Optimal strategies for identifying kidney disease in diabetes: properties of screening tests, progression of renal dysfunction and impact of treatment systematic review and modelling of progression and cost-effectiveness

Definitions of microalbuminuria

Microalbuminuria is conventionally defined as levels of albumin ranging from 30 to 300 mg in a 24-hour urine collection (Table 2). 15 Overt albuminuria, macroalbuminuria or proteinuria is defined as a UAE of ≥ 300 mg/24 hours. There is increasing recognition, however, that there is not an arbitrary boundary between levels of albumin excretion and that there is a continuum of progression in which the level of albumin provides a marker of how far the disease has advanced. 3 In recognition of this, a recent position statement has recommended that the terms microalbuminuria and macroalbuminuria be replaced, respectively, by the terms ‘high albuminuria’ and ‘very high albuminuria’. 16

In clinical practice, albuminuria is measured by the albumin-to-creatinine ratio (ACR) in a random or first-pass morning sample of urine. Because women excrete less creatinine than men, clinicians in the UK use different ACR thresholds for diagnosing microalbuminuria in men and women. 17

Are tests to detect the onset of microalbuminuria reliable?

Detecting whether levels of urinary albumin reach the threshold for diagnosis of microalbuminuria is not straightforward. First, levels of albumin may be affected by concurrent illness including virus infections and urinary tract infections. They may also be elevated after physical activity. Second, levels vary within an individual. There is evidence, at least for type 1 diabetes, that the occurrence of intermittent tests above the threshold for microalbuminuria predicts progression, and that those with levels of urinary albumin close but below the threshold have an increased risk of progression. 18 To address these issues, a standard protocol for diagnosis of microalbuminuria requires additional tests following a positive test, with at least one of the further tests required to also be above the diagnostic threshold (Figure 2). Standard protocols recommended in guidelines vary slightly;15,19 however, they are broadly consistent in aiming to optimise the delivery of targeted intensification of therapy to high-risk patients, for whom the treatments will provide a greater absolute risk reduction than they would in those at lower risk. 3

FIGURE 2.

Flow chart for confirmation of microalbuminuria. Tests need to be carried out in the absence of proteinuria/urinary tract infection. Test abnormal if ACR > 2.5 mg/mmol for men and > 3.5 mg/mmol for women. Based on the guidance in National Institute for Health and Care Excellence CG66. 19

The development and progression of diabetic kidney disease

Determinants for the development and progression of diabetic kidney disease are well established and discussed in detail below. Although there are familial elements that contribute to the risk of developing kidney disease, there are, as yet, no genetic traits that have been identified as contributing a substantial risk. However, there are some differences in the risk factors for development and progression of diabetic kidney disease between type 1 and 2 diabetes.

What risk factors affect progression to microalbuminuria?

Impact of impaired renal function on long-term outcomes

Interventions to prevent the development of microalbuminuria

The rationale for screening for microalbuminuria depends on the existence of an effective treatment for those identified as at increased risk, and the extent to which there is a difference in benefit between treating those identified and all patients. A series of large, interventional clinical trials has confirmed that intensive control of glycaemia and blood pressure among individuals with early diabetes reduces the occurrence of microalbuminuria. The DCCT trial demonstrated substantial reductions in the development of microalbuminuria with intensive glycaemic management. 45 Similarly, both the UKPDS57 and ADVANCE (Action in Diabetes and Vascular Disease: Preterax and Diamicron MR Controlled Evaluation)58 trials have confirmed that the treatment of patients with type 2 diabetes leads to a risk reduction with, for example, a relative reduction of 20% in onset of new microalbuminuria. 58 However, the extent to which treatment should differ between those with and without microalbuminuria is unclear.

Blood pressure, in particular, is a target for treatments that may have renal benefit. Conventional thresholds at which to initiate treatment to lower blood pressure are 130/80 and 125/75 mmHg for people with diabetes and those with diabetes and kidney disease, respectively. 59 Trials that evaluated lowering blood pressure below these thresholds, including UKPDS60 and ADVANCE,61 confirm that participants allocated to regimens targeting lower blood pressures had reduced risks of developing microalbuminuria. However, in the ACCORD (Action to Control Cardiovascular Risk in Diabetes) study, which targeted patients with more advanced diabetes and prior cardiovascular events, similar benefits were not seen when targeting a systolic blood pressure of 120 mmHg rather than 130 mmHg. 62

There is strong evidence that overactivity of the renin–angiotensin–aldosterone system (RAAS) is implicated in the deterioration of renal function in patients with diabetic kidney disease. Angiotensin-converting enzyme (ACE) and angiotensin 2 receptor blockers (A2RBs) slow deterioration of GFR and lower the rate of albumin excretion. The extent to which benefits occur at early stages of the disease is unknown. There is some evidence that suppression immediately reduces UAE rates, leading to a lower rate of progression of kidney disease,63 but, to date, a systematic review of the literature has not addressed this question.

We could not identify any studies that separately examine the impact of other specific interventions that are known predictors of the development of microalbuminuria, for example smoking cessation.

Risk reduction in microalbuminuria

Strategies for risk reduction in diabetes patients with microalbuminuria may be envisaged as falling into two broad and potentially overlapping categories. First, there may be therapies that are effective in reducing morbidity and mortality for any individual with diabetes; however, the overall absolute risk reduction may be greatest when individuals are already high risk because of their microalbuminuria. Second, there may be therapies for which the treatment effect differs in those with and without microalbuminuria.

Reducing blood pressure among those with microalbuminuria may result in a greater reduction in risk than a reduction in those without microalbuminuria. This raises issues of cost–benefit, requiring judgements about the benefit of absolute risk reduction and the adverse impacts of additional medication. Control of lipids and treatment goals are managed in the same way in the presence or absence of microalbuminuria. 64 Other strategies, such as low-protein diets, have been found to be effective only in moderate to severe kidney failure, not at early stages of kidney disease.

There is good evidence for targeting the overactivity of the RAAS with ACE inhibitors (ACEi) and A2RBs to reduce progression to kidney disease and reduce risks of cardiovascular events. However, the widespread use of ACEi in patients with type 2 diabetes (up to 50% of patients with diagnosed diabetes are treated for hypertension) indicates that, at least for some groups of patients, there may be limited opportunity for treatment optimisation. 64

Search strategy

Trials published up to January 2005 were identified from the reference lists of two Cochrane systematic reviews on this topic. 63,64 Additional articles published between January 2005 and August 2010 were identified as follows: MEDLINE (January 2005 to August 2010), EMBASE (January 2005 to August 2010) and the Cochrane Controlled Trials Register (issue 1, 2005, to issue 8, 2010) were searched using the search strategy described in a previous Cochrane systematic review for diabetes and randomised controlled trials,63 as well as terms to identify diabetes, urinary albumin, ACEi or A2RBs (see Appendix 2). Reference lists of identified trials and relevant reviews were also searched.

Trial selection criteria

Included trials met the following criteria: (1) randomised controlled design; (2) outcomes for patients with type 1 and type 2 diabetes were reported separately; (3) the intervention used was either an ACEi or an A2RB; the comparator group was a placebo, no treatment or antihypertensive group (other than ACEi or A2RB); (4) trial duration was ≥ 6 months; and (5) at least one outcome was separately reported for patients with normoalbuminuria or micro-/macroalbuminuria and urinary albumin levels were reported, either as a timed albumin excretion rate, or as a spot test ACR. The outcomes evaluated were (1) UAE at the end of the trial; (2) progression to microalbuminuria (30–300 mg/24 hours); (3) progression to macroalbuminuria (> 300 mg/24 hours); (4) regression to normoalbuminuria (< 30 mg/24 hours); (5) change in GFR in ml/minute/1.73 m²; and (6) death.

Data extraction

Relevant publications were identified in duplicate and the data independently extracted. Disagreements were resolved by consensus with a third author. Structured forms were used for data reviewing extraction (see Appendix 3). Unpublished trial data were sought from authors of trials published since 2000. In trials that compared two doses of the same ACEi or A2RB with a comparator (placebo, no treatment or other antihypertensive), only the data from the higher dose were included in the meta-analysis to avoid double counting. 73

When multiarm trials had both placebo arms and active comparator arms (a non-ACEi/A2RB antihypertensive), we preferentially selected the antihypertensive arm as the comparator for the ACEi/A2RB arm. In a secondary analysis designed to inform our cost-effectiveness models (see Chapters 5 and 6), we also preferentially selected the active antihypertensive comparator arms from these trials for use in a metaregression of effect size on blood pressure difference. Trials that looked at both an ACEi and an A2RB in comparison with a placebo, no treatment or other antihypertensive treatment were included in the meta-analysis by splitting the data from the comparator group between the required comparisons.

Statistical methods

Statistical analysis was carried out using Stata version 11.1 (StataCorp, College Station, TX, USA). Data on progression to micro-/macroalbuminuria and regression to normoalbuminuria were analysed as dichotomous data using a Mantel–Haenszel fixed-effects model and specifying the risk ratio as the effect measure. The analysis of continuous urinary albumin data was based on an inverse variance method that also used a fixed-effects model. 73 For comparison, we report results and CIs from random-effects analyses based on the DerSimonian and Laird method. 74 A ‘ratio of means effect measure’ was used to allow us to combine the 35 trials that measured UAE using albumin excretion rate (mg/24 hours) with the five trials that reported ACR (mg/mmol or mg/g) into a single analysis. 75–77 We took the ratio of means to be the ratio of the final value in the intervention group to the final value in the comparator group, so that a ratio of means < 1 favours intervention. A comparison with standard analysis methods using mean difference and standardised mean difference verified that ratio of means was a suitable analysis method.

The following sensitivity analyses were carried out to verify that our results were not sensitive to these approximations, or to trials in which the comparator group was not an antihypertensive group; double-blinding was not clearly stated; it was not clearly stated that at least two measurements of urinary albumin were made to categorise patients with normo-, micro- and macroalbuminuria (applied both at baseline and end point); the number of participants was < 200. Statistical heterogeneity between studies was assessed with the I² statistic. For the GFR data, we used an inverse variance, fixed-effect model and ratio of means. For a few studies, it was necessary to impute or approximate data. We undertook random-effects metaregression to evaluate the explanatory value of baseline urinary albumin (microalbuminuria vs. normoalbuminuria) and medication type (ACEi vs. A2RB) on effect size separately for type 1 and type 2 diabetes. Because a high degree of statistical heterogeneity was observed between studies in the analysis of continuous data, a post-hoc multivariate metaregression analysis was carried out to test whether the heterogeneity was explained by the following factors: study length; trial size; average participant age; country of trial; comparator type (no treatment, placebo or active comparator); medication type (ACEi or A2RB); treatment dose (categorised as high dose vs. low dose based on usual dose in the British National Formulary and excluding one study of a drug not licensed in the UK; temocapril, Acecol, Sankyo); difference in baseline UAE between the treatment and comparator groups, whether or not approximations were made; or year of trial publication. In addition, the effect of comparator type (placebo or no treatment vs. another antihypertensive) on effect size was investigated using subgroup analyses and metaregression.

Univariate and multivariate metaregression were carried out using the metareg routine in Stata version 11.1, applying the method of DerSimonian and Laird. 74 Heterogeneity was estimated from the Mantel–Haenszel model,73 and reported using the I² statistic.

Because subgroup analyses of trials comparing ACEi/A2RB with other antihypertensives suggested an effect of ACEi/A2RB on renal outcomes beyond that attributable to blood pressure reduction, we also conducted a metaregression of effect size on between-arm blood pressure difference and used this to inform the cost-effectiveness modelling in Chapters 5 and 6.

