Introduction

Diabetes affects nearly 300 million globally and contributes to over 10% of adult deaths, mainly through related cardiovascular and renal disease [1]. In the UK, direct treatment costs are estimated at £9.8 billion ($16.3 billion)—over 10% of the total National Health Service (NHS) budget—with further indirect costs of £4 billion ($6.7 billion) [2]. Type 2 diabetes accounts for 90–95% of all cases, is linked with the obesity ‘epidemic’ and is considered to be largely preventable [3].

The key requirements for management of type 2 diabetes are now well established and centre on lifestyle management (diet, weight control, smoking cessation, physical activity), primary and secondary prevention (glucose, BP and lipid control, monitoring and treatment of retinopathy, neuropathy and nephropathy) and, when necessary, referral to specialist services [4–7]. Consensus has not been reached, however, on how best to deliver these requirements or on optimal levels of control for biometric variables. For example, guidelines generally advise control of HbA1c levels to below 7.0% (53 mmol/mol) in North America and 7.5% (58.5 mmol/mol) in the UK but evidence has emerged that intensive control of blood glucose does not increase longevity, is associated with higher risk of cardiovascular death (particularly in younger patients [8]) and increases the risk of severe hypoglycaemia [9, 10]. Similarly, BP control below 130/80 mmHg was recently found not to reduce all-cause mortality in newly diagnosed diabetes patients [11]. Other studies demonstrated a U-shaped relationship between HbA1c levels and all-cause mortality, which is possibly attributable to residual confounding through unmeasured variables [12]. For example, patients in the later stages of cancer often have low plasma glucose levels, where the elevated risks of adverse outcomes are due to the cancer not the hypoglycaemia. Non-linear relationships have also been observed for the association between patient outcomes and BP (with the lowest risk of coronary events associated with systolic pressures of 120–130 mmHg [13, 14]) and lipid levels [15, 16]. Causality, however, is again difficult to establish due to confounders such as the use of statins.

Given such uncertainties, a greater understanding of the effect of risk factors on outcomes for diabetic patients in real-world settings is required. In this study we used a large longitudinal database of individual patient records to quantify the relationship between biological variables (HbA1c, BP and cholesterol levels) and key outcomes for type 2 diabetes, including complications and mortality, while controlling for time-varying patient and practice covariates (e.g. comorbidities and area deprivation). As far as we know this is the first investigation that attempts to control analyses for various comorbidities and thus limit residual confounding, while using novel methodologies to account for missing values and simultaneously examine the relationship between all key biological variables (biometrics) and outcomes.

Methods

Data source

We used the Clinical Practice Research Datalink (CPRD), a large primary care database that holds complete electronic patient records (including diagnoses, prescriptions and referrals) from participating family practices across the UK. A hierarchical clinical coding system (Read) is used to record the data. In July 2012, data were available for 644 practices and 13,772,992 patients. Full details of the database have been provided elsewhere [17].

