Original Article
Comorbidity scores for administrative data benefited from adaptation to local coding and diagnostic practices

https://doi.org/10.1016/j.jclinepi.2011.04.004Get rights and content

Abstract

Objective

The Charlson and Elixhauser indices are the most commonly used comorbidity indices with risk prediction models using administrative data. Our objective was to compare the original Charlson index, a modified set of Charlson codes after advice from clinical coders, and a published modified Elixhauser index in predicting in-hospital mortality.

Study Design and Setting

Logistic regression using two separate years of administrative hospital data for all acute nonspecialist public hospitals in England.

Results

For all admissions combined, discrimination was similar for the Charlson index using the original codes and weights and the Charlson index using the original codes but England-calibrated weights (c = 0.73), although model fit was superior for the latter. The new Charlson codes improved discrimination (c = 0.76), model fit, and consistency of recording between admissions. The modified Elixhauser had the best performance (c = 0.80). For admissions for acute myocardial infarction and chronic obstructive pulmonary disease, the weights often differed, although the patterns were broadly similar.

Conclusion

Recalibration of the original Charlson index yielded only modest benefits overall. The modified Charlson codes and weights offer better fit and discrimination for English data over the original version. The modified Elixhauser performed best of all, but its weights were perhaps less consistent across the different patient groups considered here.

Introduction

What is new?

Key findings

  1. Despite being around 25 years old and derived in the United States, the original weights for the Charlson comorbidity index perform nearly as well as the weights recalibrated for current English data. Some extra codes used more in England explained more variation, especially at high-coding hospitals, and had higher consistency of recording across admissions for the same patient.

  2. The Elixhauser score performed best of all, but its weights varied more across patient groups.

What this adds to what was known
  1. The Charlson index needed some extra codes for use in England but still performed well, although the Elixhauser index did best.

What is the implication
  1. Comorbidity index codes and weights may need to be adapted to suit local patterns.

Patient comorbidity is a common potential confounder in health services research and can be estimated from administrative data sets. It is often included in risk models of mortality and other outcomes in provider profiling [1]. The two most commonly used indices are the Charlson [2] and Elixhauser [3], originally described using International Classification of Diseases, Ninth Revision (ICD-9) with US data. Both were designed for use with databases that cannot distinguish between comorbidities and complications and tried to limit their set of conditions to those of a chronic nature or a few acute conditions that are relatively unlikely to be potential complications. The Charlson index is now more than 20 years old, and both indices need calibrating on the data set of interest (external validation); to our knowledge, very little has been published from the United Kingdom. The weights (or scores) for these two indices may be inappropriate for the United Kingdom because of differing populations and coding practices. We have to date been using an Australian version of Charlson [1] in our risk adjustment models [4], [5], but discussions with clinical coders raised questions over the suitability of some codes when used in the United Kingdom. Recent work in Canada [6] found that a modified Elixhauser [7] outperforms Charlson; the authors also concluded that it needed external validation.

Our first aim was to derive new empirical weights based on English hospital administrative data for the original Charlson codes, a modified set of Charlson codes following discussions with coders, and the modified Elixhauser. Given that secondary diagnosis recording is known to vary between hospitals, our second aim was to assess the consistency of recording of codes used in the modified Charlson and modified Elixhauser across admissions belonging to the same patient.

Section snippets

Methods

We took 2 years of Secondary Uses Service data from Hospital Episodes Statistics (HES), the administrative data set that covers all admissions to National Health Service (NHS) (public) hospitals in England, covering the period 2007–08 to 2008–09. HES currently contains 20 diagnostic fields, coded using standard International Classification of Diseases, Tenth Revision (ICD-10). The first of these is the primary diagnosis or main problem treated, with the others available to capture comorbidities

Results

In 2008–09, there were a total of 5,432,220 inpatient admissions: 5,130,552 (94.5%) of these were to our list of acute nonspecialist hospital trusts (this proportion was similar in 2007–08). Of these admissions, 73.2% had one or more secondary diagnoses entered (70.9% in 2007–08). Table 1 shows the distribution for the 2 years of the number of comorbidities by each score. Compared with the original Charlson codes, the amended Charlson gives a nonzero comorbidity score to a further 2.1% of

Summary of results

In terms of discrimination and consistency of recording, the modified Charlson performed better than the original version but its discrimination was lower than the modified Elixhauser. The original Charlson codes and weights had similar discrimination as but inferior R-squared statistics to the modified set of codes and empirical weights. Model performance was much better for all inpatients combined than for either AMI or COPD admissions alone, but the relative performance of the comorbidity

Acknowledgments

Our unit is largely funded via a research grant by Dr Foster Intelligence, an independent health care information company and joint venture with the NHS Information Center. The Dr Foster Unit is affiliated with the Center for Patient Safety and Service Quality at Imperial College Healthcare NHS Trust, which is funded by the National Institute of Health Research. We are grateful for support from the National Institute for Health Research Biomedical Research Center funding scheme.

References (16)

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

Cited by (88)

  • Ranking sets of morbidities using hypergraph centrality

    2021, Journal of Biomedical Informatics
  • Temporal Comorbidity-Adjusted Risk of Emergency Readmission (T-CARER): A tool for comorbidity risk assessment

    2019, Applied Soft Computing Journal
    Citation Excerpt :

    Moreover, Mehta et al. [4] reported c-statistics of 0.70–0.76 for CCIs and ECIs using the Texas Medicare data (2006–2011). Furthermore, Bottle et al. [5,6] benchmarked CCIs using England’s HES data (2007–2009) and reported c-statistics of 0.57–0.79. Finally, Holman et al. [7] reported c-statistics of 0.61–0.77 for CCIs, ECIs, and the Multipurpose Australian Comorbidity Scoring System (MACSS) models based on data from hospitals in Western Australia (1989–1996).

  • Correlation between potentially inappropriate medication and Alzheimer's disease among the elderly

    2020, Archives of Gerontology and Geriatrics
    Citation Excerpt :

    In accordance with the 2012 version of the Beers list (Morgan et al., 2016), the proportion of all PIM was calculated as the fraction of the prescribed days with poor outcomes (numerator) divided by the total prescribed days of cardiovascular, neurological, or psychiatric diseases by well-trained clinical pharmacists (denominator). In this study, based on the literature review and clinical suggestion, control variables included gender, age, any of the 31 comorbidities listed in the Elixhauser Comorbidity (EC) index (Bottle & Aylin, 2011; Li, Kim, & Doshi, 2010), and the health status of participants one year before being diagnosed with AD. Health status was determined using the Adjusted Clinical Groups (ACG) system developed by Johns Hopkins Bloomberg School of Public Health.

View all citing articles on Scopus
View full text