The search strategy was designed by a diabetes specialist with experience in diabetes epidemiology (AGT) with input from all other investigators in accordance with the Meta-analysis of Observational Studies in Epidemiology (MOOSE) guideline (13). MEDLINE databases from their inception until 31/MAY/2021 were searched. Controlled vocabulary supplemented with keywords was used to search for studies fulfilling the criteria of reporting on both sensory diabetic neuropathy and mortality. We imposed no limitation on regional origin or language of any findings. We used the following search phrase: ((senso* OR DSPN OR neuropat* OR “diabetic foot” OR (foot AND ulcer*) OR Charcot) AND (diabete* OR diabeti*) AND (mortality OR death)). We aimed to collect case-control and cohort studies of diabetic people where baseline DSPN status was reported and follow-up mortality by DSPN status could be retrieved. All studies had to be of original data. We did not impose restriction on the length of follow-up.

Several approaches (including questionnaires, symptoms, physical examination, and equipment-based techniques) were accepted for the diagnosis of DSPN ( Table 1 ) as long as the same validated approach was used consistently within a report. Quality assessment of the papers was completed by the Newcastle-Ottawa Scale (NOS). We translated the NOS scores into the following strata using the Agency for Healthcare Research and Quality (AHQR) standards (1): good quality: 3 or 4 stars in selection domain AND 1 or 2 stars in comparability domain AND 2 or 3 stars in outcome/exposure domain (2); fair quality: 2 stars in selection domain AND 1 or 2 stars in comparability domain AND 2 or 3 stars in outcome/exposure domain (3); poor quality: 0 or 1 star in selection domain OR 0 stars in comparability domain OR 0 or 1 stars in outcome/exposure domain (41). The corresponding authors of potentially eligible primary studies without the required information within their publication were not invited to contribute with their raw data.

Reviewers worked independently and checked all abstracts and selected full-text manuscripts for eligibility. Preselected manuscripts were evaluated by 2 independent reviewers for eligibility. The total number of included and excluded articles was documented by the reviewers, including reasons for exclusion or non-eligibility. Any disagreements were discussed with the study designer (AGT). Whenever reviewers disagreed and no consensus was found, the article was included into the full-text phase. Disagreements at full-text screening were resolved by consensus of all authors.

For all studies, we extracted information on study design, number and characteristics of participants (i.e. age, sex, type of diabetes), prevalence of DSPN, method of DSPN assessment, duration of follow-up, all-cause mortality, matching, and confounding factors. Additionally, in the case of multiple publications, we included the most up-to-date or comprehensive information.

From the given publications, we sought data on the association between DSPN and all-cause mortality with their respective 95% confidence intervals (CI) based on unadjusted (minimally adjusted) and fully adjusted models. For the unadjusted models, we collected raw numbers, odds ratios (OR), hazard ratios (HR), or incidence rate ratios (IRR) that were either unadjusted or adjusted for only age and sex. If any other variable (i.e. diabetes type, BMI, and diabetes duration, co-morbidities) was taken into account, the estimate was considered to be fully adjusted. If more than one model was reported, the one with the most co-variables was selected. At this stage, we excluded unadjusted estimates from those studies, where the population was referred specifically for the examination of sensory or autonomic neuropathy, given the high risk of collider bias in this setting. If a given paper reported adjusted estimates, those were used in the analysis.

The outcome measure of this meta-analysis was pooled all-cause mortality among individuals with DSPN compared to those without DSPN. We pooled estimates with their 95% confidence intervals (CIs) using random effect meta-analysis, as this methodology works well even when heterogeneity between studies is substantial (42). We also pre-planned to run analyses stratified by the type of diabetes (type 1 diabetes, type 2 diabetes, or undefined/mixture of type 1 and type 2 diabetes). We also tested formally for a difference between the estimates for type 1 and type 2 diabetes using meta-regression. For the main analysis, we included all cohorts that provided either unadjusted or adjusted estimates (using the unadjusted estimates if available).

To assess statistical heterogeneity, visual inspection of the forest plots was used, followed by formal testing using the I²-statistic. This provides an estimate of the percentage of variability across studies due to heterogeneity rather than chance: I² < 40% may represent low heterogeneity; 30-60% may represent moderate heterogeneity; 50-90% represent substantial heterogeneity; and 75-100% represent considerable heterogeneity (43). Publication bias was evaluated by visual inspection of the funnel plots and formally by Egger’s tests. Additionally, we tested for influence of individual studies using a meta-analysis influence test that eliminated included studies one by one.