Diabetes cohort

We extracted data from 1 April 2006 to 31 March 2012 and, for ease of reporting and analysing, aggregated information into six financial years. Within each year, practice inclusion eligibility was determined by a CPRD assessment algorithm, which informs on practices considered to be of research standard; therefore, our cohort of practices varied over time. For each research standard practice and year, we defined as eligible patients those who were registered with the practice for the full year and were aged 18 years or over in that year. From these patients, using relevant Read codes for type 2 diabetes (e.g. C10F.00: Type 2 diabetes mellitus) and excluding those treated with insulin within 2 years of diagnosis, we identified 246,544 patients over the study period. Diagnoses were not constrained to the study period and a relevant code prior to the study as well as during the study period would flag a patient from the respective year onwards. Data on sex, age and removal from the database due to deaths were available and complete for all patients. We extracted data on diabetes-related macrovascular (myocardial infarction, stroke, peripheral vascular disease or amputation) and microvascular (retinopathy, neuropathy, nephropathy [chronic kidney disease stages 4–5] or foot ulcer) complications as well as comorbidities (asthma, coronary heart disease, chronic kidney disease [excluded from microvascular analysis], chronic obstructive pulmonary disease, depression, dementia, severe mental illness, heart failure, hypertension, stroke [excluded from macrovascular analysis], cancer, epilepsy, osteoarthritis, osteoporosis and hypothyroidism). Although we aimed to include all conditions associated with diabetes, the choice was partially determined by the domains incentivised under the Quality and Outcomes Framework (QOF), for which accuracy of diagnosis is considered high [18]. Information was also extracted on smoking (never smoked, current, ex-smoker and missing data), BMI, HbA1c levels (%), cholesterol levels (mmol/l) and systolic/diastolic BP (mmHg). Biometric measurement data were cleaned and we calculated patient means for each year when more than one relevant record was available. Using product lists we determined prescription prevalence (at least one) for relevant medications: ACE inhibitors, acarbose, α-blockers, anticoagulants, antiplatelet agents, β-blockers, calcium-channel blockers, thiazide diuretics, loop diuretics, dipeptidyl peptidase-4 inhibitors, glucagon-like peptide-1 agonists, statins and other lipid-lowering drugs, meglitinides, metformin, sulfonylureas and thiazolidinediones. For approximately 60% of the practices, records were linked to Office of National Statistics (ONS) mortality data and we had access to death dates for all their patients. For these, using ICD-10 codes (www.who.int/classifications/icd/en/) we were able to estimate deaths linked to specific causes (underlying or in the top three): diabetes (E10–E16), ischaemic heart disease (I21–I22), stroke (I60–I64) or stroke excluding bleeds (excluding I63). All code lists used are available for download from www.clinicalcodes.org [19].

Statistical modelling

We used Cox (proportional hazards) regressions to investigate mortality and first new diabetes-related complication events in relation to the risk factors. For many patients, data for risk factors were less complete in the year of their death. For this reason, and to avoid reverse causality, when considering mortality we excluded data on these variables in the year of death. For first new macrovascular or microvascular diabetes complication (cases with any previous record of complication were excluded), missing data in the year of complication diagnosis was not an issue and they were modelled without a time lag. Separate analyses were performed for each complication type and mortality: all-cause (via CPRD and ONS record linkage); diabetes; coronary; stroke (ischaemic, haemorrhagic). Each model included HbA1c, cholesterol and systolic and diastolic BP as potentially explanatory variables. We included systolic and diastolic BP measurements as separate predictors in the models, rather than use an aggregate pulse pressure measure, since they demonstrated different relationship patterns with the outcomes and the correlation between them was only moderate (Pearson’s r = 0.49, after imputation), while UK, US and other international guidelines for controlling BP, _target systolic and diastolic pressures rather than pulse pressure. Additional covariates to control for possible confounding were age, sex, each comorbidity, BMI, microvascular and macrovascular complications (not if outcome), smoking and practice characteristics (diabetes prevalence, list size, region, area deprivation). We used Schoenfeld residuals to test the proportional hazards assumption and included time-varying covariates when needed to stabilise the models.

Due to the nature of the modelled variables (e.g. each comorbidity was identified by the presence of a relevant code), data was complete except for risk factor levels (BMI, HbA1c, BP and total cholesterol levels) and smoking status. For BMI, we used an interpolation algorithm to clean and impute missing values between observations over time. For all analyses, risk factor levels were categorised and for the main analyses missing data were coded as an additional category (Fig. 1 and electronic supplementary material [ESM] 1 Figs 1–7). However, we also employed ‘twofold’, a multiple-imputation algorithm for longitudinal data, which uses data within and across time to impute missing biometric measurements [20]. Through this approach we generated five complete datasets with which we conducted sensitivity analyses using multiple-imputation techniques. Additional sensitivity analyses for all-cause mortality were run in a subsample of patients aged 65 years or younger (to see if the patterns were similar in a younger population) and with a 2 and 3 year time lag on mortality (to verify that patterns were unaffected by biometric changes immediately prior to death, possibly due to frailty). Further sensitivity analyses for all-cause mortality examined whether patterns differed according to sex or to polypharmacy level. We defined three levels of polypharmacy using the 33rd and 66th centile of the count of medication groups a patient was prescribed on average within a year (medication groups are shown in Table 1). All analyses were performed using Stata v13 (StataCorp, College Station, TX, USA) and commands stcox and mi estimate. In ESM 1 and ESM 2 we provide more methodological details, discuss the role of BMI and provide the sensitivity analyses results.