Description of studies

The reference lists of two Cochrane systematic reviews identified 65 trials, 38 of which were assessed for full eligibility (a selection of the excluded studies are listed in Appendix 4). Of these, three trials could not be included in the meta-analysis because they measured urinary protein instead of urinary albumin. 78–80 A fourth trial was excluded because consistent data could not be extracted,81 leaving 34 trials which fully met our inclusion criteria. In addition, 7849 articles were identified from the search of MEDLINE, EMBASE and The Cochrane Library from January 2005 to August 2010. Of these articles, 1450 were duplicates (Figure 3), leaving 6399 articles for review. After screening the titles and abstracts, 671 articles were examined in more detail for eligibility. A further 659 articles were excluded, leaving 12 papers (15 trials) that met our inclusion criteria.

FIGURE 3.

Flow diagram of literature search.

A total of 49 trials (representing 34,082 patients) were therefore included in the meta-analysis, three of which had multiple arms,82–84 which were included as separate comparisons, resulting in a total of 52 comparisons (see Figure 3). Six trials were included in the analysis of patients with type 1 diabetes and normoalbuminuria (seven comparisons). A total of 15 trials were included in the analysis of patients with type 1 diabetes and micro- or macroalbuminuria. As one trial was included for both the normoalbuminuria and the microalbuminuria groups,85 this gives a total of 20 trials of type 1 diabetes patients. In the analysis of patients with type 2 diabetes and normoalbuminuria, 10 trials were included; 22 trials were included in the analysis of patients with type 2 diabetes and micro- or macroalbuminuria (24 comparisons). As four trials were included in both the normoalbuminuria and microalbuminuria groups, this gives a total of 28 included trials for type 2 diabetes. 61,86–88 One trial could not be included in the analysis because only one patient was left in the comparator group at the end of the trial. 89

Of the 21 trials of RAASI carried out in people with type 1 diabetes, only 3 used A2RBs (Table 6),82,88 all of which were carried out on patients with normoalbuminuria. The remaining trials of type 1 diabetes patients used ACEi. As regards type 2 diabetes, three trials of normoalbuminuric patients used A2RBs, and nine trials of microalbuminuric patients used A2RBs; the remaining trials all used ACEi (see Table 6).

All trials of type 1 diabetes included normotensive or controlled hypertensive patients, and the majority of trials allowed patients to take antihypertensive medications (other than RAASI) (see Table 6). In contrast to this, nearly all trials in type 2 diabetes and normoalbuminuria included hypertensive patients, whereas 6 of the 21 trials in type 2 diabetes patients with microalbuminuria specifically selected hypertensive patients (see Table 6). Only one trial in type 1 diabetes,85 and two trials in type 2 diabetes,86,88 directly compared patients with normo- and microalbuminuria.

Trials of type 1 diabetes

The ratio of mean UAE in treatment versus comparator groups at the end of the trial for type 1 diabetes is shown in Figure 4. In patients normoalbuminuric at baseline, the ratio of means was 0.96 (95% CI 0.94 to 0.99) in the fixed-effects model (inverse variance) and 0.94 (95% CI 0.79 to 1.12) in the random-effects model (DerSimonian and Laird). This equated to a 4% and 6% lower UAE in the treatment group versus the comparator groups, respectively. In microalbuminuric patients, the ratio of means was 0.40 (95% CI 0.36 to 0.44) using a fixed-effects model (60% lower) and 0.33 (95% CI 0.23 to 0.46) using a random-effects model (67% lower).

FIGURE 4.

Ratio of mean UAE in treatment and comparator arms of trials. Ratio of mean UAE at the end of trials of RAASI treatment versus comparator (boxes) and pooled estimates across trials (diamonds) calculated by the fixed-effects inverse variance (I-V) method and by the DerSimonian and Laird random-effects (D+L) method, in patients with (a) type 1 and (b) type 2 diabetes, stratified by baseline urine albumin status. Horizontal bars and diamond widths denote 95% CIs and box sizes indicate relative weight in the I-V analysis. ATLANTIS, ace-inhibitor trial to lower albuminuria in normotensive insulin-dependent subjects; BENEDICT, Bergamo nephrologic diabetes complications trial; CCB, calcium channel blocker; DIRECT, DIabetic REtinopathy Candesartan Trials; EUCLID, a study comparing cardiovascular effects of ticagrelor and clopidogrel in patients with peripheral artery disease; NT, no treatment; RENAAL, reduction in end points in noninsulin-dependent diabetes mellitus with the angiotensin II antagonist losartan; ROADMAP, randomized olmesartan and diabetes microalbuminuria prevention; SMART, shiga microalbuminuria reduction trial.

A subgroup analysis by comparator type was conducted to explore the difference between RAASI and an antihypertensive comparator, and RAASI and placebo versus no treatment. These subgroup analyses were not possible in patients with normoalbuminuria, as all trials used placebo as the comparator. In trials of microalbuminuric patients, subgroup analyses gave a ratio of means of 0.44 (95% CI 0.39 to 0.48) in 12 placebo or no treatment trials, and a ratio of means of 0.20 (95% CI 0.14 to 0.27) in two active comparator trials. This and other sensitivity analyses are listed in Table 7.

TABLE 7 - Analyses of subgroups of trials to explore sensitivity to assumptions made

Results in bold are from inverse variance fixed-effects analysis; results not in bold are from DerSimonian and Laird random-effects analysis UAE.

Results are presented as the ratio of mean UAE levels (intervention/comparator) at the end of the trial.

Analysis	Type 1 normoalbuminuria		Type 1 microalbuminuria		Type 2 normoalbuminuria		Type 2 microalbuminuria
	Number of trials	Ratio of mean UAE levels (95% CI)a	Number of trials	Ratio of mean UAE levels (95% CI)a	Number of trials	Ratio of mean UAE levels (95% CI)a	Number of trials	Ratio of mean UAE levels (95% CI)a
Main analysis (Figure 2)	7	0.96 (0.94 to 0.99)	14	0.40 (0.36 to 0.44)	7	0.88 (0.84 to 0.92)	20	0.77 (0.74 to 0.80)
		0.94 (0.79 to 1.12)		0.33 (0.23 to 0.46)		0.79 (0.68 to 0.93)		0.73 (0.62 to 0.86)
Trials with antihypertensive comparator group	–	–		0.20 (0.14 to 0.27)		0.65 (0.51 to 0.83)		0.82 (0.78 to 0.86)
	–	–	2	0.20 (0.14 to 0.27)	2	0.65 (0.51 to 0.83)	7	0.82 (0.64 to 1.06)
Trials with duplicate urinary albumin measurements for categorising urinary albumin status of patients		0.96 (0.94 to 0.99)		0.36 (0.31 to 0.41)		0.88 (0.84 to 0.92)		0.69 (0.65 to 0.72)
	7	0.94 (0.79 to 1.12)	12	0.35 (0.23 to 0.53)	5	0.81 (0.69 to 0.95)	15	0.69 (0.60 to 0.79)
Trials with double-blinding		0.97 (0.94 to 0.99)		0.43 (0.37 to 0.50)		0.88 (0.84 to 0.92)		0.71 (0.66 to 0.75)
	6	0.95 (0.79 to 1.14)	10	0.43 (0.28 to 0.65)	5	0.81 (0.69 to 0.95)	11	0.69 (0.57 to 0.83)
Trials n ≥ 200		0.96 (0.94 to 0.99)	–	–		0.92 (0.88 to 0.96)		0.72 (0.67 to 0.77)
	3	0.96 (0.94 to 0.99)	–	–	4	0.90 (0.83 to 0.98)	3	0.73 (0.54 to 0.98)

Other outcomes

Type 1 diabetes

Results for progression and regression of albuminuria are presented in Figure 5. RAASI treatment of normoalbuminuric patients led to no significant difference in the number of patients who progressed to microalbuminuria, with a relative risk of 0.96 (95% CI 0.76 to 1.23, p = 0.25, I² = 24%). All trials using ACEi favoured the treatment group, although the effect was not significant. In contrast, all trials using A2RBs favoured the comparator group, although, again, not significantly. ACEi treatment of patients with microalbuminuria resulted in fewer progressing to macroalbuminuria (UAE > 300 mg/24 hours), with a relative risk of 0.39 (95% CI 0.23 to 0.64, p = 0.0005, I² = 0%). Treatment with ACEi also resulted in more patients regressing from microalbuminuria to normoalbuminuria, with a relative risk of 5.81 (95% CI 2.05 to 16.43, p = 0.001, I² = 0%). In those studies in which it was assessed, there was no significant effect of treatment on mortality (see Figure 5) or GFR (data not shown).

FIGURE 5.

Relative risk of progression and regression of albuminuria in treatment and comparator arms of trials in type 1 diabetes. Relative risk of progression and regression of albuminuria during trials of RAASI treatment vs. comparator (boxes) and pooled estimates across trials (diamonds) calculated by the fixed-effects inverse variance (I-V) method and by the DerSimonian and Laird random-effects (D+L) method, in patients with type 1 diabetes, stratified by baseline urine albumin status. Horizontal bars and diamond widths denote 95% CIs and box sizes indicate relative weight in the I-V analysis. (a) Progression from normo- to microalbuminuria; (b) progression from micro- to macroalbuminuria; and (c) regression from micro- to normoalbuminuria. ATLANTIS, ace-inhibitor trial to lower albuminuria in normotensive insulin-dependent subjects; DIRECT, DIabetic REtinopathy Candesartan Trials; EUCLID, a study comparing cardiovascular effects of ticagrelor and clopidogrel in patients with peripheral artery disease.

Type 2 diabetes

Results for progression and regression of albuminuria are presented in Figure 6. Patients with both normo- and microalbuminuria benefited from treatment with RAASI in terms of progression or regression of urinary albumin, but the effect was larger for patients with microalbuminuria. Fewer patients with normoalbuminuria treated with RAASI progressed to microalbuminuria, with a relative risk of 0.84 (95% CI 0.79 to 0.89, p = 0.002, I² = 19%). RAASI treatment resulted in fewer patients progressing from micro- to macroalbuminuria, with a relative risk of 0.52 (95% CI 0.43 to 0.63, p = 0.00001, I² = 48%). More patients regressed from micro- to normoalbuminuria in the treated group than in comparator groups, with a relative risk of 1.20 (95% CI 1.12 to 1.29, p = 0.03, I² = 75%). In those studies in which it was assessed, there was no significant effect of treatment on mortality (Figure 7) or GFR.

FIGURE 6.

Relative risk of progression and regression of albuminuria in treatment and comparator arms of trials in type 2 diabetes. Relative risk of progression and regression of albuminuria during trials of RAASI treatment versus comparator (boxes) and pooled estimates across trials (diamonds) calculated by the fixed-effects inverse variance (I-V) method and by the DerSimonian and Laird random-effects (D+L) method, in patients type 2 diabetes, stratified by baseline urine albumin status. Horizontal bars and diamond widths denote 95% CIs and box sizes indicate relative weight in the I-V analysis. (a) Progression from normo- to microalbuminuria; (b) progression from micro- to macroalbuminuria; and (c) regression from micro- to normoalbuminuria. BENEDICT, Bergamo nephrologic diabetes complications trial; DIRECT, DIabetic REtinopathy Candesartan Trials; ROADMAP, randomized olmesartan and diabetes microalbuminuria prevention; SMART, shiga microalbuminuria reduction trial.

