Comorbidity indices in observational studies on cardiovascular risk

Abstract

Objective

To analyse comorbidity measures in relation to cardiovascular disease risk, using data from Swedish healthcare registries. The aim was to evaluate the performance of different indices in predicting incidences of four outcomes: coronary heart disease (CHD), myocardial infarction (MI), heart failure (HF) and stroke.

Methods

Study population: All individuals in Sweden born between 1936 and 1975, with a look-back for diagnosis data 2010–2014 and followed-up for outcomes from 2024 to 2019, n=4 454 895 individuals. Two age groups: 40–64 and 65–79 years old were studied. We used the area under the receiver operating characteristic curves (AUROC) of age-and-sex, the Nordic Multimorbidity Index (NMI), a set of five cardiovascular risk factors as indicator variables (CV-IV) and explored measures based on inpatient care and numbers of filled prescriptions.

Results

In the age group 40–64 years, the AUROCs for age-and-sex alone were higher than those for the indices alone in all analyses: CHD (0.739); MI (0.736); HF (0.730) and stroke (0.685). CV-IV alone showed a higher AUROC for HF (0.694; 95% CI 0.690 to 0.697) than that for NMI (0.643; 95% CI 0.639 to 0.647), and for numbers of filled prescriptions (0.671; 95% CI 0.667 to 0.675).

Highest AUROC was observed for HF, when taking age-and-sex into account, for CV-IV (0.781; 95% CI 0.778 to 0.784) followed by numbers of filled prescriptions (0.776; 95% CI 0.773 to 0.780) and NMI (0.771; 95% CI 0.768 to 0.774). Furthermore, AUROC was higher for CV-IV, when taking age-and-sex into account, than that of age-and-sex alone for all outcomes in both age groups.

Conclusion

AUROC for age-and-sex alone was higher than that for any other single measure alone for all outcomes. The highest AUROC observed was for CV-IV adjusted for age-and-sex for HF. Thus, simple indicators measuring a few well-established cardiovascular risk factors outperformed a complex index such as the NMI. Similar results were obtained for numbers of filled prescriptions implying possible use as a proxy measure for comorbidity.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Several cardiovascular risk scores are available for early prediction and identification of cardiovascular risk in clinical practice. However, most risk models still operate within a single-disease paradigm, which can lead to an underestimation of risk in patients with comorbidities.

WHAT THIS STUDY ADDS

• A composite of five register variables representing diabetes, hypertension, dyslipidaemia, obesity and chronic obstructive pulmonary disease, as a surrogate for tobacco smoking, and numbers of filled prescriptions, respectively, outperformed a more complex Nordic Multimorbidity Index for 5 years risk for coronary heart disease, myocardial infarction, heart failure and stroke.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• Among the available scores and indices, it is important to identify the most appropriate measure or combination thereof for comorbidity adjustment that best aligns with the study population and data source. The number of prescriptions filled needs to be further explored and compared with the frequently used established cardiovascular scores.

Introduction

Several cardiovascular disease (CVD) risk scores are available for early prediction and identification of CVD risk factors for treatment in clinical practice.^{1 2} However, a recent scoping review found that most CVD risk models still operate within a single-disease paradigm, which can lead to an underestimation of risk in patients with comorbidities.³ In CVD studies, comorbidity indices that are used as covariates in regression models can aid in measuring the impact of coexisting conditions on patient outcomes, thereby improving risk stratification.³ The Nordic Multimorbidity Index (NMI), which is intended for use in registry-based observational studies, was recently developed in a general population setting using Danish health registry data. According to the authors, it outperformed several conventionally used indices in mortality prediction.⁴ With ageing populations and better survival rates, the prevalence of CVD with comorbidity increases, highlighting the need to integrate comorbidity measures to improve CVD risk stratification.⁵