Fig. 1
figure 1

Hazard ratios (CIs) for HbA1c level (a), cholesterol level (b), systolic BP (SBP) (c) and diastolic BP (DBP) (d) on all-cause CPRD mortality in the following year (main analysis). The standard dataset with time-varying covariates was used. The second category shown in all graphs represents the reference category. The missing category for diastolic BP has been omitted as it is perfectly collinear with the missing category for systolic BP. To convert values for HbA1c in DCCT % into mmol/mol, subtract 2.15 and multiply by 10.929

Full size image
Table 1 Characteristics for the study population, 2006–2012
Full size table

Results

Patient characteristics

The characteristics of the cohort are summarised in Table 1. Overall, recorded type 2 diabetes prevalence rose from 2.8% (148,570 patients) in 2006/2007 to 3.3% (166,718 patients) in 2011/2012 with the increase being greater in more deprived areas. In 2011/2012 prevalence ranged from 2.9% in the most affluent fifth of areas to 4.0% in the most deprived. Prevalence rates were highest in Wales, the East Midlands and the North-West of England. Between 2006/2007 and 2011/2012, the mean age of patients increased slightly from 66.1 to 66.4 years; the percentage of female patients remained constant at 44% to 45%; the mean diabetes duration increased from 6.8 to 7.8 years and the mean age at diagnosis dropped from 59.3 to 58.5 years.

The recorded prevalence of the first occurrence of a new microvascular complications rose from 21.9% in 2006/2007 to 33.0% in 2011/2012, while first new macrovascular complications dropped from 18.3% to 16.9% over the same period. Annual mortality rates varied between 3.3% and 3.7%: approximately a quarter of deaths were linked to diabetes; 10% to ischaemic heart disease and 10% to stroke. Over the study period, mean HbA1c levels remained stable at between 7.3% (56.3 mmol/mol) and 7.4% (57.4 mmol/mol), mean systolic BP fell from 138 to 136 mmHg, mean diastolic BP fell slightly from 77 to 76 mmHg and mean cholesterol levels fell from 4.4 to 4.3 mmol/l. BMI averaged around 31 in all years, while 15% of patients were recorded as current smokers and 65% as ex-smokers.

Prevalence of key medications and comorbidities are reported in Table 1. Statins, ACE inhibitors and metformin were the most widely prescribed drugs, while prescriptions for dipeptidyl peptidase-4 inhibitors and thiazolidinediones increased over time. Hypertension was by far the most commonly recorded comorbidity (>60%), followed by osteoarthritis (≈27%), depression (≈25%), chronic kidney disease (≈20%) and coronary heart disease (≈ 19%). More information on numbers of comorbidities is provided in ESM 1 Table A1.

Main analyses

Adjusted hazard ratios for mortality and complication events vs the differing levels of HbA1c, total cholesterol and BP are shown in Table 2.

Table 2 Hazard ratios (95% Cis) from Cox proportionate hazards survival analysis for all CPRD deaths, all ONS coronary heart disease-related deaths, all ONS cerebrovascular-related deaths and new diabetes complications (main analysis)a
Full size table

We observed U-shaped relationships between all-cause CPRD mortality and the mean levels of vascular risk factors (HbA1c, total cholesterol and BP), with both low and high levels associated with more deaths (Fig. 1). The HbA1c level associated with the minimum mortality risk was 7.5 ± 0.25% (56–61 mmol/mol). Compared with patients with HbA1c levels of 6.5 ± 0.25% (45–50 mmol/mol), hazard ratios for mortality were greater for patients with mean HbA1c levels <6.25% (45 mmol/mol) (1.17; 95% CI 1.12, 1.22), 8.0 ± 0.25% (61–67 mmol/mol) (1.09; 95% CI 1.03, 1.15) and >8.25% (67 mmol/mol) (1.20; 95% CI 1.15, 1.26). The total cholesterol level associated with the minimum mortality risk was 4.0 ± 0.5 mmol/l. Compared with patients with mean total cholesterol levels in the reference range 3.0 ± 0.5 mmol/l, hazard ratios were greater for patients with mean levels <2.5 mmol/l (1.40; 95% CI 1.29, 1.53) and >6.5 mmol/l (1.15; 95% CI 1.05, 1.25). For BP, less evident U-shaped relationships were observed. The systolic and diastolic BP associated with the lowest mortality risk was 140 ± 5 and 85 ± 2.5 mmHg, respectively. Higher systolic BP was associated with lower risk compared with the reference category (115–125 mmHg), although the risk seemed to increase for very high systolic BP. However, the highest risk was observed for pressures below 115 mmHg (1.22; 95% CI 1.17, 1.28). For diastolic BP, only pressures below 75 mmHg were linked with a statistically significant increase in mortality (1.07; 95% CI 1.03, 1.11), compared with the reference range (72.5–77.5 mmHg).

