Effect of timeframes to define long term conditions and sociodemographic factors on prevalence of multimorbidity using disease code frequency in primary care electronic health records: retrospective study

Data source

We used data provided from the Clinical Practice Research Datalink Aurum dataset, which includes routinely collected, patient level data from general practices using EMIS health web software in England.11 We included all patients who were deemed research acceptable (a marker used by the Clinical Practice Research Datalink for patients meeting specific quality criteria12) registered to a general practice on 1 January 2020 and aged 18 years or older with at least one year of continuous practice registration. Observation data for each patient are recorded in the form of Medcodes, along with the date of observation and entry date. Any observation codes occurring after the 1 January 2020 were excluded and we included all historical observation codes. Data were linked to the 2019 Index of Multiple Deprivation of a patient’s area of residence (categorised into tenths), which was used as a patient level marker of socioeconomic deprivation.13

Diagnostic codes and data cleaning

Multimorbidity was defined as the presence of two or more chronic conditions from a set of 212 conditions.4 We included 211 conditions selected by Head and colleagues, based on a set of 308 disease phenotypes developed by Kuan and colleagues.3 14 We reviewed these code lists and made changes for diabetes disease lists (online supplemental appendix) and also included "chronic primary pain" as a new condition. In the study by Head and colleagues, most conditions were defined by a diagnostic code ever being recorded, but for some conditions, the authors required three occurrences of a code within a 12 month period. Two clinicians (TB and JC) reviewed the codes and reassigned 41 diseases to require multiple codes where a single code might otherwise indicate an acute condition (see online supplemental appendix for further details and rationale). For diabetes, we assigned the most likely diabetes type (type 1, type 2, or other/unspecified) based on the frequency of codes. Further information on the long term condition categories and cleaning of ethnic group codes are given in the online supplemental appendix. Our code lists are available from: https://tbeaney.github.io/MMclustering/

Disease definitions

The number of chronic conditions for each person was calculated as of the 1 January 2020 under each of the following five definitions (figure 2). These definitions were selected to reflect the variety used in previous multimorbidity research and definitions of chronic conditions:

1. Any disease code in the record for all 212 conditions, which was used as the reference definition (single code);14

2. two codes at least three months apart for the 41 conditions requiring multiple codes,7 or a single code for the remaining 171 conditions;

3. two codes at least 12 months apart for the 41 conditions requiring multiple codes,7 8 or a single code for the remaining 171 conditions;

4. three codes within any 12 month period for the 41 conditions requiring multiple codes,3 or a single code for the remaining 171 conditions; and

5. any code appearing within the past 12 months for the 41 conditions requiring multiple codes,9 or a single code for the remaining 171 conditions.

Request permission

We also compared a definition of multimorbidity only including conditions defined as active problems (referred to here as problems). Diagnostic codes entered into the record can be marked by clinicians as problems, which then appear on a patient’s summary during a consultation.11 These may be more visible to the clinician than historical diagnostic codes, which would require manual searching through the electronic notes and are also pulled across onto referral letters.10 Problems are recorded as either active or past problems and can be marked as ended if no longer active, with the end date recorded.11 We included only those problems that were active, that is, those that did not have an associated end date before 1 January 2020 and were marked as active. The problem definition was applied to all 212 long term conditions and not only the 41 requiring multiple codes. Further details are given in the online supplemental appendix page 10.

Statistical analysis

We calculated the prevalence of patients with one, two, or three or more chronic conditions based on each definition, stratified by age group, gender, ethnic group, and socioeconomic deprivation. Age (years) was categorised into 18-29, 30-39, 40-49, 50-59, 60-69, 70-79, and 80 or more. For gender, we used the categories defined in Clinical Practice Research Datalink, as male, female, or indeterminate (where gender cannot be classified as male or female). No data for age or gender were missing due to our inclusion criteria and Clinical Practice Research Datalink's definition of research acceptable patients. In all analyses, missing data were assigned as categories to ensure no patients were excluded.

