Introduction

Diabetic peripheral neuropathy (DPN) is a common complication of both type 1 and type 2 diabetes. It is a major cause of foot ulcers and subsequent amputations, and is also associated with higher rates of all-cause and cardiovascular mortality compared with individuals without DPN [1,2,3]. Poor glycaemic control is the strongest risk factor for the development of DPN; however, age, duration of diabetes, smoking and the presence of other comorbidities (dyslipidaemia, hypertension, diabetic retinopathy, nephropathy) are also associated with DPN [4, 5]. Previous studies have shown that DPN is present in 30–50% of diabetic individuals [6, 7]. Therefore, screening and timely diagnosis of DPN is important for improving foot care and highlighting the risk of adverse cardiovascular outcomes.

While electromyography, electroneurography and sural nerve biopsy are the gold standard for diagnosing DPN, the Michigan Neuropathy Screening Instrument (MNSI), introduced in 1994, is considered to be an alternative diagnostic tool. The MNSI includes two separate assessments: a 15-item self-administered questionnaire about symptoms, and a lower extremity examination that includes inspection and assessment of vibratory sensation and ankle reflexes [8]. The MNSI was developed to diagnose DPN in clinical practice and has also been used in large clinical trials [9,10,11]. However, its value in risk stratification has not been ascertained. Therefore, we investigated whether the MNSI questionnaire offers additional information about risk of death and major cardiovascular events in high-risk individuals with type 2 diabetes and chronic kidney disease (CKD) and/or cardiovascular disease (CVD).

Methods

Participants

The Aliskiren Trial in Type 2 Diabetes Using Cardio-Renal Endpoints (ALTITUDE, NCT00549757) was a double-blind, randomised, controlled trial conducted among 8561 participants with type 2 diabetes and CKD and/or CVD who were randomly assigned to receive aliskiren (300 mg/day) or placebo, added to an ACE inhibitor or angiotensin II receptor blocker. Participants were followed for a median of 2.6 years (interquartile range 2.0–3.2). The trial design and results have been published previously [12, 13]. The trial was approved by the ethics committee or institutional review board at each participating center and was carried out in accordance with the Declaration of Helsinki as revised in 2008. All patients provided written informed consent.

Of the 8561 randomised participants, 8463 completed the self-administered MNSI questionnaire (the questionnaire is included in the electronic supplementary material [ESM Methods]) at baseline and were included in this post hoc analysis (the remaining 98 participants did not complete the questionnaire). Based on the date of randomisation, we divided this cohort into independent training (n = 3252, randomised 2007–2008) and validation sets (n = 5211, randomised 2009–2011).

Study design

In the training set, we identified specific questions that were independently associated with the cardiovascular composite outcome (cardiovascular death, resuscitated cardiac arrest, non-fatal myocardial infarction/stroke, heart failure hospitalisation). These questions were then evaluated in the validation set in models adjusted for the randomised study treatment (model 1) and baseline covariates (model 2) to confirm their potential independent association with clinical outcomes.

Statistical analysis

Baseline characteristics were stratified by study group (training set/validation set). Continuous data are presented as the mean ± SD, except for the triacylglycerol and urinary albumin-to-creatinine ratio data, which are presented as median (interquartile range). Categorical variables are expressed as proportions and were compared by the χ 2 test, while continuous variables were compared using t tests or Wilcoxon rank-sum tests as appropriate. Using the training dataset, forward stepwise-selection techniques were used with threshold p values at 0.05 to identify specific questions that were independently associated with the cardiovascular composite outcome without adjustment for any other variables. We further tested the null hypothesis that all selected questions were equally associated (i.e. equal HR) with the outcome of interest vs the alternative that one or more of the selected questions were differentially related to the outcome (i.e. one or more HRs different from the others). These selected questions were then tested as predictors in adjusted Cox proportional hazards models using the validation dataset. We used these questions in two models. First, we estimated the risk associated with an affirmative answer to any of the three questions compared with those with no affirmative answers. Next, we estimated the risk associated with each specific number of affirmative answers compared with a reference of zero affirmative answers. Proportional hazards regression models were used to assess the association between the questions and cardiovascular outcomes. Model 1 was adjusted for the randomised study treatment. Model 2 was adjusted for baseline covariates: age, sex, race, smoking status, systolic blood pressure, eGFR, urinary albumin-to-creatinine ratio, history of heart failure, myocardial infarction, stroke, atrial fibrillation, diabetic nephropathy, diabetic retinopathy, amputation, claudication, unstable angina, coronary revascularisation, duration of diabetes, HbA1c and randomised treatment [14]. Harrell’s C statistics (compared using a transformation of the equivalent Somers’ D parameters [15]), continuous net reclassification improvement (NRI) and integrated discrimination improvement (IDI) were obtained by comparing the described multivariable Cox models with and without the inclusion of the variable identifying the number of ‘yes’ responses from among the chosen MNSI questions and was assessed at 3 years post-randomisation. Analyses were repeated for subgroups of participants based on CVD history as well as CKD status at randomisation. Assessment of the proportional hazards assumption was performed using Schoenfeld residuals. Two-sided p values <0.05 were considered to be statistically significant. Analyses were performed using Stata version 13.1 (Stata Corporation, College Station, TX, USA).