The aim of our study was to analyse comorbidity measures in relation to CVD risk. We used the area under the receiver operating characteristic curves (AUROC) of age-and-sex alone, the NMI and a set of five well-established CVD risk factors as indicator variables (CV-IV). This was made as clinical data on lipids and blood pressure used in, for example, SCORE2 (1) or PREVENT (2) is not generally available in health registers, not in the Nordic health registers, nor in Sweden or in Denmark where the NMI was developed. We also explored measures based on inpatient care, that is, the number of hospitalisations, days hospitalised and the number of unique International Classification of Diseases, 10th Revision (ICD-10, three-character diagnoses as well as the number of filled drug prescriptions using the WHO Anatomical Therapeutic Chemical (ATC) Classification,⁶ that is, the unique ATC four-character level codes; ATC-4). This was made for four specified diagnoses as outcomes: coronary heart disease (CHD), myocardial infarction (MI), heart failure (HF) and stroke, as detailed below using Swedish healthcare registries.⁷ A further aim was to study possible differences when using several look-back periods of 1–5 years of health register data and a 1-year look-back for filled prescriptions from the Swedish Prescribed Drugs Register.

Study design and setting

Our study population comprised all individuals born from 1936 to 1975 who had continuously lived in Sweden between 2010 and 2014. The baseline population comprised: 4 454 895 individuals with a median age of 58 years (range 40–79), of whom 50% were women. These individuals were then monitored for the four outcomes for a further 5 years, from 2015 to 2019, that is, before the onset of the SARS-CoV-2 pandemic. All individuals with a previous history of the respective outcome under study were excluded at baseline.

Methods

We present AUROC with 95% CIs, calculated using 1, 3 and 5 years look-back periods of historical data for NMI, CV-IV, diagnoses in inpatient or specialised outpatient care and a 1-year look-back for filled prescriptions.

The NMI was calculated as described in the original article.⁴

The CV-IV represented the occurrence of the following diseases at baseline as dichotomous variables (1/0): diabetes (ICD-10 codes E10-E14 or ATC-code A10); hypertension (I10, I11, I15 or C02, C03, C07-C09); dyslipidaemia (E78 or C10); obesity (E66 or A08) and chronic obstructive pulmonary disease (COPD) (F17, J44 or N07BA).

The four outcomes evaluated were the 5 years incidences of: CHD (ICD-10: I20-I25); MI (I21); HF (I50); and stroke (I60-I64, I69), respectively.

This was done in three models per outcome:

1. Age-and-sex alone.

2. Each index and explored measures alone.

3. Each index and explored measures taking age-and-sex into account.

Results

Baseline prevalences in this sample from 2010 to 2014 were as follows: diabetes (7.4%); hypertension (31.3%); dyslipidaemia (15.7%); obesity (2.0%); and COPD (2.2%). The three most frequent filled prescriptions (18.7%) at baseline were CVD-preventive medicines (specific ATC-codes: C07, beta-blockers; C08, calcium channel blockers; and C09, agents acting on the renin–angiotensin system).

Figure 1 presents the AUROC for 5-year follow-up on HF with a 5-year look-back period, for those aged 40–64 years and 65–79 years. In the younger age group, the highest AUROCs were observed for HF when taking age-and-sex into account in analyses of respective index and the highest AUROCs observed was the one for the CV-IV (0.781; 95% CI 0.778 to 0.784) followed by those for the numbers of filled prescriptions; ATC-4 (0.776; 95% CI 0.773 to 0.780) and the NMI (0.771; 95% CI 0.768 to 0.774). When not taking age-and-sex into account, the CV-IV alone showed a higher AUROC for HF (0.694; 95% CI 0.690 to 0.697) than that for the NMI alone (0.643; 95% CI 0.639 to 0.647), and for numbers of filled prescriptions alone, ATC-4 (0.671; 95% CI 0.667 to 0.675). For age-and-sex alone, AUROC was higher than those for the indices alone (0.730; 95% CI 0.727 to 0.733). In the higher age group, AUROC were generally lower than those observed in the younger age group and age-and-sex showed lower AUROC than those for the NMI, CV-IV and ATC-4.