Results were similar for all-cause ONS mortality and broadly similar for diabetes-related mortality (ESM 1 Figs 5, 6 and Table 2). However, for the latter, hazard ratios for HbA1c levels more resembled a J-shape pattern, with increased risks for high levels, especially above 8.25% (67 mmol/mol) (1.49; 95% CI 1.33, 1.70). The pattern was different for coronary heart disease-related mortality, with statistically significant increased risks observed only for high values of HbA1c and cholesterol (ESM 1 Fig. 1). For HbA1c levels above 8.25% (67 mmol/mol) the hazard ratio was 1.32 (95% CI 1.10, 1.59), and for cholesterol levels above 5.5 and up to 6.5 mmol/l or above 6.5 mml/l, the hazard ratios were 1.61 (95% CI 1.27, 2.05) and 2.62 (95% CI 1.96, 3.52), respectively. For cerebrovascular (stroke)-related mortality high values of diastolic BP, cholesterol and HbA1c suggested higher risk, but CIs were much wider due to fewer deaths (ESM 1 Fig. 2). Only HbA1c levels above 7.75 and up to 8.25% (61–67 mmol/mol) or above 8.25% (67 mmol/mol) demonstrated a statistically significant increase in risk, with hazard ratios of 1.37 (95% CI 1.08, 1.73) and 1.31 (95% CI 1.07, 1.61), respectively. Patterns of results were similar when we excluded cerebrovascular deaths attributed to haemorrhage (ESM 1 Fig. 7, Table 2).

For both new macrovascular and new microvascular diabetes complications, patterns were almost linear for HbA1c, with increased levels associated with higher risk of developing at least one complication (ESM 1 Figs 3, 4). For example, compared with an HbA1c level of 6.25–6.75% (45–50 mmol/mol), hazard ratios for levels above 8.25% (67 mmol/mol) were 1.46 (95% CI 1.36, 1.57) for macrovascular and 1.27 (95% CI 1.23, 1.31) for microvascular complications, respectively. The relationship with systolic BP seemed linear for new microvascular complications, but U-shaped for macrovascular complications. Compared with the reference range (115–125 mmHg), levels below 115 mmHg were associated with higher macrovascular risk (1.17; 95% CI 1.06, 1.29) and levels above 175 mmHg were associated with increased risk for both macrovascular (1.64; 95% CI 1.37, 1.95) and microvascular complications (1.45; 95% CI 1.32, 1.59). Cholesterol levels above 3.5 and up to 6.5 mmol/l indicated a small but statistically significant risk reduction compared with the reference category (2.5–3.5 mmol/l), for both complication types. However, values above 6.5 mmol/l were associated with increased risk for macrovascular complications (1.29; 95% CI 1.14, 1.47), while values below 2.5 mmol/l were associated with hazard ratios of 1.44 (95% CI 1.23, 1.70) for macrovascular and 1.29 (95% CI 1.18, 1.41) for microvascular complications. The picture was less clear for diastolic BP, with values above 77.5 and up to 92.5 mmHg suggesting a small reduction in macrovascular risk, while a smaller but statistically significant microvascular risk reduction was observed for values above 92.5 mmHg.

Missing values for cholesterol and, especially, BP were found to be significant predictors of outcome in most analyses, while missing HbA1c values were predictors of complications. This implies that the missing completely at random (MCAR) assumption of a complete-case analysis could not be justified in some survival models and sensitivity analyses with multiple-imputation techniques would be informative. In most cases, missing data were associated with better outcomes, compared with the reference category, implying that measurements are more likely to be missing when patients are healthier.