Given the expected heterogeneity of the eligible studies, pre-planned sensitivity analyses (stratified by type of diabetes) were also carried out by restricting the analysis to (1) studies that adjusted for potential predictors of mortality (other than age and sex) (2), population-based investigations (3), studies that used semi-quantitative (physical examination including monofilament or tuning fork or pinprick tests) or quantitative (vibration or current perception threshold or nerve conduction test) methods for DSPN assessment, and to studies that were deemed to be of good quality according to the Newcastle-Ottawa Scale (41). A further sensitivity analysis was done using meta-regression with adjustment for diabetes duration to investigate whether a risk difference between type 1 and type 2 diabetes would be explained by different length of diabetes duration.

All statistical tests were two sided and used a significance level of p<0.05. We used STATA 15.1 (StataCorp, CollegeStation, TX) for all statistical analyses.

The flow chart of the search strategy is presented in Figure 1 . An initial search produced 4904 articles ( Figure 1 ). Based on titles and abstracts, we excluded 4844 articles that reported no original data that didn’t investigate DSPN or didn’t report mortality data, or had no control population (without DSPN), leaving 60 papers for full text retrieval. All these papers were written in English. We excluded a further 31 papers due to different reasons leaving 29 papers that could be included in the meta-analysis. Of the 29, two studies (17, 37) reported estimates separately for type 1 and type 2 diabetes and one (14) reported estimates separately for two population-based cohorts (ADDITION and DD2) that allowed us to use them as altogether six separate cohorts. In contrast, we excluded 2 studies from the unadjusted analysis (37, 44), due to the potential of collider bias. Our final sample for the main analysis thus included 31 cohorts and for the fully adjusted analysis 17 cohorts (12, 14–40).

Detailed characteristics of the included cohorts are summarized in Table 1 . Out of the included 31 cohorts, 10 reported on type 1 diabetic patients, 16 on type 2 diabetic patients. In the remaining 6 papers, type of diabetes was either not reported, or estimates could not be separated for patients with type 1 and type 2 diabetes.

Nine of the cohorts included population-based samples, while most were performed in tertiary care (14 in single, 2 in multiple centers). Six cohorts used data from randomized trials or their follow-up cohorts ( Table 1 ).

Diagnosis of DSPN varied widely between the different cohorts. Five cohorts used medical records as the source of data, leading to potential bias related to the non-standardized ways of examination. 3 cohorts used non-standardized symptom-checks only for diagnosis, 6 cohorts standardized questionnaires, 8 physical examination, and 7 studies used different equipments that provide quantitative measures of neuronal function. Only 5 studies used a combination of the above methods to provide the diagnosis of DSPN. When quantitative or semi-quantitative methods were considered as more reliable tools for the diagnosis of DSPN, 18 studies fulfilled this requirement ( Table 1 ).

Altogether over 150 thousand diabetes patients (54% male) were included in the main analysis of whom 12.3% died during the median 7.5 (range 1.9-15.0) years of follow-up. The median proportion of participants with DSPN was 27.4% (range 8.8%-73.6%).The median age of participants was 60 years (range 20.8-68.0 years), diabetes duration ranged 0 to 25 years (median 11.1 years) in the different cohorts ( Table 1 ).

While all studies provided some estimates of mortality, unadjusted (including those only adjusted for age and sex) estimates were available for 29 cohorts, while multiple adjusted estimates were reported in 17 cohorts ( Table 1 ).

According to the NOS classification, 18 studies had good, 1 study had fair, and 9 poor quality ( Supplementary Table S1 ).

In the pooled primary analysis, the presence of DSPN at baseline was significantly associated with an almost doubled risk of mortality (pooled HR 1.96, 95%CI 1.68-2.27). While the point estimates in the individual studies were over one for all but one studies, the heterogeneity between studies was large (I² = 91.7%, p<0.001) ( Figure 2 ).

According to our primary hypothesis we stratified the analysis between DSPN and mortality by the type of diabetes. Among patients with type 1 diabetes, the mortality was 3.64 (95%CI 2.55-5.20) times higher in the presence of DSPN compared to those without DSPN. While the I² value (75.4%, p<0.0001) was somewhat lower in these studies compared to the overall value, heterogeneity remained large ( Figure 2 ).

The pooled association between DSPN and mortality was weaker in patients with type 2 diabetes (pooled HR 1.63, 95%CI 1.37-1.93), although DSPN still was a strong predictor of all-cause mortality. There was large heterogeneity in the estimates between studies reflected by the high I² value (91.8%, p<0.0001), partly related to the study by Foryoung et al. that reported a decreased risk of mortality in patients with DSPN ( Figure 2 ).

