Data and participants
We analysed data from Evaluation of the Methods and Management of Acute Coronary Events, EMMACE-3 and EMMACE-4, which are multicentre longitudinal national cohort studies of outcomes following MI combining survey data with national clinical registration data (ClinicalTrials.gov NCT01808027 and NCT01819103) . The study included patients aged 18 years and over who were hospitalized with MI, defined by the third universal definition as either ST-elevation myocardial infarction (STEMI) or non-STEMI (NSTEMI) . Participants were recruited from 77 National Health Service (NHS) hospitals in England between November 1, 2011, and June 24, 2015. They were consented to participate by a trained researcher during their hospital admission (data flow is shown in Additional file 1: Figure S1).
Patients at a terminal stage of any illness, and those for whom follow-up would be inappropriate or impractical, were excluded. Consenting patients were asked to complete a self-administered questionnaire at the time of enrolment in hospital, and at 1, 6, and 12 months following discharge from hospital. This included information about HRQoL assessed using the three level EuroQol 5-dimension (EQ-5D-3L) instrument . For non-responders who were alive and who had not withdrawn from the study, repeat questionnaires were sent by post on up to three occasions before the date of the next follow-up contact. Data for consenting patients were linked to the national clinical register of MI admissions in the UK (Myocardial Ischaemia National Audit Project, MINAP ) to gather information about patients medical history including presence of hypertension, diabetes mellitus, angina, asthma or chronic obstructive pulmonary disease (COPD), cerebrovascular disease (CVSD), peripheral vascular disease (PVD), chronic heart failure, chronic renal failure, type of MI (NSTEMI or STEMI), and in-hospital as well as post-discharge treatments and medications.
Health-related quality of life
The outcome of this study was HRQoL assessed using self-reported EQ-5D-3L . This contains two subscales: a descriptive system (EQ-5D) and a visual analogue scale (EQ-VAS). EQ-5D comprises five dimensions: mobility, self-care, usual activities, pain/discomfort, and anxiety/depression. Each domain has three levels (3L): no problems, some problems, and extreme problems. The EQ-5D-3L dimensions data may be summarised as a single index score ranging from −0.5 to 1, with scores less than 0 indicating states ‘worse than death’, 0 indicating no quality of life, or ‘death’, and 1 indicating full health and therefore no problems in any domain. The index score was standardised to the UK population . The EQ-VAS score ranges from 0 to 100, with 0 denoting the worst and 100 the best, health state imaginable. The EQ-5D questionnaire has previously been validated in patients following MI . A difference in score of 7 for VAS and 0.05 for EQ-5D score are regarded as clinically important , and these thresholds were used to define a clinically important change between subgroups.
The exposure was multimorbidity clusters based on 7 pre-existing long-term health conditions recorded in the MINAP registry including hypertension, diabetes mellitus, asthma or chronic obstructive pulmonary disease (COPD), cerebrovascular disease (CVSD), peripheral vascular disease (PVD), chronic heart failure, and chronic renal failure. In the data source used for the study, the Myocardial Ischaemia National Audit Project (MINAP)  registry information on multi-morbidities is given as binary variables. No further detail is given beyond this; therefore, we were restricted to use the information as recorded.
Sociodemographic, health characteristics, and clinical variables included age, sex, ethnicity (white vs other), smoking status (never vs ex or current), body mass index (BMI) (kg/m2), past medical history of MI, angina, diagnosis (STEMI or NSTEMI), revascularisation (percutaneous coronary intervention [PCI] vs. no PCI; coronary artery bypass graft [CABG] surgery vs no CABG surgery), medications (aspirin, β blockers, statins, and ACE inhibitors), and referral for cardiac rehabilitation (yes/no).
Patient characteristics according to multimorbidity clusters were described using frequencies and percentages for categorical data and for continuous data as means and standard deviation. Chi-square test and ANOVA were used to assess univariate associations between categorical and continuous patients’ characteristics and multimorbidity clusters, respectively. We corrected for multiple testing in the tables using the Hochberg correction, using a false discovery rate of 0.05.
Latent class analysis (LCA)  using Mplus software version 8 was used to identify clusters of multimorbidity for 7 pre-existing long-term health conditions recorded in the MINAP registry at hospital admission including hypertension, diabetes mellitus, asthma or chronic obstructive pulmonary disease (COPD), cerebrovascular disease (CVSD), peripheral vascular disease (PVD), chronic heart failure, and chronic renal failure.
Latent class analysis (LCA) is a statistical technique used to determine subgroups within populations that share certain outward characteristics . In LCA, class membership is based on probability of belonging to a class membership given the pattern of responses they have on indicator variables, and there are no clear cut assignments. LCA is a “person-centred” approach of deriving typologies, unlike “variable-centred” tradition that uses arbitrary cut-offs for classifying individual cases . We used LCA instead of cluster analysis  because unlike cluster analysis or k-means clustering, LCA is model-based and an evaluation of how well a proposed LCA model represents the data can be conducted  using Bayesian information criterion (BIC) , Akaike’s information criteria (AIC) , and Bootstrap likelihood ratio test (BLRT)  p values. We fitted several latent class models varying the number of classes up to 5 classes to identify the best class solution with 1000 random starting values each with 100 iterations. Bootstrap p values based on 500 replications were used to assess model fit. The model goodness of fit statistics, entropy, classification matrix, class frequencies, and class conditional probabilities is reported in Additional file 1: Table S1, Table S2, and Table S3. The utility of using LCA has been demonstrated by other researchers who used it to identify clinical phenotypes with differential treatment responses  and patient outcomes [4, 38, 39].