FIGURE 7.

Relative risk of all-cause mortality in treatment and comparator arms of trials. Relative risk of all-cause mortality during trials of RAASI treatment versus comparator (boxes) and pooled estimates across trials (diamonds) calculated by the fixed-effects inverse variance (I-V) method in patients with (a, b) type 1 and (c, d) type 2 diabetes, stratified by baseline urine albumin status. Horizontal bars and diamond widths denote 95% CIs and box sizes indicate relative weight in the I-V analysis. (a) Normoalbuminuria; (b) microalbuminuria; (c) normoalbuminuria; and (d) microalbuminuria. ATLANTIS, ace-inhibitor trial to lower albuminuria in normotensive insulin-dependent subjects; BENEDICT, Bergamo nephrologic diabetes complications trial; DIRECT, DIabetic REtinopathy Candesartan Trials; RENAAL, reduction in end points in noninsulin-dependent diabetes mellitus with the angiotensin II antagonist losartan; ROADMAP, randomized olmesartan and diabetes microalbuminuria prevention.

Metaregression

In a univariate metaregression of type 1 diabetes studies, it was found that treatment effect in trials of microalbuminuric patients was significantly greater than in trials of normoalbuminuric patients (p = 0.006). In studies of type 2 diabetes, differences in treatment effect between microalbuminuria and normoalbuminuria were not significant (p = 0.65). Metaregression of the type of medication (ACEi or A2RB) could not be performed for type 1 diabetes, as too few trials used A2RB; however, medication type was not found to significantly affect urinary albumin levels in patients with type 2 diabetes (p = 0.43). Comparator type (antihypertensive medication vs. no antihypertensive medication) did not significantly alter treatment effect either for type 1 diabetes (p = 0.34) or for type 2 diabetes (p = 0.86). In multivariate metaregression, the effect of the following factors on heterogeneity were explored: baseline urinary albumin, medication type, trial country, age, differences in baseline urinary albumin levels, trial size, data extraction method and dose of RAASI. None of these factors individually reduced the I² statistic by > 20%, with trial country having the biggest impact on the heterogeneity in both type 1 and type 2 diabetes trials.

Data sources

We evaluated progression of ACR, eGFR and associated parameters in a cohort of newly diagnosed type 1 diabetes patients drawn from ORPS. A comprehensive description of the population and the research design is provided in a previous study report. 48 Briefly, between 1986 and 1996 young people, living in Oxfordshire, UK, and diagnosed with type 1 diabetes before the age of 16, were invited to participate in a longitudinal study. Children with diabetes secondary to another condition were excluded. Annual assessment of clinical and biochemical measures included three consecutive early morning urine specimens. Urine samples were assayed for albumin and creatinine in a central laboratory and the urinary ACR was calculated for each urine sample. Albumin was measured by an enzyme-linked immunosorbent assay, and creatinine using the modified Jaffe method. Serum creatinine samples were used to calculate eGFR using the Schwartz formula in patients < 18 years of age,138 and the MDRD formula for those ≥ 18 years. 14 Annual assessments continued for up to 20 years. Written consent was obtained from parents and children were asked to give assent before commencement of the study. Ethical approval was obtained from the local ethics committees.

Statistical methods

The parameters describing the progression of urine ACR, its variability and its test performance were derived from the repeated measurement in the ORPS data set. We used WinBUGS 1.4 (MRC Biostatistics Unit, Cambridge, UK)139 to fit the linear random-effects model described below, and report model parameters with 95% credible intervals as approximations to 95% CIs. We considered the parameters describing the progression of ACR under three main headings: (1) the between-subject variation in baseline ACR (often referred to as the intercept), for which the tendency for patients’ ACR to systematically differ from other patients’ was modelled as a mean and standard deviation of log-ACR; (2) the average change in ACR over time, which was modelled as the change in log-ACR per year; and (3) the difference between the observed ACR and the underlying ACR for a single individual, conceptualised as the within-person variability in any individual ACR measurement arising from the analytic variability in the assay and short-term biological variability. The first two of these components use random-effects models with normal distributions to describe the underlying log-ACR for each patient over time, whereas the third component uses a t-distribution to summarise the within-measurement variability. An inverse t-distribution with three degrees of freedom was used to describe the within-measurement variability of log-ACR. The t-distribution was used because the variation in individual measurements was too great to be described by a normal distribution; in particular, some measurements were much further from the mean than the standard deviation. The model included effects of age and sex on underlying ACR. A previous report from ORPS suggested that the age effect could be partly attributable to puberty, but, after adjustment for age, we found no significant effect on ACR (p = 0.4) and, therefore, did not include it in the final model.

We also fitted models for trends and variation in eGFR using the same methods. Residual plots indicated an approximate normal distribution for eGFR and, therefore, no log transformation was required. We allowed mean eGFR and mean rate of change to vary between men and women, but held other parameters the same for both men and women.

Results are reported as mean and standard deviation for log-ACR scale and eGFR and geometric mean of ACR with CIs. The average signal-to-noise ratio is also reported, with the signal being the average change over time, and the noise being the standard deviation of the t-distribution for component within-person variability in any individual ACR measurement.

As previously described,140 these estimates of between- and within-person trends and variability in log-ACR allow us to estimate the proportions of diagnoses of microalbuminuria attributable to the onset of kidney disease (true-positive results), and the proportion attributable to within-measurement variability (false-positive results). Likewise, we are able to estimate the number of people correctly or incorrectly classified with normoalbuminuria (true negative and false negative, respectively). These estimates can be derived only by modelling methods such as simulation. The high variability in any measurement of urine ACR makes it impossible to classify any individual patient in a real data set as a ‘true’ or ‘false’ positive for microalbuminuria. However, simulation allows us to classify individual patients in the simulated data set and, hence, estimate the proportion of patients who would be thus classified in a real cohort. The statistical validity and clinical utility of this approach have been established previously. 134,140

A simulated population of 75,000 individuals generated with the same distribution of age, sex, ACR and duration of diabetes was created. In each patient, the simulated ‘true ACR’ value at each year is calculated from his or her initial ACR value and annual rate of change. The true ACR value is assumed to be unobservable to patient or clinician. The measured ACR value in any given urine sample is simulated from the true value at that time and the model estimate of the within-measurement variability of ACR. We assumed that the screening process would follow the NICE guidance: (i) microalbuminuria defined by ACR > 2.5 mg/mmol for men or 3.5 mg/mmol for women; (ii) testing follows a sequence of tests following on from an initial positive test; (iii) screening starts at 12 years of age. 19 Therefore, an individual of age ≥ 12 years in the simulated population was assumed to be diagnosed with microalbuminuria if his or her first measured ACR was above the threshold, and at least one of two further ACR tests were also above the threshold that year. We then classified a diagnosis of microalbuminuria as ‘true positive’ if the true ACR was above the threshold or ‘false positive’ if the true ACR was below the threshold. Similarly, when a patient was not diagnosed with microalbuminuria – either because the first measured ACR was below the threshold or because the first measured ACR was above the threshold but not confirmed by subsequent measured ACR in the same year – we classified this as ‘true negative’ if the true ACR was below the threshold and ‘false negative’ if the true ACR was above the threshold. We assumed that patients not diagnosed with microalbuminuria continue to be monitored in subsequent years and hence may subsequently be diagnosed with microalbuminuria.

We carried out simulation modelling for annual, biennial (every 2 years) and triennial (every 3 years) screening. We also compared long-term predictions from the model with previously published reports on the prevalence of albuminuria at different ages from large studies of type 1 diabetes. There were insufficient data to study eGFR monitoring in type 1 diabetes because there were no cases of eGFR < 60 ml/minute during ORPS follow-up.

Albumin-to-creatinine ratio

Table 9 reports the parameters used in defining the model that estimates trends and variation in log-ACR. The mean log-ACR of 0.031 (95% CI –0.053 to 0.110) in males and 0.20 (95% CI 0.10 to 0.29) in females is equivalent to geometric mean ACR of 1.03 (95% CI 0.95 to 1.12) and 1.22 (95% CI 1.11 to 1.34), respectively. The average changes in log-ACR per year of 0.035 (95% CI 0.02 to 0.049) in males and 0.047 (95% CI 0.031 to 0.063) in females correspond to average increases in ACR of 4% and 5% per year, respectively. However, the standard deviation around this average, 0.11 on the log-ACR scale, corresponds to wide variation in this rate of change: the estimated 25th and 75th centiles of the change in ACR are a 4% per year decrease in males (3% in females) and 12% per year increase in males (13% in females), respectively. The estimated standard deviation of within-measurement variability in log-ACR is 0.79 (95% CI 0.73 to 0.86), which corresponds to a within-measurement coefficient of variation in ACR of > 100%. The average signal-to-noise ratio is 0.044 in males and 0.059 in females for annual screening, 0.088 in males and 0.120 in females for biennial screening, and 0.13 in males and 0.18 in females for triennial screening.

The results of modelling the rate of identification of microalbuminuria for males and females separately are shown in Tables 10 and 11. Each section of the tables shows the cumulative rate of positive test results, the percentage of those tests that are ‘false positive’ and the percentage of those repeatedly testing negative that are ‘false negative’. Each of these outcomes is shown under assumptions of annual, biennial and triennial screening. Initially assuming annual screening, Table 10 (males) shows the modelled cumulative rate of identified microalbuminuria for the first year after diagnosis of diabetes, and the cumulative diagnoses of microalbuminuria over subsequent years. Over the first 6 years after diagnosis of diabetes, annual, biennial and triennial screening would identify 327, 270, and 244 cases of microalbuminuria per 1000 males with type 1 diabetes, respectively. Over 18 years, annual, biennial and triennial screening would identify 575, 529 and 503 cases of microalbuminuria per 1000 males with type 1 diabetes.

Table 10 (columns 5–7) shows the percentage of those inaccurately identified with microalbuminuria, that is the number of false positives as a percentage of the total number of positive tests. For an initial test, an estimated 53% of microalbuminuria diagnoses would be true positive, and 47% would be false positive, with the inaccuracy attributable to the variability in ACR measurements. We estimate that, by year 6, with a regimen of annual, biennial and triennial screening, 48%, 41% and 38% of individuals identified with microalbuminuria would be false positive, and, by year 18, 27%, 22% and 20%, respectively. The estimated false-negative rates, again with a regimen of annual, biennial and triennial screening, are, respectively, 0.7%, 2.1% and 3.4% by year 18.

Table 11 shows the corresponding estimates for females with type 1 diabetes. Over 18 years, a regimen of annual, biennial and triennial screening would identify 484, 436 and 411 cases of microalbuminuria per 1000 females with type 1 diabetes, 33%, 27% and 24% of which would be false-positive diagnoses. Of those repeatedly testing negative for microalbuminuria, 0.5%, 1.7% and 2.7% would be false negatives.

In summary, in males and females with type 1 diabetes, 48% (95% CI 43% to 53%) and 55% (95% CI 48% to 61%) of patients, respectively, tested annually for microalbuminuria over 6 years would be false positive, that is inaccurately identified with microalbuminuria.

The estimated true- and false-positive test rates reported in Tables 10 and 11 are based on the parameter estimates shown in Table 9. Figure 8 compares estimates from the fitted model with actual data by overlaying the predicted mean log-ACR, adjusted for duration of diabetes and gender, with the actual data derived from the ORPS. Figure 9 also shows the performance of the simulation model in predicting the development of microalbuminuria, macroalbuminuria and regression to microalbuminuria in comparison with the original data on which the model is based.