Sensitivity analyses

Results with multiple-imputation methods were broadly similar to results without imputation although in some cases the U-shaped relationships were slightly more pronounced and the effects of extreme ranges were marginally higher (ESM 1 Tables 3, 4, ESM 1 Figs 8–15). The only exception was all-cause mortality vs diastolic BP, for which the estimated risk for high BP was markedly high and statistically significant (ESM 1 Figs 8, 13). For example, for diastolic BP above 92.5 mmHg, the hazard ratio was 1.26 (95% CI 1.14, 1.41) and 1.28 (95% CI 1.12, 1.46) in the CPRD and ONS deaths, respectively.

Relationship patterns were generally similar in the 2 and 3 year mortality analyses, except for systolic BP where the U-shape relationship became much more pronounced with an increase in mortality for higher values. For the younger subcohort (aged ≤65 years), U-shaped relationships for HbA1c, total cholesterol and systolic BP appeared more extreme, while the pattern for diastolic BP was similar. Patterns were similar for male and female patients (with the exception of high systolic BP values) and different polypharmacy levels, indicating that the risk for confounding by severity is small (ESM 2).

Discussion

We found that higher levels of three key factors (HbA1c, total cholesterol and BP) were generally associated with higher risks of adverse outcomes for patients with type 2 diabetes, but we also saw several U-shaped relationships—particularly with mortality—even after adjusting for important confounders.

Strengths and limitations of the study

Our analysis advances current knowledge in several areas: (1) it assessed risks for mortality and several diabetes-related complications associated with levels of three modifiable risk factors; (2) it used a large longitudinal database of individual patient records; (3) it adjusted for several important comorbidities and time-varying covariates; and (4) it dealt with missing data using sophisticated multiple-imputation techniques. The observed prevalence of diabetes was lower in our study than has been reported elsewhere [21] but we excluded patients with type 1 diabetes and those receiving insulin within 2 years of diagnosis. Without these exclusions, prevalence rates were closer to national estimates.

Although the cohort size gave us the statistical power to investigate cause-specific mortality and to incorporate important covariates, some important limitations exist. First, diabetes and other comorbidities were identified from practice populations using relevant code lists and algorithms and were not verified cases. Although we do not expect false positives to be a serious issue [22], we will not have identified undiagnosed patients or those who rarely visit their general practice. Second, the list of comorbidities used was not exhaustive. Third, measurements of biological variables are subject to error, particularly in the case of BP, which is measured using instruments of variable accuracy and is prone to digit bias in recording [23]. Also, since 2004, UK practices have had financial incentives for controlling BP and there is evidence to show that recordings of systolic BP have been biased downwards for patients with values just above the _target levels (_target levels for diabetic patients were 145 mmHg from 2004/2005 to 2010/2011 and 140 mmHg from 2011/2012) [24]. Fourth, some risk factor measurements might be missing not at random (MNAR), as our analyses assume, although multiple imputation offers some protection against biased estimates [25, 26]. Fifth, late registration of deaths through the ONS meant that the proportion of deaths captured is lower for later years in the coverage period, which might have led to an underestimation of the risks. However, the CPRD death algorithm provides a more complete picture for later years and the results for these two death outcomes agree closely. Sixth, we ignored the effect of covariates within the year of death, since we would have many missing values for that year and imputing them all would have been problematic. Seventh, we only assessed the development of first complications, and the risks associated with subsequent complications may not follow the same pattern. Eighth, each risk factor could have a different time-varying relationship with each outcome, confounded by unmeasured confounders. Ninth, we used a yearly time window, which is suboptimal for survival analysis, but our choice was driven by QOF recording practices. A smaller time window would introduce more missing data under MNAR mechanisms. Tenth, CPRD practices use a single computer system and differences have been found across systems used in England [27]. Finally, data are observational and relationships between risk factor levels and clinical outcomes may be affected by unmeasured confounding.