The optimal LCA model was selected based on Bayesian information criterion (BIC), Akaike’s information criteria (AIC), Bootstrap likelihood ratio test (BLRT) p values, and clinical interpretation. For the BIC and AIC the optimal model is the model with the smallest value, and for the BLRT, the optimal number of clusters is where the p value becomes non-significant at significance level 0.05. A three-class multimorbidity solution provided the best latent class model fit based on BIC, AIC, and BLRT p values. Patient allocation was based on posterior probabilities of belonging to a class. In order to determine the adjusted association of baseline patient characteristics with multimorbidity cluster membership, we fitted multinomial logistic regression models and reported using odds ratios and their corresponding 95% confidence intervals.
Multilevel linear regression analysis  of longitudinal changes in EQ-VAS scores (in hospital, 1 month, 6 months, and 12 months) was performed to investigate the associations of multimorbidity clusters and temporal changes in patient perceptions of health. The outcomes data are repeated measurements overtime, and patients are clustered within hospitals; therefore, data are not independent, and a multilevel linear model was used to account for the clustering in the data. The multilevel models were fitted in steps, first an unconditional means model was used to determine the significance of the 2 random-effect terms (hospital and patient). To check whether the linear model was appropriate, we examined the distribution of residuals to check that there were approximately normally distributed. The normal probability plots are reported in Additional file 1: Figure S2. Where the normality assumption was violated, the multilevel Tobit regression  models were fitted and compared to the multilevel linear model results and the results were similar. Tobit regression models are commonly used to analyse patient reported outcome measures data with ceiling and floor effects.
The analysis adjusted for time (categorised as baseline, 1 month, 6 months, and 12 months), age, sex, ethnicity (white vs other), smoking status (never vs ex or current), past medical history of MI, angina, diagnosis (STEMI or NSTEMI), revascularisation (percutaneous coronary intervention [PCI] vs. no PCI; coronary artery bypass graft [CABG] surgery vs no CABG surgery), medications (aspirin, β blockers, statins, and ACE inhibitors), referral for cardiac rehabilitation (yes/no), and interactions of time and multimorbidity class. The confounders were selected based on clinical consideration and previous research [4, 42].
Multilevel linear regression analysis of longitudinal changes in EQ-5D scores (in hospital, 1 month, 6 months, and 12 months) was performed to investigate the associations of multimorbidity clusters and temporal changes in HRQoL. Effect sizes (regression coefficients) and their corresponding 95% confidence intervals were used to assess the adjusted magnitude of the difference in EQ-VAS, EQ-5D scores across multimorbidity classes.
To investigate the association of multimorbidity clusters and changes in HRQoL measured by EQ-5D dimensions, five multilevel logistic regression models were fitted for the EQ-5D dimensions (mobility, self-care, activities, pain, and anxiety, and depression) adjusting for age, sex, ethnicity (white versus other) smoking status (never vs ex or current), past medical history of MI, angina, diagnosis (STEMI or NSTEMI), revascularisation (percutaneous coronary intervention [PCI] vs. no PCI; coronary artery bypass graft [CABG] surgery vs no CABG surgery), medications (beta-blockers, statins, angiotensin converting enzymes (ACE), aspirin), cardiac rehabilitation (yes/no), and interactions of time and multimorbidity.
The ‘extreme problem’ category for some domains of the EQ5D measure was endorsed by few individuals for some domains (e.g., self-care and mobility); therefore, we combined the EQ-5D levels ‘some problems’ and ‘extreme problems’ and the responses were binary (no problems vs some/extreme problems) and adjusted odds ratios (OR) and their corresponding 95% confidence intervals were used to assess the adjusted associations of multimorbidity classes and EQ-5D dimensions.
In longitudinal studies, missing data are commonly encountered, subjects can be missed at a particular assessment time; therefore, subjects may provide outcome data at some, but not all study time points resulting in incomplete data. Participants might drop out of the study or could be lost to follow up. In this study, there was missing outcome data over time. A sensitivity analysis was conducted comparing baseline characteristics of patients with complete data and those that dropped out (Additional file 1: Table S4). The drop outs were not significantly different from the followed up subjects in sex, ethnicity, previous angina, chronic renal failure, and PVD but were significantly different in age, diagnosis (NSTEMI, STEMI), Index of Multiple Deprivation (IMD), smoking status, history of previous acute MI (AMI), PCI, prevalence of diabetes, baseline EQ-5D, and EQ-VAS scores. In order to include patients with incomplete outcome data and to mitigate against biases which may arise as a result of such an omission, we used a multilevel model that includes all participants even if they were not assessed at all 4 time points.
All statistical tests were two-sided, and statistical significance was considered at p <0.05. Analyses were conducted using stata (IC) version 15
Whilst no patients were involved in setting the research question or the study design, we have co-produced this research manuscript with a patient with prior MI who provided input into the interpretation of the research findings, gave a critical review of the manuscript, and will work with our research team in ensuring its widespread dissemination.