FIGURE 8.

Mean (dots), interquartile range (vertical bars) and 95th centiles (+) of ACR in male (top) and female (bottom) patients with type 1 diabetes in the ORPS compared with model estimates (dashed lines).

FIGURE 9.

Estimated prevalence of normo-, micro- and macroalbuminuria, and estimated incidence of apparent regression from microalbuminuria to normoalbuminuria (dots), with 95% CIs (vertical lines), in the ORPS compared with model estimates (dashed lines). (a) Normoalbuminuria; (b) microalbuminuria; (c) macroalbuminuria; and (d) regression to normoalbuminuria.

Because ORPS was primarily a study of young people with type 1 diabetes (maximum age 35 by end of follow-up), we compared model predictions to the prevalences of microalbuminuria and macroalbuminuria after diagnosis reported by other large studies of people up to 40 years (Figure 10). 1,51,141–147 The studies used different criteria (e.g. number of confirmatory tests) for diagnosing microalbuminuria.

FIGURE 10.

Proportion of patients who test positive for microalbuminuria or macroalbuminuria by duration of diabetes estimated from parameters obtained from the ORPS (simulation model) compared with studies providing comparator data.

To illustrate the sensitivity of our conclusions to the fitted parameter values, Figure 11 shows how the estimated false-positive rate would vary in hypothetical populations with higher rates of progression of albuminuria. This illustrates the change in the number of false-positive tests when the average rate of change in ACR per year is increased above that observed in our cohort.

FIGURE 11.

Percentage of microalbuminuria diagnoses in male patients with type 1 diabetes that would be false positive under an annual monitoring scheme, when the average increase is as estimated from ORPS (circles; corresponding to columns 5–7 of Tables 10 and 11), or in hypothetical populations in which the average change is 50% (triangles), 100% (pluses) or 200% (crosses) higher than estimated from ORPS. Illustration of the sensitivity of results to key parameters: the results of Table 10, recalculated for four different values of the rate of change parameter. Note that the circles correspond to the analysis in Tables 10 and 11 and other lines correspond to hypothetical subpopulations with increased rates of change.

Glomerular filtration rate

During 463 person-years of follow-up on 386 patients (aged 5–27 years) with a total of 752 eGFR measurements, there were no cases of eGFR < 60 ml/minute. Table 12 reports the parameters for change in eGFR in the cohort. The estimated average decrease in eGFR was 2.1 (95% CI 0.39 to 3.7) ml/minute per year in males and 3.0 (95% CI 1.8 to 4.3) ml/minute per year in females. The estimated between-person standard deviation of the decrease was 3.3 (95% CI 1.8 to 4.7) ml/minute per year, so 74% of male patients and 82% of female patients will have a decreasing eGFR. The within-measurement variability of eGFR has an estimated standard deviation of 22 (95% CI 20 to 23) ml/minute, giving average signal-to-noise ratios for annual, biennial and triennial screening, respectively, of 0.095, 0.19 and 0.29 in males and 0.14, 0.27 and 0.41 in females.

Data

The randomised trial CARDS was a trial of atorvastatin versus placebo in people with type 2 diabetes. The protocol and the characteristics of the study cohort have been described in full elsewhere. 151–153 Briefly, 132 clinical centres in the UK and Ireland recruited 2832 people with type 2 diabetes aged 40–75 years with low-density lipoprotein cholesterol ≤ 4.14 mmol/l, fasting triglycerides ≤ 6.78 mmol/l and at least one of the following conditions: hypertension, retinopathy, micro- or macroalbuminuria and current smoking status. Exclusion criteria included previous coronary heart disease or stroke, previous surgery for peripheral vascular disease, plasma creatinine ≥ 150 µmol/l, and HbA_1c ≥ 12%. Patients were recruited between 1997 and 2001. Safety parameters, including urine albumin and urine creatinine, were measured 1, 2, 3 and 6 months after randomisation, and at 6-monthly intervals thereafter. Albumin and creatinine were measured using a Hitachi 747 autoanalyser (Boehringer Mannheim Corp, Indianapolis, IN) on urine samples that were confirmed free of infection by dipstick testing. At the recommendation of the data monitoring board, the study was stopped in June 2003 following an interim analysis of the primary cardiovascular end point, which showed a highly significant (p < 0.001) benefit of atorvastatin.

Statistical analysis

As described in Chapter 3, we fitted longitudinal models for trends and variation in log-ACR. Log-ACR at diagnosis of diabetes was modelled as a patient-level, normally distributed random effect. Change over time in log-ACR was modelled as a patient-level, normally distributed random effect, correlated with actual value of ACR. Within-measurement variability in log-ACR was modelled using a t-distribution with three degrees of freedom, uncorrelated with actual level of ACR. The model parameters were estimated using WinBUGS (MRC Biostatistics Unit, Cambridge, UK). 139 We also fitted models for trends and variation in eGFR using the same methods as described above, except for the modelling of the within-measurement variability, for which we used a normal distribution. Residual plots indicated an approximate normal distribution for eGFR and, therefore, a log transformation was not required. We allowed mean eGFR and mean rate of change to vary between men and women, but held other parameters the same for men and women. Models were stratified by baseline urine albumin status. We report means and standard deviations of log- and eGFR, geometric means of ACR and the average signal-to-noise ratio as defined in Chapter 3.

Although we did not have data with which to validate individual tests for microalbuminuria in the CARDS database, the model estimates of within-person variability allowed us to estimate the proportion of positive tests that were false positive – that is, attributable to measurement variability rather than to true change in underlying renal function. With the same methodology as described in Chapter 3, we used simulation modelling to calculate the proportion of those identified with true- and false-positive tests in a population similar to that of CARDS, under the assumption of annual screening according to NICE guidelines. Our primary analysis models an annual screening programme as described above, and compares this to 2-yearly screening (biennial) and 3-yearly screening (triennial), while maintaining the NICE protocol for use of confirmatory samples when the first sample measures above threshold. In sensitivity analyses, we assess the degree of change in model parameters that would be required to reach different clinical conclusions. We also used simulation to estimate rates and accuracy of diagnoses of stage 3b kidney disease, defined by eGFR < 45 ml/minute in normoalbuminuric patients.

Albumin-to-creatinine ratio

Table 14 reports the parameters used in defining the model that estimates trends and variation in log-ACR. The mean log-ACR of 0.41 (95% CI 0.28 to 0.53) in men and 0.29 (95% CI 0.11 to 0.48) in women are equivalent to geometric mean ACR at diagnosis of diabetes of 1.51 (95% CI 1.32 to 1.70) mg/mol and 1.34 (95% CI 1.12 to 1.62) mg/mol, respectively. The average change in log-ACR per year of 0.018 corresponds to a 2% per year increase in ACR in both sexes (Figure 12). The standard deviation around this average, 0.022 on the log-ACR scale, corresponds to variation in the rate of change: the estimated 25th and 75th centiles of the annual increase in ACR are a 0.3% and 3.3% increase per year, respectively. The estimated standard deviation of within-measurement variability in log-ACR is 0.85 (95% CI 0.74 to 1.00), which corresponds to a within-measurement coefficient of variation in ACR > 100%. The average signal-to-noise ratios, defined as average change divided by within-measurement standard deviation, are therefore 0.021, 0.042 and 0.063 for annual, biennial and triennial screening, respectively.

FIGURE 12.

Estimated cumulative proportion of patients with microalbuminuria and macroalbuminuria drawn from a cohort based on CARDS using parameters obtained from CARDS (modelled ACR) compared with data obtained directly from the CARDS cohort. The circles and lines are mean (95% CI) of ACR by duration (0–1, 1–2, 2–3, etc.) stratified by gender (men top, women bottom). The numbers are based on the same 535 who made up the model.

Tables 15 and 16 report the estimated rates of identification of microalbuminuria modelled from these estimates. Results for men and women are shown separately (see Tables 15 and 16). The tables show the cumulative rate of positive test results, the percentage of those that are false positive and the percentage of those repeatedly testing negative that are false negative. Each of these outcomes is shown under assumptions of annual, biennial and triennial screening.

Over the first 6 years after diagnosis of diabetes, annual, biennial and triennial screening would, respectively, identify 451, 394 and 365 cases of microalbuminuria per 1000 men with type 2 diabetes. Over 18 years, annual, biennial and triennial screening would identify 582, 534 and 505 cases, respectively.

In an initial screening test, an estimated 78% of those identified with microalbuminuria would be true positive, and 22% would be false positive. By year 6, with assumptions of annual, biennial and triennial screening, respectively, 36%, 30% and 27% of diagnoses of microalbuminuria in men would be false positive, and, by year 18, 35%, 29% and 26%. By year 18, estimated false-negative rates for men were 0.1%, 0.8% and 1.7% for annual, biennial and triennial screening, respectively.

Table 16 shows the corresponding estimates for women with type 2 diabetes. Over 18 years, we estimate that annual, biennial and triennial screening, respectively, would identify 431, 382 and 354 cases of microalbuminuria per 1000 women with type 2 diabetes, 50%, 44% and 40% of which would be false positive. Of those testing negative for microalbuminuria, 0.1%, 0.3% and 0.7% would be false negative.

In summary, in men and women with type 2 diabetes, 36% (95% CI 32% to 42%) and 48% (95% CI 41% to 55%), respectively, of patients tested annually for microalbuminuria over 6 years would be inaccurately identified as having microalbuminuria. If screening intervals were extended to 3-yearly, the corresponding figures would be 27% (95% CI 23% to 31%) and 37% (95% CI 30% to 44%).

Figure 13 shows the degree to which the model reproduces the CARDS data by comparing model predictions to observed data. Figure 14 examines the extent to which the resulting model can be generalised by comparing model-predicted rates of microalbuminuria identification to observed prevalence of microalbuminuria in other cohort studies, with the limitation that different studies have used different criteria (e.g. number of confirmatory tests) for diagnosing microalbuminuria.

FIGURE 13.

Estimated prevalence of normo-, micro- and macroalbuminuria, and estimated incidence of apparent regression from microalbuminuria to normoalbuminuria (dots), with 95% CIs (vertical lines), in the CARDS compared with model estimates using parameters obtained from CARDS (modelled ACR) (dashed lines). (a) Normoalbuminuria; (b) microalbuminuria; (c) macroalbuminuria; and (d) regression to normoalbuminuria.

FIGURE 14.

Proportion of patients who test positive for microalbuminuria or macroalbuminuria by duration of diabetes estimated from parameters obtained from the CARDS (simulation model) compared with other cohort studies in type 2 diabetes. Bars and whiskers denote 95% CIs.

To illustrate the sensitivity of our conclusions to the exact parameter values, Figure 15 shows how the estimated false-positive rate would vary under different assumptions about the average rate of progression of log-ACR. This illustrates the change in the number of false-positive tests when the average rate of change in ACR per year used within the model is increased above that observed in our cohort.

FIGURE 15.

Sensitivity of model predictions to variation in parameter estimates. The lower line of the graph plots the percentage of positive tests that are false (numbers in the fifth column of Table 15) for males with type 2 diabetes with annual monitoring. Other lines represent the percentage of positive tests that are false when the average slope (1.8%) is multiplied by factors of 1.5, 2 and 2.5.