To calculate associations with the expected probability of a change in having multimorbidity under each definition, we used mixed effects logistic regression. We included only patients with multimorbidity defined using the single code definition and created a binary outcome variable as 1 if multimorbidity status changed and 0 if multimorbidity status did not change, compared with the single code definition. We included age, gender, ethnic group, and Index of Multiple Deprivation as covariates in the models and random intercepts for general practice to account for clustering and potential coding differences between practices. We then calculated the marginal probability of a change in multimorbidity status overall, and separately for each covariate, holding other covariates at their mean values and assuming random effects of zero.

To calculate associations with the change in expected total count of conditions under each definition compared with the single code definition, we used mixed effects negative binomial regression. A negative binomial model was chosen because the conditional variance was found to be significantly larger than the conditional mean for each model (P<0.001) making a Poisson model unsuitable.15 Models included the same covariates and random effects as in the mixed effects logistic regression models. We then calculated the marginal expected change in the count of long term conditions overall and separately for each covariate, assuming other covariates in the model were held at their mean values and random effects of zero. All regression models assumed the single code definition as the reference group, given that this method will always result in the highest estimated prevalence. Further details, including equations for both models, are given in the online supplemental appendix.

Python version 3.10.6 and Pandas version 1.4.3 were used in data processing and plots and Stata version 17.0 was used for regression analyses.

Multimorbidity prevalence

The prevalence of people with one or more, two or more, or three or more long term conditions varied substantially according to definition used (table 2). Under the single code definition, the prevalence of multimorbidity (two or more long term conditions) was 73.9%, reducing to 41.4% using the three codes within 12 months’ definition and 35.2% using only conditions recorded as active in the problem list.

Of the 41 conditions included that were required to have multiple codes, elevated concentrations of total cholesterol, triglycerides, and low density lipoprotein; dermatitis; and enthesopathy and synovial disorders were the most common conditions, with a prevalence of at least 20% using the single code definition (online supplemental appendix figure A1). The prevalence of most of these conditions was substantially lower with each of the alternative definitions from two codes over three months to two codes over 12 months to three codes within 12 months, and any code in the past 12 months. However, reductions in the prevalence were not consistent between conditions, as evidenced by the prevalence of obesity, which was higher under the any in past 12 months definition compared with the three within 12 months definition.

For problem codes, the pattern was mixed according to condition. Hypertension, type 2 diabetes, thyroid disease, chronic kidney disease, atrial fibrillation, and chronic obstructive pulmonary disease had a similar prevalence applying the problems definition compared with using only the single code definition (online supplemental figure A2). However, those diagnoses based only on blood test results (raised concentrations of cholesterol or triglycerides) were represented rarely in the problem table (prevalence <0.1%; Online supplemental figure A2).

The prevalence of multimorbidity under each definition stratified by demographic group is shown in figure 3 (percentages are given in online supplemental table A2). Within all sociodemographic strata, multimorbidity prevalence was lowest applying the problems definition, from 12.1% in the 18-29 year age group to 87.8% in the 80 years and older age group, compared with 48.5% and 98.1%, respectively, applying a single code definition. Comparing the definitions using alternative timeframes, the absolute difference in prevalence between definitions was large for younger age groups but became smaller with increasing age. Within categories of gender, ethnic group, and Index of Multiple Deprivation group, absolute differences in multimorbidity prevalence between definitions remained similar in size (figure 3).

Request permission

Demographic differences in multimorbidity status according to disease timeframe

In people with multimorbidity based on the single code definition, we compared the expected probability of a change in multimorbidity status under each different timeframe, holding age, gender, ethnic group and Index of Multiple Deprivation at their mean values. The expected probabilities by definition of a person being reclassified as not multimorbid (compared with the single code definition) were 18.7% for two codes over three months, 21.9% for two codes over 12 months, 40.4% for three codes within 12 months, 37.5% for any code in the past 12 months, and 51.6% for problems.