Results

The 8463 participants were divided into a training (n = 3252) and a validation group (n = 5211). Baseline characteristics in these two groups are presented in Table 1. Participants included in the training and validation groups were qualitatively similar. All pairwise correlations between MNSI questions were less than +0.40. In the training dataset, we identified three questions which were independent predictors of the cardiovascular composite outcome: question 1, Are your legs and/or feet numb? (HR 1.25 [95% CI 1.04, 1.50], p = 0.020); question 8, Have you ever had an open sore on your foot? (HR 1.44 [95% CI 1.18, 1.76], p < 0.001); and question 12, Do your legs hurt when you walk? (HR 1.62 [95% CI 1.35, 1.95], p < 0.001) (Table 2). The associations of the three questions with specific cardiovascular outcomes in the training dataset are shown in ESM Table 1. After adjusting for other baseline covariates, we found no evidence that the HRs associated with any of the three questions were significantly different from the others with respect to any of the outcomes.

Table 1 Baseline characteristics by study group
Full size table
Table 2 Association between the three questions and cardiovascular composite outcome, models 1 and 2 in participants randomised in 2007–2008 (training group n = 3252)
Full size table

In the validation set, 3079 participants (59.1%) answered ‘yes’ to at least one of these three questions (29.7% answered positively to one, 18.8% to two and 10.6% to three questions).

We then analysed the associations between different combinations of individual questions and cardiovascular outcomes in the validation dataset (Table 3, Fig. 1). In model 1, participants who answered ‘yes’ to at least one of the three questions demonstrated a higher risk of the cardiovascular composite outcome (HR 1.77 [95% CI 1.48, 2.11], p < 0.001), all-cause death (HR 1.48 [95% CI 1.19, 1.84], p < 0.001), cardiovascular death (HR 1.51 [95% CI 1.14, 2.01], p = 0.004), heart failure hospitalisation (HR 2.00 [95% CI 1.50, 2.68], p < 0.001), myocardial infarction (HR 2.20 [95% CI 1.51, 3.21], p < 0.001), stroke (HR 1.94 [95% CI 1.36, 2.78], p < 0.001) and three-point major adverse cardiovascular events (MACE) (cardiovascular death, non-fatal myocardial infarction, non-fatal stroke) (HR 1.75 [95% CI 1.42, 2.15], p < 0.001) relative to those who answered ‘no’ to all of the questions. After multivariable adjustment, the excess risk associated with a positive answer to a question was reduced by 25–50%, but remained clinically important and retained statistical significance (with the exceptions of all-cause death and cardiovascular death). These associations became even stronger if participants answered positively to multiple questions in relation to all cardiovascular outcomes, except for stroke (ESM Table 2). The addition of the total number of affirmative responses to existing models significantly improved Harrell’s C statistic for the cardiovascular composite outcome (0.70 vs 0.71, p = 0.010), continuous NRI (+22% [+10%, +31%], p = 0.027) and IDI (+0.9% [+0.4%, +2.1%], p = 0.007). For all other outcomes, C statistic was improved by +0.00 to +0.02, NRI by +17% to +25% and IDI by +0.3% to +0.8% (Table 4). All results were generally consistent across subgroups of participants defined by the presence or absence of CVD and CKD (ESM Tables 3–10).