Estimated glomerular filtration rate

Data on 2012 CARDS participants with normoalbuminuria at baseline and 550 CARDS participants with microalbuminuria at baseline were used in obtaining model estimates. Table 17 reports the parameters used in defining the model that estimates trends and variation in eGFR in people with type 2 diabetes and normoalbuminuria. The estimated mean eGFR at 10 years since diagnosis of diabetes is 69 (95% CI 69 to 70) ml/minute in men and 62 (95% CI 61 to 64) ml/minute in women, and the between-person standard deviation is 8.7 (95% CI 8.1 to 9.3) ml/minute. The estimated average change in eGFR per year is 0.26 (95% CI 0.13 to 0.40) ml/minute in men, and 0.24 (95% CI 0.061 to 0.42) ml/minute in women. The estimated between-person standard deviation of the reduction per year is 0.50 (95% CI 0.28 to 0.69) ml/minute. The estimated standard deviation of within-measurement variation in eGFR is 4.9 (95% CI 4.8 to 5.1) ml/minute. The signal-to-noise ratio is therefore 0.053 for annual monitoring, 0.11 for biennial monitoring and 0.16 for triennial monitoring. Table 17 also shows the corresponding parameter estimates for participants with normoalbuminuria and microalbuminuria at baseline. Figure 16 illustrates the observed mean eGFR by duration of diabetes compared with the modelled estimates of progression and shows how the modelled rate of change matches the original cohort data.

FIGURE 16.

Observed mean (dot) and 25th and 75th centiles (vertical lines) of eGFR in men (top) and women (bottom) with type 2 diabetes and normoalbuminuria, and model estimate of trends (dashed line).

Table 18 shows the modelled estimates of the proportion of normoalbuminuric men classified as having stage 3b kidney disease (eGFR < 45 ml/minute) over time for annual, biennial or triennial eGFR screening. In annual screening, for example, there would be 20 men identified with stage 3b kidney disease (95% CI 16 to 26) per 1000 after 18 years, and 67% (95% CI 53% to 83%) of these would be false positive. Similar estimates were obtained for women (data not shown).

Model structure

The type 1 diabetes simulation model is built around a series of ‘key health states’ and ‘interlinking dependencies’. Figure 17 shows the different states of the model and how a type 1 diabetes patient can progress through its different states over their lifetime. The aim of this model is to estimate the first occurrence of a cardiovascular event (including stroke or a myocardial infarction) for type 1 diabetes patients, either with or without end-stage kidney disease and death, in order to estimate lifetime outcomes and quality-adjusted life expectancy associated with different screening programmes. The arrows indicate dependencies linking key health states. For example, the occurrence of kidney disease elevates the risk of myocardial infarction and death.

FIGURE 17.

Diagram depicting the structure and relationship among equations, disease progression, screening and treatment steps used in the type 1 diabetes simulation model.

Data sources

The data sources used in the model construction are summarised in Table 19 and the functional form and coefficients used in each numbered equation are described in Table 20. Two equations for CVD-related death are used to capture both the acute and long-term impact on mortality of this disease. The key studies used to construct this model report time to first event and, therefore, the model focuses on these states (i.e. there are no equations to estimate second or subsequent events). 45,167

Translation of reported results

When published equations to estimate transition rates were not available (e.g. equations describing the transition to CVD and kidney disease), we obtained data about the underlying hazard by estimating survival times from published graphs. This involved using a program called Dagra 2.05 (Blue Leaf Software, Hamilton, New Zealand) to trace out the reported survival curves and export them as time-to-event data. 172 Using this approach we were able to replicate the cumulative hazard of a CVD event for both the conventional and the intensive treatment groups in the DCCT. 45 This involved fitting a parametric proportional hazards model. Tests were consistent with intensive therapy having a proportional effect on the hazard, reducing the risk of a CVD event when compared with the conventional treatment group.

A Weibull proportional hazards regression model (equations 1 and 2) was used to build the simulation model for type 1 diabetes, as this had previously been shown to fit the data well for people with type 2 diabetes. Coefficients used in the Weibull model included the following: the hazard ratios of microalbuminuria and albuminuria based on reported estimates involving a Cox model. 45 As parameters for the impact of age and sex are not reported in the literature, we used the parameters estimated from the UKPDS 68. 68 Patients were assumed to have a 70% probability of having a myocardial infarction and a 30% probability of having a stroke when patients had a cardiovascular event, based on two large diabetes trials. 68,173

For myocardial infarction and stroke, we used Hayes and colleagues’ logistic regression (equations 3–5) to estimate the higher risk of mortality 1 month post event and their Gompertz proportional hazards regression model (equations 6 and 7) to estimate the risk of mortality for surviving patients. 166 A Gompertz parametric form produces an exponential rise in death rates in older patients and is widely used to model human survival. 174

Proportional hazard ratios for age and type 1 diabetes and age and gender were used in the myocardial infarction and stroke logistic regression model, respectively. A sex-specific Gompertz proportional hazards model was used to represent patients with history of a CVD event and estimate the probability of survival post event.

Patients were at risk of progressing to the end-stage renal disease (ESRD) state in the model only if they had micro- or macroalbuminuria. We estimated a Gompertz proportional hazards model (equations 8 and 9) to determine the incidence of patients progressing to ESRD from published age- and sex-specific cumulative incidence graphs,167 using Dagra to estimate the data. These techniques were also applied to the reported Kaplan–Meier survival curves for cardiovascular events and all-cause death for type 1 patients with ESRD to determine rates of progress to a CVD event or death, respectively (equations 10 and 11). 168

Model testing and validation

All estimated proportional hazards models were tested for accuracy by comparing the estimated survival curves with the published survival curves, ensuring that the model provided the best fit. We carried out a validation exercise comparing simulation estimates of life expectancy among type 1 diabetes patients with reported all-cause mortality rates among a general non-diabetic population from an analysis of the UK General Practice Research Database (GPRD). 175

To carry out this validation exercise, life expectancy estimates were derived using a hypothetical group of 10,000 patients, who were simulated for 85 years, with a mean age at entry into the type 1 simulation model of 15 years.

Estimated mean life expectancy separated by sex and treatment group is shown in Figure 18. Men have a mean life expectancy of 66.0 (95% CI 60.7 to 71.3) years in the intensive treatment group and 55.4 (95% CI 47.8 to 63.0) years in the conventional treatment group. Life expectancy for women was estimated at 63.6 (95% CI 56.4 to 70.7) years for the conventional treatment arm and 72.2 (95% CI 68.7 to 75.6) years for the intensive treatment arm.

FIGURE 18.

Type 1 simulation outcomes for life expectancy separated by sex. Data extracted from Soedamah-Muthu 2006. 175

These results were compared to those of the UK type 1 diabetes GPRD study: life expectancies lay within one standard deviation of those obtained by the external study. 175 Women have a higher life expectancy across all treatment groups, which is reflected in the life expectancy of the general population. 176 Women with type 1 diabetes have a higher risk of mortality, and this is reflected in the literature. 175,177–181

To examine the validity of these model estimates, life expectancy was calculated using life table methods by adjusting the UK life table to take into account the higher relative risk of mortality among people with type 1 diabetes. The relative risks for this analysis were taken from a study that selected 7713 type 1 diabetes patients and compared them to 38,518 non-diabetic patients. 175 The reported overall hazard ratios of 3.3 for men and 4.5 for women were applied to UK life tables for 2007–9176 by multiplying the age-specific mortality rates of the general population by the appropriate hazard ratios at 5-yearly intervals. This gave figures for life expectancy at birth of 63.8 years for males and 65.6 years for females. Compared with the mean life expectancy of 77.7 years in the general population of the United Kingdom,176 our simulation model, and the external study,175 suggests a loss of approximately 14 years of life attributable to type 1 diabetes.

Uncertainty

There is uncertainty around the estimates of health outcomes and costs derived from any simulation model. Two forms of uncertainty are addressed within this modelling exercise: uncertainty in the individual patient predictions generated by the system of model equations and uncertainty in the estimated parameters of those equations. We addressed the former – the random variation between simulations – by carrying out large numbers of Monte Carlo replications until the estimates of the outcome of interest were stable. To model CIs around estimates we sampled from normal distributions using bootstrapping methods with different sets of regression coefficients around key parameters, based on the standard errors of coefficients reported in the literature. These analyses allow us to report CIs for the mean difference in QALYs and costs.

To examine the degree to which different elements of the model contribute to uncertainty, 1000 draws from normal distributions for the major parameters of the model were taken and used to estimate 95% CIs around life expectancy. In each draw, the mean value was used in the model and the variance calculated using reported measures of standard error. The same model simulations were repeatedly run with an increasing number of equations subject to parametric uncertainty.

Figure 19 shows the increased uncertainty in the life expectancy estimates in the form of 95% CIs as uncertainty for each parameter is included incrementally. This comprises, reading from left to right:

• point estimates

• CIs when just the CVD incidence parameters were assigned distributions

• CIs when CVD and ESRD incidence parameters were adjusted

• all parameters were assigned distributions around their point estimates.

FIGURE 19.

Uncertainty in the life expectancy estimates (95% CIs) when parameters were assigned distributions incrementally (see text for a detailed explanation).

This was calculated for the life expectancy for the whole population, which included both conventional and intensive treatment arms. Figure 18 shows that the CVD incidence parameters contributed the most uncertainty in the current version of the model.

Limitations

There are a number of limitations that require acknowledgement in our model. Multiple CVD events are not included in the model, as the source equation from the DCCT reports only the first occurrence of a CVD event. Second, although the DCCT/EDIC study has a comparatively long follow-up of around 17 years, this is a relatively short period when compared with a patient’s overall lifetime. Hence, there is considerable uncertainty surrounding the lifetime extrapolation relating to the data. This is reflected in the reported uncertainty surrounding the life expectancy estimates when allowing for parameter uncertainty. When accounting for uncertainty in our model, we assumed independence between: (1) variables in the same risk equation; and (2) different risk equations. If the covariance between parameters is significant, it could impact on the level of uncertainty.

Our model also does not take into account recently published studies of risk in type 1 diabetes patients, such as two recently published equations for CVD risk from Sweden182 and the United States. 183 Furthermore, the transition probabilities for ESRD patients progressing to CVD and death have been estimated using non-type-1 diabetes populations168 because of the lack of published studies in the literature.

Model structure

The UKPDS outcomes model is a computer simulation model for forecasting quality-adjusted life expectancy and other outcomes for people with type 2 diabetes. It was developed using patient-level data from the UKPDS and was primarily intended as a tool to assist in the evaluation of new interventions to manage patients with type 2 diabetes. 68

The outcomes model involves probabilistic discrete-time computer simulation and is based on an integrated system of parametric proportional hazards risk equations. These were estimated over a median period of 11 years from diagnosis of diabetes using individual patient data. 68 The model includes both macrovascular (e.g. myocardial infarction, other ischaemic heart disease, congestive heart failure, stroke) and selected microvascular (e.g. blindness) complications. It can be used to assess the impact of the disease on morbidity and mortality, and can also be used to estimate health-care costs associated with the disease. The existing risk factors and complications covered by the model are summarised in Figure 20. These equations are used to estimate the probability of occurrence of different complications given risk factors such as a patient’s age, sex, duration of diabetes, systolic blood pressure, HbA_1c, lipid levels and smoking status. The relative risk for each of these factors by type of complication is also reported in Figure 21.

FIGURE 20.

Summary of model equations for the type 2 diabetes simulation model (based on the UKPDS outcomes model)68 showing event-related dependencies and hazard-to-odds ratio for each risk factor.

FIGURE 21.