Figure 1. AUROC characteristics with 95% CIs for 5-year follow-up for heart failure (I50) in the general Swedish population, divided into two groups aged 40–64 and 65–79 years, who were free from heart failure at baseline. From top to bottom, for the NMI and a set of five CV-indicator variables representing the presence or not of: diabetes (ICD-10 code E10-E14 or ATC-code A10); hypertension (I10, I11, I15 or C02, C03, C07–C09); dyslipidaemia (E78 or C10); obesity (E66 or A08) and COPD/nicotine replacement drugs against smoking (F17, J44 or N07BA). Further, numbers of distinct three-character ICD-10 codes for the main diagnoses, numbers of hospitalisations (N-Visits), numbers of days hospitalised for inpatient care (N-Days) and numbers of distinct four-character ATC-codes on filled prescriptions with a 1-year look-back period (as in the NMI). The models were based on (1) age-and-sex alone (red symbols), (2) the index or measure alone (light blue symbols) and (3) the full model, that is, age-and-sex and the index (dark blue symbols). ATC, Anatomical Therapeutic Chemical; AUROC, area under the receiver operating characteristic curves; COPD, chronic obstructive pulmonary disease; CV, cardiovascular; ICD-10, International Classification of Diseases, 10th Revision; NMI, Nordic Multimorbidity Index.

Figure 2 shows the AUROC for 5-year follow-up on HF with a 5-year look-back period, for NMI and CV-IV for those aged 40–64 years and those 65–79 years. The CV-IV combined with age-and-sex rendered an AUROC of 0.781 (95% CI 0.778 to 0.784) compared with 0.771 (95% CI 0.768 to 0.774) for NMI and 0.730 (95% CI 0.727 to 0.733) for age-and-sex alone for those aged 40–79 years and 0.719 (95% CI 0.717 to 0.721), 0.705 (95% CI 0.703 to 0.706) and 0.651 (95% CI 0.649 to 0.653), respectively, for those aged 65–79 years.

Figure 2. AUROC characteristics for 5-year follow-up on heart failure with a 5- year look-back period in the general population, aged 40–64 years (upper panel) and 65–79 years (lower panel) who were free from heart failure at baseline. The left panel shows the Nordic Multimorbidity Index and the right panel shows a set of five CV-indicator variables representing the presence or absence of: diabetes (ICD-10 code E10-E14 or ATC-code A10); hypertension (I10, I11, I14 or C02, C03, C07–C09); dyslipidaemia (E78 or C10); obesity (E66.0 or A08) and COPD/smoking (J44, F17 or N07 BA). The models were based on (1) age-and-sex alone (red line), (2) the index or measure alone (light blue line) and (3) the full model, that is, age-and-sex and the index (dark blue line). ATC, Anatomical Therapeutic Chemical; AUROC, area under the receiver operating characteristic curves; COPD, chronic obstructive pulmonary disease; CV, cardiovascular; ICD-10, International Classification of Diseases, 10th Revision.

In figure 3 and online supplemental appendix table A1 to A4, we present AUROC, for the varying look-back periods, 2010–2014, for 5-year follow-up on the respective four outcomes, for those aged 40–64 years at baseline. For age-and-sex alone, AUROCs were higher than those for the indices alone in all analyses: CHD (0.739; 95% CI 0.737 to 0.741); MI (0.736; 95% CI 0.733 to 0.739); HF (0.730; 95% CI 0.727 to 0.733) and stroke (0.685; 95% CI 0.682 to 0.688). An increasing trend in AUROC was observed for all outcomes with an increasing look-back period for indices and explored measures of inpatient care alone, see figure 1. This trend was most pronounced for HF.