Older people were substantially less likely to be reclassified as not multimorbid compared with younger people after controlling for gender, ethnicity and deprivation, under each of the four alternative definitions compared with a single code definition (figure 4 and online supplemental table A3). People aged 80 years or older had only a 1-10% probability of being reclassified under the different definitions, compared with those aged 18-29 years, who had a 50-75% range of being reclassified. Within each definition, the expected probabilities of a change in multimorbidity status were similar for men and women. People of South Asian ethnicity had the lowest expected probability of a change in multimorbidity status, under all alternative definitions, except for the three codes within 12 months and problems definition, where the probability was lowest for people of white ethnicity. People of other, or unknown ethnic groups had the highest expected probabilities of no longer being multimorbid under all alternative definitions. The data showed a strong gradient towards a lower chance of being reclassified as not multimorbid with increasing deprivation across all alternative definitions.

Request permission

Demographic differences in the number of chronic conditions according to disease timeframe

We used mixed effects multivariable negative binomial models to estimate the expected reduction in the number of long term conditions under each definition compared with the single code definition, holding age, gender, ethnic group, and Index of Multiple Deprivation at their mean values. The mean expected reduction in the number of long term conditions from the single code definition was 1.9 for two codes over three months, 2.1 for two codes over 12 months, 3.1 for three codes within 12 months, 3.0 for any code in the past 12 months, and 3.6 for problems.

Younger people had a smaller expected reduction in the number of long term conditions than older people, after controlling for gender, ethnicity and deprivation, relative to the single code definition. This finding was apparent across all alternative definitions but was more pronounced in the problems, three within 12 months and any in past 12 months definitions, with people of 18-29 years expected to have 2.0 conditions fewer by each definition, compared with people 80 years or older expected to have 6.4, 3.9 and 4.0 conditions fewer, respectively, when compared with the single code definition (online supplemental appendix figure A3 and table A4). Men had smaller expected reductions in the count of conditions compared with women across all definitions. People of South Asian ethnicity had the largest expected reduction in the number of long term conditions, while people of black, other, or unknown ethnic groups had smaller reductions in the expected count. Data were trending towards larger expected reduction in counts of long term conditions in people living in areas of higher deprivation.

Strengths and limitations

A strength of our study is the use of large sample of adults registered to primary care in England, with previous studies of the Clinical Practice Research Datalink Aurum data finding it to be representative of the national population.11 We adopted a broad set of disease codes, many of which have particular relevance to primary care, and applied different definitions of long term condition timeframe using the same set of conditions to the same patient cohort. To an extent, these definitions are arbitrary, and our analysis presented a range of approaches used in the literature before, rather than attempting to analyse all possible definitions. For example, a condition lasting more than six months is sometimes used to define a long term condition,6 but we did not include this definition because estimates would lie between our range of three and 12 months.

A limitation of using routinely collected healthcare data is that some conditions may not be recorded, either because a person does not present with a condition or because when presenting, the condition is not coded by a clinician. The likelihood of missing codes is unlikely to be random with respect to either patients or diseases. Although a previous systematic review found good agreement in diseases recorded in Clinical Practice Research Datalink with other sources,23 comparison against cancer diagnoses in the Clinical Practice Research Datalink with cancer registry data found a range of 9–26% for different cancers were missing in the Clinical Practice Research Datalink.24 The financial incentives offered by quality and outcomes framework have improved data collection and coding in primary care22 and may therefore inflate the prevalence of long term conditions requiring multiple codes for conditions included in the framework relative to other conditions.