Algorithm used for the UKPDS model. Source: UKPDS 68. 68

Simulation algorithm

Figure 21 comprises an algorithm that illustrates the sequence of modelling events for the existing model. In each cycle, the probability of experiencing each of the diabetes-related complications and death is calculated using the estimated risk equations. At the beginning of each cycle, the order of the equations is randomised to account for competing risks (i.e. if a patient dies in any cycle of the model, he or she cannot experience an event defined by equations lower in the order). The probability is compared with a random number ranging from zero to one from a uniform distribution to determine whether an event occurs. If the model predicts that a death has occurred, the years lived and quality-adjusted years lived are calculated. If the patient is predicted to have survived that cycle, the risk factor equations are used to update the current risk factor values (e.g. systolic blood pressure) and these are carried forward to the next cycle of the model along with the updated event history.

Although the model can estimate the risk over time of each complication and death in this study, it will primarily be used to estimate life expectancy and QALYs.

Reduced rates of complications may reduce health-care costs, producing savings that may offset some of the costs of improving treatment. The sources of costs and utility values are discussed in Chapter 6. To capture these benefits, the simulation model was also used to estimate lifetime health-care costs of diabetes-related complications. Non-diabetic morbidities are not specifically included in the simulation model, although non-diabetic mortality is included.

The methods described above to deal with uncertainty in the type 1 simulation model were also used to address the uncertainty in the type 2 model. The aim, as before, was to facilitate reporting of the CIs for the mean difference in QALYs and costs.

How the model has been adapted for this project

The population parameters of the UKPDS outcomes model were adjusted to more accurately reflect the type 2 diabetes population in the UK, including ethnic distribution. The data used to carry out this adjustment were extracted from the UKPDS,184 including data that compared intensive glucose therapy to conventional dietary therapy among 4209 newly diagnosed type 2 patients. Patient characteristics used were mean age, proportion of males, body mass index, HbA_1c, blood pressure and cholesterol, and ethnic distribution.

The simulation model reported in Chapter 445 was integrated into the UKPDS outcomes model for the purposes of this project. This involved integrating programming codes in Stata version 11 that classify patients according to whether they have either hypertension or kidney disease. Increased risk of kidney disease or risk of CVD and mortality was incorporated into the UKPDS outcomes model. 169,185 Patients detected as having kidney disease were assumed to be treated with ACEi and A2RB (see Chapter 2 for further discussion) and benefit from the mortality reductions was quantified for these treatments.

The utilities derived from a meta-analysis for diabetes patients with no complications, myocardial infarction, stroke and end-stage kidney disease were integrated into the model,170 whereas for all other health utilities (ischaemic heart disease, congestive heart failure, amputation and blindness) the original values as reported in the original UKPDS outcomes model were used. 68 Costs for the analysis of renal screening were also integrated into the model, and these will be reported in Chapter 6.

Economic evaluation

Incremental cost-effectiveness analyses were carried out for each of the various screening strategies. The incremental net cost and net cost-effectiveness were calculated in relation to the comparator, and expressed as a ratio for each strategy. As the economic evaluation perspective was that of the health-care purchaser, only direct health service costs were included. Both costs and effects were discounted at 3.5% per year for the first 30 years and 3% subsequently, in line with current guidelines. 186

All outcomes are expressed as QALYs, as this measure captures both increases in life expectancy and improved quality of life resulting from prevention of complications, providing a composite outcomes measure of fatal and non-fatal events that permits comparisons among a wide range of interventions.

In the main analyses, we vary the screening interval between 1 and 10 years, initially using 5-year and then 1-year increments. Where possible, results are reported as mean values and standard deviations, or mean differences with CIs, and as ‘cost-to-clinical-effectiveness’ ratios. The effect of uncertainty surrounding costs and the utility values used in the study were examined using sensitivity analyses.

Model simulations

Using the models outlined in Chapter 5, simulations were undertaken at a patient level. The simulations include (as treatment effects) estimates of the proportion of true-positive/true-negative/false-positive/false-negative tests and the consequences of these outcomes in terms of additional benefit and cost of treatment. These models were used to examine the impact of various monitoring policies on the incidence of important health states (e.g. CVD). The results presented are the average of 1000 simulations, which stabilise the expectation of different individual Monte Carlo simulations.

Resource data and costs (type 1 and the United Kingdom Prospective Diabetes Study outcomes model)

The economic evaluation was carried out from the perspective of the health-care provider and, therefore, the following costs were included in the analysis: cost of monitoring; cost of further investigations; cost of treating diagnosed kidney disease in people with diabetes; and the cost of subsequent complications.

All costs included were obtained from published UK sources (individually referenced below) and the values and sources of cost data are described in the following sections. Costs are expressed in 2011 pounds sterling and data from previous years have been adjusted to 2011 value using the Hospital and Community Services pay and price index. 186

Table 21 summarises the main sources of information on therapy unit costs, implementation unit costs and immediate (and subsequent) costs of CVD complications. As described in previous chapters, a maximum of three urinary ACR tests are taken during each year of monitoring. To determine whether a patient with normoalbuminuria has progressed to microalbuminuria or macroalbuminuria, two positive tests are required before patients are confirmed to have progression of kidney disease requiring intensification of treatment. The cost of the ACR test is estimated at £2.16 according to the Chronic Kidney Disease Costing Report, published by the National Institute for Health and Care Excellence in 2008. 187 ACR tests are administered by the practice nurse, and we approximate each test to take 10 minutes of the nurse’s time. According to the Unit Costs for Health and Social Care 2010,188 a 10-minute clinic visit costs £13.50. Once a patient is confirmed to have progressed from normoalbuminuria to microalbuminuria, a GP is consulted to discuss what action to take and the administering of an ACEi, which we estimate to take around 5 minutes of a GP’s time (£15.41). 188

If patients are confirmed to require treatment intensification through use of the ACR test, they are recommended to start an ACEi, or A2RB if they cannot tolerate ACEi. Pursuant to Prescription Cost Analysis England 2010,189 ACEi and A2RB cost £27.74 and £191.12 per year, respectively.

Furthermore, UKPDS 65 provides estimates of inpatient and outpatient costs for complications related to type 2 diabetes in the UK for fatal/non-fatal myocardial infarction, fatal/non-fatal stroke, no complications and a history of cardiovascular events. 190 These were included in the model to estimate potential cost offsets from lower rates of CVD.

End-stage kidney disease costs were divided into two parts. The National Costing Report estimated a mean annual cost of £29,800 per patient for haemodialysis, which is the main mode of renal replacement therapy in the UK. 192 A previous study, using UK guidelines, estimated the additional costs associated with diagnosis to be £11,479,191 which takes into account the cost of initial investigations, total referral cost and follow-up. For patients who move to end-stage kidney disease, the model includes a cost of around £40,000 in the first year (i.e. combining both diagnosis and dialysis costs), and £29,800 for every subsequent year they survive in that state.

Utility values

The economic evaluation for screening people with either type 1 or type 2 diabetes uses QALYs to adjust length of life for quality of life. A QALY value of 1 is equivalent to being in full health and 0 is equivalent to death. The values used are based on a recent meta-analysis. 170 Patients without complications were assumed to have a QALY of 0.81, reflecting that people with diabetes are on average in less than full health. In the type 1 simulation model, which uses a composite cardiovascular event, we assumed two-thirds had a myocardial infarction (reducing their utility to 0.75) and one-third of patients had a stroke (reducing their utility to 0.59) over their remaining lifetime. ESRD patients are assumed to have a utility of 0.48. We also incorporated these utility values into the UKPDS outcomes model, and used the original utility values for all other diabetes complications (Table 22).

Model baseline parameters

For the type 1 diabetes evaluation, a cohort of 10,000 patients was progressed through the type 1 simulation model (see Chapter 5). Screening for albuminuria commenced when patients were aged ≥ 12 years, in keeping with UK guidelines. The frequency of the screening was varied from annual to 10-yearly using the model described in Chapter 3.

A similar evaluation was undertaken for screening in patients with type 2 diabetes using the adapted version of the UKPDS outcomes model described in Chapter 5. For these simulations, the mean age of patients was assumed to be 62 years and the baseline characteristics were based on the patients in the post-study monitoring phase of UKPDS. Following UK guidelines, screening occurred annually beginning from time of diagnosis for type 2 diabetes patients. Again, the incremental cost-effectiveness ratios (ICERs) are calculated for different screening intervals from annual to 10-yearly using the model described in Chapter 4.

Analysis

Results are reported as mean values and standard deviations, or mean differences with CIs, and as cost-to-clinical-effectiveness ratios. To provide a visual representation of the results, the costs and health outcomes were mapped onto the cost-effectiveness plane and reported as acceptability curves. 193 The effect of uncertainty surrounding some aspects of cost and the utility values used in the study were examined using sensitivity analyses.

A series of sensitivity analyses were carried out to examine whether the results in the main analysis were robust to a series of assumptions:

• the effect of costs: higher ACR test cost

• the effect of changing progression values of ACR in the simulation models outlined in Chapters 3 and 4 to fixed values at the upper and lower end of the reported 95% CIs

• the impact of assumptions about glycaemic control: estimating ICERs if all patients were simulated over a lifetime in two scenarios using each of the conventional and intensive treatment groups of the DCCT.

Screening people with type 1 diabetes: results of the type 1 model

The characteristics of the patients in the type 1 cohort are summarised in Table 23. For the base-case scenario over a lifetime, around 40% of patients are estimated to have a cardiovascular event, and 25% progress to end-stage kidney disease. Life expectancy is estimated at 63.1 years, which is similar to the figure obtained in Soedamah-Muthu and colleagues’ study on type 1 diabetes patients in the UK. Females have a higher risk of mortality, and this is reflected in the literature. 175,177–181

The base-case scenario examined was annual screening of newly diagnosed type 1 diabetes patients ≥ 12 years old; the model determined total cost and QALYs gained over the duration of the simulation for the base case and for screening at greater intervals. Results of the simulation using the type 1 model are shown in Table 24. The columns of the table represent an incremental comparison of reducing the screening interval from 5 years progressively to 1 year. The costs associated with screening increase in proportion to its frequency, with the cost of annual screening being around five times greater than 5-yearly screening. The benefit in terms of incremental QALYs gained increases at a slower rate and hence the ICER increases from £3612 (£6586) to £9601 (£34,112), but all of these ratios are well within NICE’s QALY threshold of £30,000 per QALY.

The probabilistic uncertainties surrounding these estimates are shown in Figure 22, which illustrates a cost-effectiveness plane comparing biennial with annual screening intervals, and Figure 23 shows the corresponding cost-effectiveness acceptability curve. Figure 23 shows the probability that annual screening is cost-effective based on different thresholds for a QALY. The probability the intervention is cost-saving is around 25%, and it has around an 80% chance of being below the cost-effectiveness threshold of £30,000.

FIGURE 22.

Cost-effectiveness plane when comparing biennial screening intervals with annual screening intervals for patients with type 1 diabetes.

FIGURE 23.

Cost-effectiveness acceptability curve for biennial screening intervals vs. annual screening intervals for patients with type 1 diabetes.

Screening people with type 2 diabetes: results of the type 2 model

Using this model, 1000 patients were simulated for 30 years; the mean age of patients newly diagnosed with type 2 diabetes was 62 years. Baseline characteristics are reported in Table 25. According to UK guidelines, screening occurred annually beginning from time of diagnosis for type 2 diabetes patients.

