar chart) by panel. The line plots demonstrate a Pareto-like (right-skewed) distribution of each population across the morbidity spectrum and a near-linear, monotonic increase in risk from the lowest to the highest CCI score level. There was a very slight increase in mortality versus CCI score slope from low to high CCI score level. Table 3 also shows that the risk for mortality increased monotonically as the CCI score increased, and the 95% CIs were nonoverlapping. After adjustment for sex and race, the same pattern persisted for relative risk estimates, except for a slight overlap between CCI score levels 13 and 14 in the ICD-10 panel (Table 4). All relative risk estimates were significant (P <.0001).
The comparison of scores and prevalence according to the Deyo system and CDMF ICD-9 system showed the expected differences (Table 5). For individuals identified according to the CDMF ICD-9 system as having HIV infection but not AIDS, the mean CCI score was lower according to the CDMF ICD-9 score than according to the Deyo score. Nonoverlapping CIs showed these differences to be significant. The Deyo and the CDMF ICD-9 scores were similar in the group identified by the new ICD-9 system as having AIDS.
In a similar comparison in the subpopulation with renal disease, the CDMF ICD-9 system CCI score was lower than the Deyo score in the group of individuals identified by the CDMF system as having mild or moderate renal disease, and the CDMF ICD-9 score was higher than the Deyo score in the group with severe renal disease; the CIs did not overlap.
In other words, compared with the Deyo system, the CDMF ICD-9 system differentiated more sharply between individuals with mild-to-moderate renal disease and severe renal disease: a difference of 2.84 points (CDMF ICD-9 system) versus a difference of 0.81 (Deyo system). Compared with the Deyo system, the CDMF ICD-9 system identified slightly more individuals with diabetes and was slightly more likely to identify individuals as having severe diabetes, but the CIs overlapped considerably.
The AUC-ROC values for the logistic regression models for near-term mortality were similar for the CDMF ICD-9 system (0.8) and the Deyo system (0.791). A model of the CDMF ICD-10 system had an AUC-ROC value of 0.804.
This study tested the use of a new CCI coding and scoring scheme (CDMF CCI) in 3 population panels of demographically and clinically similar patients in a Medicare Advantage plan. The third panel represented the first 12 months after the adoption in the United States of the ICD-10 codes system. Unadjusted and sex- and race-adjusted analyses of the association of CCI score with same-year inpatient admissions and with near-term mortality showed similar patterns and demonstrated the validity of the CDMF scheme. A comparison of mean scores based on the Deyo system and the CDMF ICD-9 system revealed that by not differentiating between HIV infection and AIDS, the Deyo system may overestimate mortality risk for people with HIV infection only, which is inconsistent with the intent of the original CCI system to specifically identify the mortality risk associated with AIDS.
A similar performance comparison also suggested that the new system differentiates well between individuals with mild-to-moderate renal disease and those with severe renal disease. A significant difference in the ability of the CDMF ICD-9 and the Deyo systems to identify individuals with diabetes was not detected, perhaps because of the relatively low prevalence of secondary diabetes. The overall accuracy of the CDMF and the Deyo CCI schemes was very similar.
The ultimate validation of any claims-based CCI instrument would entail a comparison with medical records data. The degree to which the new instrument proposed here, or any previously published claims-based instrument, is consistent with what might have been detected in the medical record is beyond the scope of this publication. Rather, it was the intent of this effort to lay the groundwork for such an investigation by anchoring the new instrument on the original Charlson instrument.
Choosing Medicare Advantage populations of individuals aged ≥65 years made it possible to investigate the prevalence of a condition in which morbidity was likely. A strength of the CDMF coding scheme is the size and heterogeneous morbidity of the test populations used to evaluate the scheme. The greatest advantage of testing in this population was revealed by the wide morbidity spectrum in each population panel and the consistent distribution across the panels. Although slight demographic trends were observed across time, the differences were small enough to eliminate concern when comparing the prevalence of a condition between years.
No attempt was made to compare this instrument’s condition prevalence performance with that of other claims-based CCI instruments described in the literature. Rather, the primary goals were to make this new instrument reflect Charlson’s original chart-review instrument as closely as possible and reflect the current understanding of mortality risk associated with specific conditions and severities of conditions. Given the lack of a distinction between AIDS and HIV infection and the limited use of condition hierarchies in previously published CCI instruments,1,3,7-9 a fresh start with the original CCI version1 as the benchmark constituted a simpler approach for our new scheme.
Although the prevalence of a condition was generally similar across the 3 time periods, a few differences were observed between the ICD-9 and ICD-10 eras. The most substantial discontinuity had to do with diabetes, where the condition was more likely to be classified as severe in the ICD-10 era. The recently introduced ICD-10 in the United States forces providers and coders to specify the diabetes type and increases the number of categories from 3 to 5, which could lead to greater scrutiny of the medical record by the coder and thus a greater likelihood of accurate documentation of diabetes severity. In another example of discontinuity, the ICD-10 population showed a greater prevalence of metastatic cancer, which was likely a result of the dramatically greater specificity of site required by ICD-10.
A third discrepancy between the ICD-9 and ICD-10 populations of this study was a slight decrease in the prevalence of cerebrovascular disease. The breadth of conditions in this category, however, is large, which may have made it especially difficult to create comparable categories between the 2 code sets. Alternatively, the slight decrease in prevalence may represent diminished coding accuracy, which could affect the ability of the CCI score to predict mortality as a result of cerebral morbidity. However, to our knowledge, there is no evidence that the ICD-10 codes for cerebrovascular disease do not perform as well as the ICD-9 codes.