Figure 3. AUROC characteristics with 95% CIs for 5-year follow-up in the general population, aged 40–64 years who were free from the respective outcome at baseline. From left to right, coronary heart disease (I20–I25), myocardial infarction (I21), heart failure (I50) and stroke (I60–I64, I69) with a 1-year, 3-year and 5-year look-back period for diagnoses, from top to bottom, for the NMI and a set of five CV-indicator variables representing the presence or not of: diabetes (ICD-10 code E10-E14 or ATC-code A10); hypertension (I10, I11, I15 or C02, C03, C07–C09); dyslipidaemia (E78 or C10); obesity (E66 or A08) and COPD/smoking (F17, J44 or N07BA). Further, numbers of distinct three-character ICD-10 codes for the main diagnoses, numbers of hospitalisations (visits), numbers of days hospitalised for inpatient care and numbers of distinct four-character ATC-codes on filled prescriptions with a 1-year look-back period. The models were based on (1) age-and-sex alone where the red line in each panel represents the AUROC for age-and-sex, (2) the index or measure alone (light blue symbols) and (3) the full model, that is, age-and-sex and the index (dark blue symbols). ATC, Anatomical Therapeutic Chemical; AUROC, area under the receiver operating characteristic curves; COPD, chronic obstructive pulmonary disease; CV, cardiovascular; CVD, cardiovascular disease; HF, heart failure; ICD-10, International Classification of Diseases, 10th Revision; MI, myocardial infarction; NMI, Nordic Multimorbidity Index.

In Figure 4 and online supplemental appendix table A5 to A8, we present corresponding results to figure 3 for those aged 65–79 years at baseline. Similar patterns as in the younger age group were observed, however, with lower AUROC values.

Figure 4. AUROC characteristics with 95% CIs for 5-year follow-up in the general population, aged 65–79 years who were free from the respective outcome at baseline. From left to right, coronary heart disease (I20–I25), myocardial infarction (I21), heart failure (I50) and stroke (I60–I64, I69) with a 1-year, 3-year and 5-year look-back period for diagnoses, from top to bottom, for the NMI and a set of five CV-indicator variables representing the presence or not of: diabetes (ICD-10 code E10-E14 or ATC-code A10); hypertension (I10, I11, I15 or C02, C03, C07–C09); dyslipidaemia (E78 or C10); obesity (E66 or A08) and COPD/smoking (F17, J44 or N07BA). Further, numbers of distinct three-character ICD-10 codes for the main diagnoses, numbers of hospitalisations (visits), numbers of days hospitalised for inpatient care and numbers of distinct four-character ATC-codes on filled prescriptions with a 1-year look-back period. The models were based on (1) age-and-sex alone where the red line in each panel represents the AUROC for age-and-sex, (2) the index or measure alone (light blue symbols) and (3) the full model, that is, age-and-sex and the index (dark blue symbols). ATC, Anatomical Therapeutic Chemical; AUROC, area under the receiver operating characteristic curves; COPD, chronic obstructive pulmonary disease; CV, cardiovascular; CVD, cardiovascular disease; HF, heart failure; ICD-10, International Classification of Diseases, 10th Revision; MI, myocardial infarction; NMI, Nordic Multimorbidity Index.

Numerical results for all analyses performed are those shown in the online supplemental appendix tables A1 to A8. Outcome incidence rates are shown in the online supplemental appendix table A9.

Discussion

For discussion, AUROC were higher for CV-IV and numbers of filled prescriptions, that is, ATC-4 when taking age-and-sex into account than that of age-and-sex alone for all outcomes for both age groups studied, which was not the case for the other indices analysed, as seen in figures3 4. The highest AUROC, when taking age-and-sex into account, was observed for the CV-IV and for filled prescriptions, that is, ATC-4 for HF in the younger age group, and with a trend for higher AUROC the longer the look-back period. While a 1-year, 3-year or even a 5-year look-back period may not accurately reflect the causal period of interest in the development of HF, filled prescriptions and/or comorbidities during these assessment periods may represent conditions of longer duration that are not captured in inpatient register data. Filled prescriptions capture the frequently prescribed CVD-preventive medicines. The observed association with filled prescriptions may thus be considered as a marker for CVD risk. However, the important exogenous CV-risk factor of tobacco smoking is not captured in our health register data. We used COPD combined with medications used in nicotine dependence as a proxy measure, realising however, that it will not fully capture smoking habits. Tobacco smoking and its impact on the progressive atherosclerotic process ultimately leading to HF development is thus not directly possible to detect in the health registries used.