Table 3 Association of the questions and outcomes, models 1 and 2 in participants who were randomised in 2009–2011 (validation group n = 5211)
Full size table
Fig. 1
figure 1

The association of answer ‘yes’ to (a) any of the three questions and (b) all three questions with adverse outcomes in participants randomised in 2009–2011 (validation group n = 5211). A logarithmic scale has been used on the x-axis. CV, cardiovascular

Full size image
Table 4 Harrell’s C statistic, continuous NRI and IDI: comparison of models with and without three MNSI questions
Full size table

Discussion

This study identified three out of the 15 simple questions from the MNSI that provide additional prognostic information about the risk of all-cause death, cardiovascular death and a composite cardiovascular outcome, as well as its components, in individuals with type 2 diabetes and CKD and/or CVD. While the MNSI questionnaire has previously been used to screen for DPN [9,10,11], this is the first time these questions have been used as predictors of risk of cardiovascular events. A positive answer to any one of the three identified questions was associated with a higher risk of adverse cardiovascular events, and the relationship was even stronger if the answers to multiple questions were positive. Furthermore, when participant characteristics and other cardiovascular risk factors were considered, a positive answer to each of these three questions provided additional prognostic information concerning the risk of the outcomes described.

The DCCT/EDIC study investigators examined the performance of each item in the MNSI questionnaire and examination in confirming the diagnosis of DPN. These investigators concluded that a reduced index of four questions performed nearly as well as the more extensive instrument [11]. Two of these four questions (question 1 ‘Are your legs and/or feet numb?’ and question 8 ‘Have you ever had an open sore on your foot?’) were also identified by us as important predictors of cardiovascular outcomes.

As cardiovascular risk factors and microvascular disease are associated with both DPN and CVD in people with type 2 diabetes, it is perhaps unsurprising that instruments designed to detect DPN might also predict future CVD. That the risks associated with simple questions are strong and only marginally attenuated by multivariable adjustment that accounts for variables associated with the risk of experiencing the primary outcome of the study is more striking. The aetiological explanations for this are unclear. Alternatively, or additionally, these questions may capture unmeasured pathways (e.g. chronic inflammation in association with a history of an open sore) or more global aspects of vascular damage in association with type 2 diabetes. A similar hypothesis has been invoked to account for the repeated observation that abnormalities of retinal vascular architecture, which are thought to mimic patterns in the cerebrovascular territory, predict outcomes independent of other vascular risk factors [16, 17].

It is well known that cardiovascular events are a leading cause of increased morbidity and mortality rates in people with type 2 diabetes. Therefore, in clinical settings, there is a growing need for tools which could help to accurately assess the risk of adverse events in these individuals. While study of residual confounding and alternative pathways are of interest in understanding mechanisms to identify new therapeutic targets, current clinical practice should be more concerned with identification of high-risk individuals to whom existing therapies can be targeted. In clinical practice, complication assessment will also suffer from similar levels of imprecision as in ALTITUDE and other studies, and performance of additional tests, such as markers of inflammation, are time consuming and expensive. Inclusion of just three simple, yet strongly predictive, questions may therefore be of substantial importance in clinical practice. Prediction models are widely used in medicine as potential aids in clinical understanding and therapeutic decision-making, as well as better assessment of prognosis. However, before a new risk prediction model could be adapted in clinical practice and widely used it needs to be externally validated to assess its generalisability [18,19,20]. In this analysis, we demonstrated internal validation by showing that three questions identified from the derivation set provided statistically significant prognostic information in an independent validation set.

There are several limitations of this study that need to be noted. First, all participants were included in ALTITUDE based on same inclusion criteria and, although divided into statistically independent derivation and validation groups, the two groups were similar overall. Therefore, the lack of external validation, as well as the fact that these findings may not be generalisable to other populations of individuals with type 2 diabetes, is considered an important limitation. In addition, we are limited by the fact that the cohort involved in this clinical trial included only high-risk participants who were therefore more susceptible to adverse cardiovascular outcomes.

In conclusion, we believe that these three questions represent a simple, non-invasive and inexpensive tool which could potentially provide additional prediction information in clinical practice. If externally validated in people with type 2 diabetes who have a lower cardiovascular risk, these questions may be integrated into the clinical history to augment the prediction of cardiovascular events in high-risk individuals.