A final discrepancy between the ICD-9 and ICD-10 coding systems is related to our new instrument’s ability to improve specificity by introducing the 6 explicitly specified condition hierarchies. The designation of AIDS as a condition category and HIV-positive status as a lesser condition within that disease spectrum was a particularly significant contribution. The mean AIDS scores for the 3 populations were relatively close, but there was a slight decrease in score with the introduction of ICD-10. This decrease could be the result of an actual change in morbidity, but it could also be a result of the newness of ICD-10 to the healthcare system.
The phenomenon of “diagnosis-related group creep” is an example where the payment model created incentives for increasing coding intensity as providers became accustomed to the model. Given the financial incentives that payment models have implemented in more recent times (eg, episode-based payment for hospital stays and risk-adjusted payments to providers and health plans), a slight upward drift in CCI scores based on the new instrument may occur as familiarity with ICD-10 increases.
Within each population, the right-skewed distribution of CCI scores supports the use of the CDMF CCI instrument in triaging patients for possible disease- and case-management programs. The smooth, curvilinear relationship between CCI level and the risk for resource consumption or mortality, as well as the generally nonoverlapping CIs for estimates of these associations, suggest that individual CCI levels are associated with unique levels of risk.
The performance of this new instrument as a predictor of mortality over a longer time frame, which is consistent with the validation methods used by Charlson and colleagues,1 remains unknown and cannot be tested with ICD-10 data for some time. This remains an area for future study, and an opportunity to adjust the new instrument’s condition point values. Another useful next step would be to adapt this new instrument to one that could be administered either in a chart review setting or in patient triage, such as at a renal dialysis center.
The limitations of this study include those inherent to analyses based on claims data, including incorrect coding and missing data. Validation of the new coding scheme was based on prediction of same-year inpatient admissions and near-term mortality, whereas the original instrument was validated against long-term mortality.
Our ability to test association with mortality was hampered by the lack of long-term follow-up data in the ICD-10 era, as was previously noted.
Future studies should aim to validate this new scoring scheme with long-term mortality data and to compare its performance in long-term prediction with earlier algorithms. External validation of this new instrument in other populations with Medicare Advantage coverage and in traditional Medicare populations is required to confirm its general usefulness.
Our new, more granular, and clinically updated claims-based coding scheme for the CCI, using ICD-9 and ICD-10 versions, yielded condition-specific point values and overall scores that were consistent across 3 panels of individuals in a Medicare Advantage plan and were positively associated with same-year inpatient admissions as well as near-term mortality. The distribution of resulting CCI scores supports the use of the CDMF CCI scoring instrument in segmenting and prioritizing patients for case-management support. In addition, this new instrument allows for a more precise understanding of chronic disease at a population level, thus allowing health systems and health plans to design services and benefits to meet multifactorial clinical needs. This research sets the stage for further testing with long-term follow-up data and for adaptation of the CDMF coding scheme to a chart review instrument.
The authors wish to thank medical writer Teresa L. Rogstad, MPH, for her assistance in the preparation of the manuscript.
Author Disclosure Statement
Dr Glasheen, Mr Cordier, Mr Gumpina, Mr Haugh, Dr Davis, and Dr Renda were employees of Humana during the manuscript preparation.
- Charlson ME, Pompei P, Ales KL, MacKenzie CR. A new method of classifying prognostic comorbidity in longitudinal studies: development and validation. J Chronic Dis. 1987;40:373-383.
- D’Hoore W, Bouckaert A, Tilquin C. Practical considerations on the use of the Charlson comorbidity index with administrative data bases. J Clin Epidemiol. 1996;49:1429-1433.
- Deyo RA, Cherkin DC, Ciol MA. Adapting a clinical comorbidity index for use with ICD-9 -CM administrative databases. J Clin Epidemiol. 1992;45:613-619.
- Centers for Disease Control and Prevention. International Classification of Diseases, (ICD-10-CM/PCS) transition - background. www.cdc.gov/nchs/icd/icd10cm_pcs_background.htm. 2015.Prevention. Accessed May 2, 2018.
- Centers for Disease Control and Prevention. International Classification of Diseases, Tenth Revision, Clinical Modification (ICD-10-CM). www.cdc.gov/nchs/icd/icd10cm.htm. Accessed April 11, 2018.
- Canadian Institute for Health Information. Canadian Coding Standards for ICD-10-CA and CCI. https://secure.cihi.ca/estore/productSeries.htm?pc=PCC189. Accessed May 2, 2018.
- Quan H, Sundararajan V, Halfon P, et al. Coding algorithms for defining comorbidities in ICD-9 -CM and ICD-10 administrative data. Med Care. 2005;43:1130-1139.
- Romano PS, Roos LL, Jollis JG. Adapting a clinical comorbidity index for use with ICD-9 -CM administrative data: differing perspectives. J Clin Epidemiol. 1993;46:1075-1079; discussion 1081-1090.
- D’Hoore W, Sicotte C, Tilquin C. Risk adjustment in outcome assessment: the Charlson Comorbidity Index. Methods Inf Med. 1993;32:382-387.
- ICD10Data.com. 2019 ICD-10-CM codes. www.icd10data.com/ICD10CM/Codes. 2018. Accessed April 18, 2018.
- Zou G. A modified Poisson regression approach to prospective studies with binary data. Am J Epidemiol. 2004;159:702-706.
- Glasheen WP, Renda A, Dong Y. Diabetes Complications Severity Index (DCSI)—Update and ICD-10 translation. J Diabetes Complications. 2017;31:1007-1013.