Findings compared with previous studies

Recent trials have found that intensive glucose control is associated with increased risk of cardiovascular death in younger patients [8] and observational studies have generally demonstrated U-shaped relationships between levels of HbA1c in diabetic patients and death [28, 29], possibly attributable to residual confounding through unmeasured variables [12]. In our study we adjusted for several important confounders, including comorbidities such as heart failure, chronic obstructive airways disease and chronic kidney disease. We also adjusted for BMI, low values of which are often associated with severe comorbidity, frailty and hypoglycaemia. Nevertheless, we still found U-shaped relationships between HbA1c levels and cerebrovascular and all-cause mortality—HbA1c values below 6.25% (45 mmol/mol) and above 7.75% (61 mmol/mol) were associated with increased risk. The risk of coronary heart disease-related death increased with HbA1c levels in a more linear fashion, as did the risk of complications.

We found similar patterns for total cholesterol, with a minimum risk of complications and death for patients with levels in the 3.5–5.5 mmol/l range. Interpreting these findings is more difficult, as clinical trials of aggressive cholesterol lowering have not demonstrated increased risk of mortality [30, 31] and there is no obvious explanation for low cholesterol causing diabetes-related microvascular complications. Residual confounding is a possible explanation. Patients with very low total cholesterol levels are more likely to be receiving high-dose statin therapy for severe (generalised) vascular disease and in our cohort 86.8% of patients with mean annual cholesterol levels below 2.5 mmol/l were prescribed statins, compared with 54.5% of those whose levels were above 6.5 mmol/l. Severe vascular disease is associated with poorly controlled long-duration diabetes, and this may be the underlying cause of the higher risk for microvascular complications. However, we observed almost identical results after excluding patients who were prescribed statins.

Aggressive BP control in type 2 diabetes has been found not to reduce major cardiovascular events [32], and U-shaped relationships have recently been described between BP and mortality and cardiovascular outcomes [11, 33]. We found that these relationships persisted after adjusting for important confounders. This may be attributable to under-perfusion of vital organs, but we observed a higher risk of mortality for systolic BP in the 125–135 mmHg range, which would not be expected to cause under-perfusion. Residual confounding from unmeasured variables, such as frailty and undiagnosed heart failure, may therefore have contributed to these findings.

Conclusions

Our findings from this inclusive population-level investigation confirm and expand on trial evidence from more selected populations [34, 35], concerning the management of HbA1c and BP but not total cholesterol. Further corroboration is necessary, particularly in prospective longitudinal trials but, if validated, the findings have several implications for type 2 diabetes management. Current clinical guidelines aim to control HbA1c to below 7.0% (53 mmol/mol) in North America and 7.5% (58 mmol/mol) in the UK, and to keep BP below 140/80 mmHg [36, 37]. The UK QOF pay-for-performance scheme uses _targets of 7.5% (58 mmol/mol), 8% (64 mmol/mol) and 9% (75 mmol/mol) for HbA1c, 140/80 and 150/90 mmHg for BP and 5 mmol/l for total cholesterol. Our results suggest that achieving these _targets may not optimise patient outcomes. For example, we found that HbA1c values above 7.75% (61 mmol/mol) were associated with higher risks of adverse outcomes, suggesting that two of the QOF quality _targets (HbA1c ≤8% [64 mmol/mol] and ≤9% [75 mmol/mol]) may reward practices for achieving levels that could increase patient risk of complications and death.

Current guidelines only aim to keep HbA1c below a _target value. The ACCORD (Action to Control Cardiovascular Risk in Diabetes) study indicated that aggressive glucose lowering in type 2 diabetes (to normalise HbA1c below 6% [42 mmol/mol]) leads to increased mortality compared with a conventional HbA1c _target of 7–8% (53–64 mmol/mol) [10]. In our data, strong U-shaped relationships between HbA1c and mortality persisted even after adjusting for important confounders, raising concerns about potential harms from aggressive glucose lowering. Our data, combined with trial data, would argue for controlling HbA1c within a given range (e.g. 6.0–7.5% [42–58 mmol/mol]) rather than below an upper limit. Recent EASD and ADA guidelines recommend individualised HbA1c _targets based on factors including age, diabetes duration and the presence of vascular disease and comorbidity [38]. Our results suggest that range-based _targets may also be more appropriate for other diabetic risk factors, including total cholesterol and BP. Optimal management of these multiple biological variables in patients with type 2 diabetes is complex and there are still many unknowns to be addressed before we can be confident that we are getting it right.