Results of the cost-effectiveness analysis are shown in Table 26. The differences in costs and QALYs between biennial and annual screening were reported as £209 (standard deviation £309) and 0.42 (standard deviation 0.12), resulting in an overall ratio of £606 (standard deviation £1782), which makes the intervention highly cost-effective. When the screening interval is increased beyond 5 years, there is a minimal difference between QALYs, suggesting that increasing the screening interval further has little impact on QALYs and that, potentially, a ‘sojourn period’ is reached.

Similarly, the differences in costs reflect the decreasing costs of ACR tests, but the increasing number of patients being treated for hypertension and kidney disease. Figures 24 and 25 represent the uncertainty surrounding the cost-effectiveness of renal screening in patients with type 2 diabetes. The cost-effectiveness plane is illustrated in Figure 24, showing a positive incremental benefit and incremental cost for all observations. This suggests that, although annual screening provides added benefits in terms of QALYs, it comes at a higher cost. Figure 25 shows a very high probability (97%) of annual screening being below the cost-effectiveness threshold of £30,000 per QALY. Based on these results, annual screening appears to be a cost-effective option.

FIGURE 24.

Cost-effectiveness plane when comparing biennial screening intervals with annual screening intervals for patients with type 2 diabetes.

FIGURE 25.

Cost-effectiveness acceptability curve for biennial screening intervals vs. annual screening intervals for patients with type 2 diabetes.

Sensitivity analysis results

The sensitivity analysis results for both type 1 and type 2 models are shown in Tables 27–29. We compared biennial screening intervals to annual screening intervals in all our analyses to help determine what factors drive our model. The results of these analyses are reported below.

Albumin-to-creatinine ratio progression

Changing the values for progression of ACR in the simulation models outlined in Chapters 3 and 4 had a large impact on the results for both diabetes models. Changing the ACR progression variables to their lower 95% CI value saw a small proportion of patients progress to requiring treatment and thus ESRD, resulting in a reduction in the difference in costs and QALYs. The ICER calculated of £40,801 lies outside NICE’s recommendation, suggesting that annual screening is not cost-effective. The UKPDS model shows a similar result of £51,324 per QALY, rendering this particular scenario cost-ineffective.

When adjusting the variables to the upper 95% CI limit, annual screening was very effective in identifying patients because of their quick progression to requiring treatment, and thus ESRD. Hence, the ICERs for both models were lower when using the higher CI limit than when using the lower CI limit: £26,945 and £25,447 per QALY for type 1 diabetes and type 2 diabetes patients, respectively.

Figures 26 and 27 show the results of modelling progression of the ACR over time from diagnosis. The modelling integrates the equations that estimate the true progression of renal function in the simulation models for type 1 and type 2 diabetes patients. The simulation uses estimates of the upper and lower 95% CI limits of coefficients for the observed mean rate of ACR progression to present two alternative scenarios for progression. The figures show that, for the upper 95% CI values, progression occurs very quickly and, therefore, annual screening would identify more patients with microalbuminuria than biennial screening. However, it might be too late to treat some of these patients before they develop end-stage kidney disease.

FIGURE 26.

Sensitivity analysis for rate of ACR progression based on running simulations for the 95% CI of the average rate of progression for type 1 diabetes.

FIGURE 27.

Sensitivity analysis for rate of ACR progression based on running simulations for the 95% CI of the average rate of progression for type 2 diabetes.

In comparison, modelling with mean ACR progression estimates based on the lower limit of the 95% CI shows that the progression of kidney disease is very slow, and thus neither annual nor biennial screening intervals would reliably identify the small number of patients progressing to microalbuminuria. Therefore, it is likely to be less cost-effective than the upper 95% CI limit scenario, and possibly cost-ineffective.

Main conclusion

Annual screening appears to be a cost-effective option for both type 1 and type 2 diabetes compared with biennial screening and other health-care interventions. For type 1 diabetes, screening produces benefit at a cost, but this cost is well below accepted thresholds used for other types of health care. For type 2 diabetes screening, the cost-effectiveness ratio is highly favourable. The data on which our analysis was based did not provide information about more frequent screening intervals than 1 year; however, the results did indicate that the benefit might be greater if the underlying rate of progression of ACR is higher than the average in the population.

Comparisons with previous literature

We have found renal screening for people with type 1 and type 2 diabetes to be cost-effective in a UK context with cost-effectiveness ratios that would compare favourably with many other funded health interventions. With regard to the benefits, it would appear that screening produces comparable or greater outcomes to that observed in a general population screening. For example, it has been previously estimated that biennial mammography for women aged 50 produces less than 1 additional month of survival and Pap smears less than 3 months. 194 Here, particularly with type 2 diabetes, the gains in QALYs range from a few months to over half a year.

The findings of this study are also consistent with two studies of the cost-effectiveness of blood pressure treatment in type 2 diabetes. A study based on an analysis alongside the UKPDS indicated that tight blood pressure control policies (including use of ACE inhibitors) produces several months’ increase in QALYs at a cost of only £300 per QALY. 195 Similarly, an analysis of the ADVANCE study showed treatment with a fixed combination of blood pressure therapies (again, including ACE inhibitors) produced significant increases in life expectancy and was cost-effective. 196

The higher cost-effectiveness ratio in screening type 1 diabetes patients is a result of a number of factors impacting on relative outcomes and costs of screening among these groups. First, when benefits are not discounted, the gains in QALYs of screening are similar in type 1 diabetes and type 2 diabetes, but the long duration between the commencement of screening and when the majority of patient experience cardiovascular events significantly reduces discounted QALYs. Another reason for the differences across these patients is the effect of treatment on ACR. In the case of type 2 diabetes, we have assumed an average 52% reduction in ACR (see Chapter 2), whereas in type 1 diabetes the effect is 32% (see also Chapter 2). The greater incremental effect in type 2 diabetes is a result of screening older patients who are at a higher risk during the entire screening period and the treatment being more effective in slowing progression to renal failure.

The average difference in lifetime screening costs is also much higher for type 1 patients, as, currently, the guidelines for screening recommend annual screening from diagnosis with either type of diabetes. Assuming that the age at diagnosis in type 1 diabetes is around 15 years, whereas in type 2 diabetes it is 62 years, patients with type 1 diabetes will have many more tests over their lifetime.

Further, these screening costs will be occurred at a constant rate, whereas any cost offsets associated with CVD or kidney disease are likely to occur after the age of 40 years. In contrast, type 2 diabetes patients are generally already at risk of complications at diagnosis and, therefore, the cost offsets will occur much sooner.

Clinical and research implications

While these simulations provide useful insights relating, in particular, to the additional benefits of universal treatment over screening, it is difficult to draw firm policy conclusions regarding the merits of their adoption in practice because of the limited number of studies reporting continued long-term ACE inhibitor use. For example, although the ADVANCE study randomly allocated all patients with ACE inhibitors in combination with another blood pressure medication, it involved screening patients’ toleration for this treatment through an active run-in. 61 Around 14% of patients withdrew during the 6-week active run-in period and the study had a mean follow-up duration of only 4.3 years. Therefore, it is necessary to collect more evidence surrounding the benefits and compliance associated with much longer periods of use before universal treatment can be considered as a practical policy option. This should be examined in future work.

A screening interval of 1 year appears optimal based on evidence of progression to renal failure of typical diabetic patients in the UK. The sensitivity analysis suggests that rates of progression would have to be substantially higher or lower than the average to have an important effect on our results. For example, within some ethnic groups (e.g. the south Asian population; see Table 5) rates of progression towards diabetic kidney disease may be substantially higher. 44,197 Further empirical work could help to establish whether more frequent monitoring of ACR might be justified, or whether alternative technologies might be required to identify those at risk. Similarly, if individuals at low risk of progression could be prospectively identified, then for future implementation of protocols for personalised care it might be possible to consider screening at less frequent intervals.

Strengths and limitations

Our type 1 diabetes model is subject to limitations, such as the exclusion of other major diabetes-related complications (peripheral vascular disease, blindness and neuropathy), because of the timeframe and available secondary data sources. Nevertheless, our model appears to make predictions for type 1 diabetes patients’ life expectancy that match another large type 1 diabetes mortality study in the UK. 175

The type 1 model includes important sources of major costs, including renal replacement therapy for patients who have either true-positive or false-negative test results but progress to end-stage kidney disease. Decrement of quality of life through occurrence of an ACE-related cough is not included, as the model assumes that these individuals are treated with A2RB. We have also, based on the results in Chapter 2, not added a disutility for treating patients with a false-positive test. We did not carry out a sensitivity analysis to explore the impact of the reduction in costs of A2RB with a move to generic prescribing, as this is a relatively small part of the overall costs and, in any case, would further increase the cost-effectiveness of annual screening. The impact of non-adherence to therapy is not included in this model.

The type 1 diabetes model is a synthesis of multiple data sources from a variety of different countries. Although the majority of sources are based on type 1 diabetes populations, some include type 2 diabetes patients and non-diabetes-related patients as well, which is a limitation when trying to inform NHS clinical practice for a UK-based screening programme. However, owing to the lack of published data specifically regarding type 1 diabetes populations, this was unavoidable. Further validation does, however, need to be conducted to reduce the uncertainty surrounding our estimates.

Strengths and limitations

Our analyses are based on statistical modelling rather than trials directly comparing different monitoring strategies. The inefficiency and potential invalidity of randomised trials for diagnostic strategies have been discussed elsewhere198 and apply equally to screening and monitoring strategies, making modelling studies a frequent necessity. To overcome the issues associated with modelling, we have used external validation (such as Figures 10 and 14) and sensitivity analyses (such as Tables 26 and 27). Compared with the original protocol, these sensitivity analyses have been limited by the nature of the available data sets, but we have established robustness of the models over time and to key sources of uncertainty as far as possible (discussed in more detail below).

For our analyses in type 1 diabetes, we benefited from an observational cohort study, the ORPS, with up to 20 years of follow-up; however, this included only children under 16 years at diagnosis of diabetes, and hence to ages no higher than 35 years during follow-up. Our analyses for type 2 diabetes benefited from a middle-aged, adult population, with a variety of durations of diabetes at recruitment, but the CARDS data extended to relatively few years of follow-up, and was collected in a clinical trial setting rather than routine care. In both cases, we addressed these limitations by comparing projections from our models with observations from other large studies with a wide range of ages and durations of type 1 and type 2 diabetes. Figures 10 and 14 show our models performing well against these external data; there is generally greater variation among different studies than there is discrepancy between our models and the observed data. The interpretation of these figures is complicated by the use of different protocols for classifying microalbuminuria in different studies, which is likely to exaggerate the discrepancies between studies and the model predictions. In Chapter 5, further external validation is carried out when these models are extended to use long time horizons, to consider multiple morbidities and mortality risks, and to allow for interactions between different complications of diabetes, as recommended by the ADA Consensus Panel on computer modelling. 199 As further information from trials becomes available about the extent to which these assumptions are valid, the results of this analysis will need to be revisited.

We have closely followed the ADA guidelines for diabetes modelling whenever relevant. 199 In particular, we have described our model with transparency; we have carried out external validations of the models in Chapters 3 and 4 and based our cost-effectiveness modelling of type 2 diabetes on a model with previous external validation; we have used multiple simulation runs to reduce Monte Carlo uncertainty and to quantify statistical variability; we have taken into consideration the long time horizons, multiple organ systems and multiple therapies relevant to lifetime modelling of diabetes and their impact on quality and length of life; and we have stated the cost perspective of our analyses.