Interestingly, using a 1-year look-back period, the number of filled prescriptions alone rendered AUROC for HF of a similar magnitude to that of CV-IV alone when 5 years of look-back data were used. This suggests that simply calculating the number of unique prescriptions filled could possibly be used as a proxy for disease burden instead of complex indices. This could be important in situations where clinical background data is limited. In such cases, using pharmacy dispensing data to assess comorbidity could be useful.

A recent study in 46–60 years old participants, using deep machine learning with a deep learning model incorporating 11 features outperformed traditional models in predicting a composite of CVDs with an observed AUROC of 0.764.⁸ For comparison, we observed an AUROC of 0.771 for the CV-IV and CHD in the age group of 40–64 years. However, HF was not analysed separately in the study using the deep learning model. We observed an AUROC of 0.781 for the number of prescriptions filled, adjusted for age-and-sex for HF. This rather simple measure requires further comparison to the more complex deep learning model in relation to HF and HF subgroups as HF is a heterogenous disease with distinct subgroups.⁹

Here, we excluded the respective prevalent outcomes at baseline and 5 years prior to explore comorbidity measures in relation to incidence of four CV outcomes of interest from a public health perspective. Interestingly, AUROC observed were in general lower in the older age group. This may be seen as counterintuitive at first glance. However, comorbidity is more prevalent in the older age group, and in the younger age group, the vast majority is healthier on average. Further, disease severity is not captured in the register data. Thus, complex indices that do not capture disease severity are less discriminatory when used in studies in higher age.

The NMI was developed in a Danish population aged 40 years and over. Here, we investigated two age groups: 40–64 and 65–79 years to assess possible differences, given that comorbidities impact differently at younger and older ages. This study had high statistical power, as can be seen from the very narrow CIs presented in the online supplemental appendix tables 1 to 8. Analysis of the 40–64 years age group (acting age of work-life) revealed the highest AUROC for HF, underlining the importance of effective risk-reduction in middle age, that is, the primary prevention treatment of underlying CVD risk factors including refraining from tobacco smoking, to prevent the development of CVDs and HF.

From a clinical CVD treatment perspective, comorbidities might influence outcomes such as survival, treatment tolerance and risk of complications for the individual. Taking comorbidity into account is valuable when tailoring treatment plans and stratifying patients in clinical trials.

The number of prescriptions filled needs to be further explored in pharmacoepidemiological studies and to be compared with the frequently used established CV-scores in studies where clinical laboratory data are available as needed for comparison with, for example, the clinical risk tools SCORE2 (1) and PREVENT equations (2). Among the available indices, it is important to identify the most appropriate measure or combination thereof for comorbidity adjustment that best aligns with the study population, data source and research question.

Conclusion

In conclusion, AUROC for age-and-sex alone was higher than that for any other single measure alone for all outcomes in the 40–64 years age group, highlighting the importance of always including age and sex. Furthermore, AUROC were higher for CV-IV, when taking age-and-sex into account, than that of age-and-sex alone for all outcomes for both age groups studied. Thus, simple indicators measuring a few well-established CVD risk factors outperformed the other indices. Similar results were obtained for filled prescriptions in the last year, implying a possible use of filled prescriptions as a proxy measure for comorbidity. The observed trend of increased AUROC with an increasing look-back period, most pronounced for the explored measures, shows the importance of the duration of comorbidities as risk factors for CVD, especially for HF. This should be further explored and validated in other settings.

Data availability statement

Data may be obtained from a third party and are not publicly available.