Chapter 3 : Data sources
3.9 Elements of analyses
3.9.4 Comorbidities
To determine the effects of comorbidities on patient safety outcomes in this project, and to assess the validity of two commonly applied comorbidity measures, the Charlson
Comorbidity Index and the John Hopkins Adjusted Case Group (ACG) Case-Mix System were used. A further composite comorbidity measure based on Charlson Index disease groupings was applied. I now describe these three comorbidity measures.
3.9.4.1 Adjusted Case Group (ACG) System
The ACG case mix system was created at Johns Hopkins University in the US specifically for use in ambulatory (or non-acute) care and has been applied internationally, including in English general practice.216-218 This adjustment method takes into consideration the potential for patients to have multiple diagnoses over a set period of time and the ACG system can be used to predict healthcare use.219,220 Unlike other risk adjustment methods, the ACG system takes into account clinical need of patients and the burden of diseases when assigning patients to comorbidity groups. It has been used to assess comorbidities in the UK using GPRD data.218,221
The structure of the system is as follows:
1. Adjusted clinical groups (ACGs) – 106 mutually exclusive health status categories based on morbidity, age and sex. These are used for calculating costs.
2. The ACGs are used to assign patients to Resource Utilisation Bands (RUBs), indicating severity of morbidity. The six RUB groups are:
o 0 – No or only invalid diagnoses o 1 – Healthy user
o 2– Low o 3 – Moderate o 4 – High o 5 – Very high
3. Aggregated diagnosis groups (ADGs) – All ICD-10 (and Read codes) are categorised into 32 morbidity markers per patient. These unique morbidity groupings are based on “…specific clinical criteria and demand on healthcare services”.222 Patients are assigned to single or multiple ADGs. The ADGs can be aggregated into 12 Collapsed ADGs (CADGs). Due to copyright restrictions, the mapping of ADGs to CADGs could not be reproduced in this thesis. ADG assignment is based on 5 dimensions:
o Duration o Severity
o Diagnostic certainty o Type of etiology
o Expected need of specialty care
4. Expanded diagnosis clusters (EDCs) – There are 5 MEDC types (Administrative, Medical, Surgical, Obstetric/Gynaecological and Psychosocial). Within these types, there are 27 Major EDC (MEDCs) clinical categories/disease clusters. Each ICD/Read code is mapped to one of 267 EDCs. Within each EDC, the associated ICD and Read codes share similar diagnostic (and therapeutic) characteristics.
3.9.4.2 Use of ACG measures in this project
The ACG software was applied only to the GPRD dataset. Comorbidity measures were created for the entire dataset, with no distinction made between the three AE measures or measure-specific criteria. As such, the end date used to derive ACG weights for each patient was either date of death, date of transfer out of practice or the study end date, whichever occurred first. By ignoring end dates relevant to the individual AE measures, the derived ACG variables were not valid for use in all analyses. To explain, some conditions included in developing the ACG weights will have occurred after the outcome(s) of interest. To include these conditions in analyses where the response variable is the outcome of interest would bias the results. Where variables derived from the ACG software have been used, this is denoted in the relevant sections of the thesis. It should also be noted that original US
spellings are retained for ACG derived variables and intentionally used throughout this thesis.
3.9.4.3 ACG derivations
To derive the ACG weights, only Read codes in the medical history category of the GPRD dataset (“enttype=2”) were used as the quality of coding in other categories, such as
Disease Registries and Diabetes, was unknown and potentially inconsistent and/or poor. The ACG software distinguishes between data for patients aged <65 years (labelled by ACG as
"non-elderly") and patients aged ≥65 years (labelled "elderly"). Therefore the dataset was processed in two batches, using the "lenient diagnostic certainty" option which does not limit the number of diagnoses per patient included in processing, unlike the "stringent diagnostic certainty".223 Patients were assigned up to 32 ADGs, presented as binary flags.
These flags were aggregated into CADGs, with a third binary measure of MEDC flags. Counts of the number of EDCs per patient (maximum 267) were also included in analyses, and a fifth and final ACG measure of categories derived from RUB scores.
3.9.4.4 Charlson Index
The Charlson Index was originally developed for use in the hospital setting to predict mortality within one year of admission. The index has been extensively used in healthcare research, with up-to-date translations for the ICD-10 classification system.224 In the primary care setting, the index has been adapted for use with Read and OXMIS coding.225,226
The original Charlson Index and Khan et al's adaptation for Read/OXMIS codes were derived from 17 disease categories (Table 3.2). Khan et al’s version consists of 3,156 codes and was adapted from Deyo et al’s modification of the Charlson Index.226 No changes were made to the original specification. In line with Khan et al’s methodology, the overlaps in 13
Read/OMXIS codes corresponding to diabetes and peripheral vascular disease were coded as diabetes.226 Cancers were coded into separate groups, with exclusions for benign cancer (B7), cancer in situ (B8) and neoplasms of uncertain behaviour (B9).226
Table 3.2 Disease categories used to derive Charlson Index scores Charlson disease category Score weight
AIDS 6
Cancer 2
Cerebrovascular disease 1
Chronic pulmonary disease 1 Congestive heart disease 1
Dementia 1
Diabetes 1
Diabetes with complications 2
Hemiplegia 2
Metastatic tumour 6
Mild liver disease 1
Moderate liver disease 3
Myocardial infarction 1
Peptic ulcer disease 1
Peripheral vascular disease 1
Renal disease 2
Rheumatological disease 1
There is a relative dearth of studies comparing the performance of different versions of the Charlson Indices on UK primary care data. Studies in secondary care indicate that
Elixhauser’s comorbidity index performs better than Deyo et al’s.227-229 Ideally, I would have compared two commonly used adaptations of the Charlson Index - Deyo et al’s and
Elixhauser et al’s indices.230,231 The Elixhauser modified Charlson Index contains 30 disease categories, in contrast to the 17 disease groups in the original Charlson and Deyo et al’s adapted indices.230-232 Given the potentially cumbersome nature of analyses using 30
disease groups and issues surrounding small numbers, and the lack of evidence on using the Charlson Index on non-hospital data, only Deyo et al’s version of the Index was used in this project.
3.9.4.5 Disease group flags
In addition to the comorbidity measures created using the ACG software and the Charlson Index, I applied disease flags in the analyses. These binary flags corresponded to the 17
disease categories used to derive Charlson comorbidity scores for each patient. As
mentioned in section 3.9.4, an extra 18th (composite) flag was created to indicate whether a patient had any of the 17 diseases.