We have not addressed the health economics of eGFR measurement in this report. The low prevalence of impaired eGFR in our type 1 data set limits to the degree to which such analysis is possible, but, more importantly, it also shows that eGFR measurement has only a secondary role to urine albumin measurement in screening for impaired renal function in type 1 diabetes. Although we base this on a relatively young cohort with type 1 diabetes, it appears to be confirmed by follow-up of the DCCT/EDIC cohort, in which declining eGFR is seen only in those with macroalbuminuria. 150 In type 2 diabetes, we also find that declining eGFR is a feature of patients who already have microalbuminuria or worse (see Table 16).

We could not extend our modelling to look at testing intervals of < 1 year. Both our data sets collected data at annual intervals, and studies that collect markers of diabetic kidney disease more frequently are likely to have shorter follow-up. Owing to the low signal-to-noise ratio and the difficulty in distinguishing change from underlying variation in the measure of both eGFR and ACR with a short interval between measurements (see Chapters 3 and 4), we deemed that evaluating intervals shorter than 1 year would not be justified. Future work might be focused on identifying individuals at greater risk of progression and draw data from such groups to re-estimate parameters for rates of progression and interindividual variation.

Interpretation

The cost-effectiveness results reported in Chapter 6 can be viewed as quantifying overall impressions from the systematic review and analyses of cohort data. The simulation modelling (see Chapters 3 and 4) found that false-positive diagnoses of microalbuminuria are frequent and false-negative classification of normoalbuminuria comparatively rare. The former would become less frequent, and the latter more frequent, if biennial or triennial screening were adopted in place of annual screening. The cost-effectiveness models of Chapter 5 consider how these two types of error – resulting, respectively, in overtreatment or in missed or delayed opportunities to treat – might translate into financial costs for providers and health costs for patients. In type 1 diabetes, the benefit of ACEi and A2RB treatment is strongly evident only in patients who have microalbuminuria (see Chapter 2). Our cost-effectiveness model for type 1 diabetes reports estimates and surrounding uncertainty for annual screening that are well within accepted thresholds of cost-effectiveness, and comparable to similar procedures.

We have not, because data remain unclear, included any estimates of treatment at even earlier stages of renal impairment, for example with hyperfiltration or with levels of albumin excretion in the upper tertiles of normoalbuminuria. However, in type 2 diabetes, treatment with ACEi and A2RB appears to have renal benefit even in patients who do not yet have microalbuminuria and, therefore, the overtreatment resulting from false-positive diagnoses in annual screening results in benefits as well as costs. Screening for microalbuminuria in type 2 diabetes appears, therefore, highly cost-effective.

Similarly, the findings from the statistical modelling in Chapters 3 and 4 are consistent with the overall findings of the research. For example, the short-term variability of ACR and eGFR measurements is high compared with the average annual change (see Tables 9 and 14). As a result, many diagnoses of microalbuminuria or decreased eGFR are a result of measurement error rather than true change (see Tables 10, 15 and 17). This is consistent with the previous findings, for other chronic conditions, that annual screening leads to high rates of false-positive diagnoses. 134,135 Further, because of the extreme variability of ACR measurements (despite log transformation; illustrated by the 95th centiles in Figures 8 and 13), even the practice of confirming microalbuminuria across multiple tests does not fully prevent high rates of false-positive diagnoses. Our models predict that the prevalence of microalbuminuria does not increase indefinitely with age until almost the whole population has diabetic kidney disease, but ‘levels off’ with about 50% affected (see Figures 10 and 15).

Future research

This research has identified a number of issues into which further research is needed. These include areas in which we have been unable to identify research that addresses our research questions, new issues arising from our findings and issues relating to the translation of our findings into clinical practice.

The simulation models used in this work have been developed to enable further data to be added as they become available. The type 2 model, drawing on data from UKPDS, is already in the process of revision by the UKPDS group to take account of the data from long-term follow-up of the cohort. When the UKPDS model becomes available, it will allow our findings to be reviewed and, if necessary, revised.

The type 1 diabetes simulation model is constructed from a number of sources, and as further data become available from larger type 1 studies it can also be expanded, with new equations incorporated into a revised model. It therefore offers a new tool for a wide range of cost–benefit studies which have previously been carried out with models that draw heavily on data from studies of patients with type 2 diabetes.

The uncertainty around the estimates for cost-effectiveness are largely driven by the rate at which new microalbuminuria occurs. There are a number of potential ways in which these findings might be taken forward. Our estimates of cost-effectiveness may be improved by establishing cohorts of individuals at increased risk of developing microalbuminuria and diabetic kidney disease. For example, with increasing capacity to personalise protocols for individual screening, those identified at increased or low risk for progression could be screened more, or less, frequently. Another approach might be to establish whether proactive treatment of those identified at risk might lead to longer-term benefits. However, recent evidence suggests that rates of progression of diabetic kidney disease are falling. 200 In establishing cohorts, estimates of sample size may need to be adjusted to take account of these trends.

An alternative approach to screening for microalbuminuria may be to look for other markers that can be used to identify risk of diabetic kidney disease or to monitor the impact of treatment intended to reduce risk. Among the candidates for new markers is cystatin C, following studies that have shown accurate identification of individuals with diabetes who are undergoing decline in renal function. 201 Cystatin C has been suggested as a more proximal marker of renal damage than microalbuminuria. 202 Initial studies have suggested that cystatin C may identify deterioration in renal function more accurately than eGFR in patients with type 1 diabetes,203 but cost-effectiveness studies are needed to establish whether the improved detection is of sufficient clinical importance to change treatment and, in particular, whether it adds to management, based on occurrence of microalbuminuria. Other strategies that might facilitate an assessment of risk include the measurement of uric acid, tumour necrosis factor receptors, certain advanced glycation end products and chemokines. However, further research to establish the utility of these measures and potentially further trials would be required to change clinical practice; the horizon for such developments is some way into the future.

Attempts have been made to characterise groups at increased risk of developing microalbuminuria. For example, data from a large cohort study have been used to develop a prediction rule for progression to microalbuminuria in type 1 diabetes. 33 However, it is not clear from our data whether more frequent measurement of albuminuria would be helpful. Further research is required to establish whether some of those who have false-positive screening test results are unlikely to develop true microalbuminuria, or whether the increased variability that leads to false-positive test results is a precursor of established microalbuminuria and, therefore, treatment may be justified. This strategy is currently being tested in a randomised trial. 18

Another strategy that might be explored for managing cardiovascular risk in association with screening for microalbuminuria is to investigate whether the occurrence of microalbuminuria, or increasing risk of it developing, might justify other treatments in addition to ACEi/A2RB. Chapter 1 lists a number of possibilities. One further possibility is whether aspirin might have an effect. Aspirin was previously recommended for individuals with diabetes, as the condition was viewed as a cardiovascular risk equivalent. More recent research has thrown doubt on this assumption, and a large-scale trial is currently under way to identify benefit from aspirin in diabetes. 204 Subgroup analysis of this trial based on the occurrence of microalbuminuria, where data are available, might provide information to inform assumptions around the benefits of screening included in our models. It is possible, for example, that, for many individuals with diabetes, aspirin treatment may not be beneficial, but, among those at high risk and identified with microalbuminuria, there might be a benefit from adding insulin to treatment regimens.

Qualitative work with subsequent surveys, and, if appropriate, trials, may also be helpful in establishing whether patient knowledge of renal albumin status provides an additional motivational factor sufficient to increase adherence to medication, without causing an adverse impact on well-being.

Aims and objectives

Methods

Research aims and objectives

Our aim is to develop a clinical and economic model of renal function decline, and then to estimate the effectiveness and cost-effectiveness of different models for chronic kidney disease (CKD) monitoring in type 1 and 2 diabetes.

Research methods

Research methods

MEDLINE search terms

1. exp antihypertensive agents/

2. (antihypertensive$ adj (agent$ or drug)).tw.

3. chlorothiazide.tw.

4. chlorthalidone.tw.

5. hydralazine.tw.

6. hydrochlorothiazide.tw.

7. indapamide.tw.

8. minoxidil.tw.

9. exp angiotensin converting enzyme inhibitors/

10. captopril.tw.

11. enalapril.tw.

12. cilazapril.tw.

13. enalaprilat.tw.

14. fosinopril.tw.

15. lisinopril.tw.

16. perindopril.tw.

17. ramipril.tw.

18. saralasin.tw.

19. teprotide.tw.

20. exp losartan/

21. losartan.tw.

22. imidazole$.tw.

23. irbesartan.tw.

24. candesartan.tw.

25. eprosartan.tw.

26. valsartan.tw.

27. olmesartan.tw.

28. telmisartan.tw.

29. (ace adj2 inhibitor$).tw.

30. (angiotensin adj2 receptor antagonist$).tw.

31. exp calcium channel blockers/

32. amlodipine.tw.

33. diltiazem.tw.

34. felodipine.tw.

35. nicardipine.tw.

36. nifedipine.tw.

37. nimodipine.tw.

38. nisoldipine.tw.

39. nitrendipine.tw.

40. verapamil.tw.

41. exp adrenergic beta-antagonists/

42. alprenolol.tw.

43. atenolol.tw.

44. metoprolol.tw.

45. nadolol.tw.

46. oxprenolol.tw.

47. pindolol.tw.

48. propranolol.tw.

49. exp adrenergic alpha-antagonists/

50. labetalol.tw.

51. prazosin.tw.

52. beta block$.tw.

53. exp diuretics/

54. spironolactone.tw.

55. triamterene.tw.

56. bumetanide.tw.

57. chlorthalidone.tw.

58. furosemide.tw.

59. indapamide.tw.

60. chlorothiazide.tw.

61. hydrochlorothiazide.tw.

62. or/1-61

63. exp diabetes mellitus/

64. diabetic nephropathies/

65. diabetic nephropath$.tw.

66. diabetic glomerulo$.tw.

67. ((diabetic or diabetes) and (kidney disease$ or renal disease$)).tw.

68. or/63-67

69. 62 and 68

70. randomized controlled trial.pt.

71. controlled clinical trial.pt.

72. randomized.ab.

73. placebo.ab.

74. clinical trials as topic.sh.

75. randomly.ab.

76. trial.ti.

77. 74 or 75 or 72 or 71 or 70 or 73 or 76

78. exp animals/ not humans.sh.

79. 77 not 78

80. 69 and 79

81. limit 80 to yr="2005 -Current"

A meta-analysis of health state valuations for people with diabetes: explaining the variation across methods and implications for economic evaluation

Lung TW, Hayes AJ, Hayen A, Farmer A, Clarke PM. Qual Life Res 2011;20:1669–78.

Purpose

To review published studies on the effect of diabetes and its complications on utility scores to establish whether there is systematic variation across studies and to examine the implications for the estimation of QALYs.

Methods

A systematic review was performed using studies reporting QALY measures elicited from people with diabetes including those with a history of complications. Meta-analysis was used to obtain the average utility, and metaregression was employed to examine the impact of study characteristics and elicitation methods on these values. The effect of different utility scores on QALYs was examined using diabetes simulation models.

Results

In the meta-analysis based on 45 studies reporting 66 values, the average utility score was 0.76 (95% CI 0.75–0.77). A metaregression showed significant variation due to age, method of elicitation and the proportion of males. The average utility score for individual complications ranged from 0.48 (95% CI 0.25, 0.71) for chronic renal disease to 0.75 (95% CI 0.73, 0.78) for myocardial infarction, and these differences produced meaningful changes in simulated QALYs. There was significant heterogeneity between studies.