The code lists used for raised total cholesterol, raised low density lipoprotein cholesterol, low high density lipoprotein cholesterol and raised triglycerides available from the CALIBER study include test results (rather than diagnostic codes), which were rarely coded as problems, partially explaining the substantially lower prevalence of multimorbidity overall. Studies using problem codes alone would therefore need to include diagnostic codes, but was not feasible for comparison in our study as the granularity required in low density lipoprotein cholesterol and high density lipoprotein cholesterol measurements cannot be accounted for using diagnostic codes alone. A further issue of using problem codes is in defining active problems. In Clinical Practice Research Datalink Aurum no date is recorded at which a problem is changed from active to inactive. Our data extraction occurred in May 2022, and so some codes marked as inactive were likely active at the study start date, and so our findings may represent an underestimate. However, given our focus on chronic conditions, the number of conditions to have resolved is unlikely to be large.

We used a large set of chronic conditions, which will have unequal burden on clinical outcomes, healthcare use, and quality of life. Many may not be included in other multimorbidity measures, with a recent Delphi study identifying a core set of 24 conditions to always include and 35 to usually include and with no conditions identified for exclusion.25 A smaller and standardised set of conditions can aid comparability and reproducibility but has drawbacks. Firstly, a smaller set limits the scope to detect novel associations between less common conditions. Secondly, choices over conditions are subjective and those that are less frequent can still have a large burden on individuals.

The overlap between some of the long term conditions included here may lead to double counting and over-estimation of multimorbidity, for example, combining myocardial infarction and angina as ischaemic heart disease. However, when investigating the cause of disease, more granular categories may be a benefit because a person with a myocardial infarction who subsequently develops angina may have a different trajectory and opportunities for intervention and prevention to a person with angina who subsequently develops a myocardial infarction, despite a common pathophysiology.

Combining diagnostic codes with relevant medications or treatments may help to distinguish active versus inactive conditions; for example, use of a proton pump inhibitor in gastritis. Similarly, the quality and outcomes framework excludes patients from the asthma register if no asthma related drugs were prescribed in the past 12 months.26 However, a difficulty with this approach for some diseases is that drugs can have multiple indications; for example, proton pump inhibitors being co-prescribed with non-steroidal anti-inflammatory medications, which may be more common in those with a history of gastro-intestinal problems irrespective of active symptoms.27

Implications

Despite the consensus in the medical literature on the definition of multimorbidity as the co-occurrence of two or more chronic conditions,4 which conditions should be included or how to determine chronicity has not been agreed.5 25 Our findings highlight that even when using the same set of medical codes, decisions on how to define a long term condition can change the prevalence of multimorbidity almost twofold, which has important implications for the direct comparison of estimates between studies. Results also suggest that a universally agreed metric of multimorbidity prevalence may be an unrealistic target and that estimates are highly context dependent. Rather than seeking one rigid definition of multimorbidity, we believe that researchers should instead embrace the variety of approaches, better reflecting the variety of lived experiences of people having multimorbidity. As such, we do not advocate only one of our approaches as the best, with choice dependent in part on the aims of the research. Nevertheless, some approaches may be more suitable in particular contexts, while the impact of bias, and exclusion of specific groups from multimorbidity measures, highlighted in our work, should be assessed (Box 1).

Researchers should consider which conditions to include, and may opt for a narrower set of diseases by inclusion of only conditions for which a single code would indicate chronic risk, rather than those that may also represent acute conditions. Where inclusion of a greater breadth of diseases is preferred, researchers should decide whether some of these require presence of multiple codes over time. For studies focused on cause or accumulation of diseases over time, use of any diagnostic code in the record may be the preferred approach, as a disease, even if not currently active and only recorded once, may be relevant to subsequent disease development. However, this approach may lead to inclusion of acute or inactive conditions, inflating the prevalence of multimorbidity. For studies focused on associations between current disease burden and healthcare service use and treatment, use of more contemporaneous active diseases may be preferable. For example, use of codes from the past 12 months, or by incorporating prescription data where possible for some conditions. However, any approach that uses multiple codes may be at greater risk of bias in the frequency of coding of conditions, which may depend on factors related to the patient, clinician, general practice, and coding incentives.21 Therefore, we suggest that use of a single code definition should always be considered as a sensitivity analysis to understand the effect on prevalence and differential impact between patients.