In cohorts where it was impossible to assess the risks by diabetes type, we found a similar association between DPSN and mortality as in type 2 diabetes (pooled HR 1.49, 95% CI 1.35-1.64). There was a low heterogeneity between the individual risk estimates (I² 0%, p=0.48), suggesting that these estimates may have come from a common background population (probably mostly type 2 diabetes patients) ( Figure 2 ).

When we formally tested for a difference in the estimates between type 1 and type 2 diabetes by meta-regression, we found that the risk associated with DSPN was more than 2 times higher in type 1 compared to type 2 diabetes (HR 2.22, 95%CI 1.43-3.45, p=0.001).

We also investigated the potential for publication bias in our estimates. Given the large difference in risks between diabetes types, we drew funnel plots separately for each type of diabetes. These funnel plots confirmed the large difference between type 1 and type 2 (and also undefined diabetes) groups but were incompatible with the presence of strong publication bias. The Egger tests were all non-significant (all p>0.8) further confirming the visual interpretation ( Supplementary Figure 2 ).

We also tested for influential studies by eliminating individual studies from the meta-analysis run for each type of diabetes. All these analyses showed overlapping confidence intervals with the overall (diabetes type specific) estimates arguing against the role of some influential studies in the findings. (Figures are available on request.)

Our pre-planned sensitivity analyses restricted included cohorts by level of covariate adjustment, type of study setting, diagnostic test used to define DSPN, and overall quality of studies.

First, we run separate analyses including 29 studies that reported unadjusted, and 17 studies that reported adjusted estimates. The overall risk associated with DSPN was smaller according to the adjusted estimates (pooled HR 2.0 vs. 1.6) suggesting that the included covariates explained some of the increased mortality in DSPN. In contrast, the difference between type 1 and type 2 diabetes remained statistically significant showing stronger associations between DSPN and mortality in type 1 compared to type 2 diabetes ( Table 2 ).

When we restricted the analysis to population-based studies, no cohort performed on type 1 diabetes patients remained. The overall risk estimates however confirmed the association between DSPN and mortality with an almost 2 times higher risk in patients with DSPN ( Table 2 ).

When we restricted the analysis to cohorts that used quantitative or semi-quantitative methods to define DSPN, the overall association between DSPN and mortality remained and the point estimate was similar to those of the other analyses. Furthermore, heterogeneity was somewhat smaller for both types of diabetes (<75%), suggesting that the use of different methods could increase heterogeneity in the mortality estimates. Although the absolute difference in hazard ratios between type 1 and type 2 diabetes was smaller in this analysis, the estimates remained statistically significantly different further supporting our hypothesis that a stronger association exists between DSPN and mortality in type 1 diabetes ( Table 2 ).

When we restricted our analysis to cohorts that deemed to be of good quality according to the NOS scale, we found point estimates similar to those from the main analysis, confirming the association between DSPN and mortality. Although the difference between type 1 and type 2 diabetes pointed to the same direction as in the main analysis, the difference between hazard ratios was statistically not different probably due to the limited number of studies included in this analysis ( Table 2 ).

Finally, in a meta-regression adjusted for diabetes duration, we found a similarly increased mortality in type 1 diabetes compared to type 2 diabetes as in the main analysis, although due to the wide confidence intervals, it lost its statistical significance.

Based on a meta-analysis of 31 cohorts comprising over 150 thousand diabetic patients, we found an almost doubling (HR 1.96, 95%CI 1.68-2.27) of all-cause mortality among diabetes patients with distal symmetric polyneuropathy compared to those without DSPN. The association was somewhat attenuated but still statistically and clinically significant (HR 1.60, 95%CI 1.37-1.87) in those studies that take into account multiple risk factors of vascular disease and all-cause mortality. The observed association was more than two times stronger in type 1 compared to type 2 diabetes (HR 2.22, 95%CI 1.43-3.45), most likely reflecting the much smaller background mortality of younger patients without DSPN. The observed findings were robust in several sensitivity analyses that aimed to remove bias related to different study designs, methods for DSPN definitions, and overall study quality. Furthermore, we found no sign of publication bias, or any individual study with a major influence on the estimates.

The mechanism responsible for the association between DSPN and mortality is not well investigated. In general, DSPN is thought to be a determinant of reduced quality of life through disturbed sleep, physical functioning, recreation, and diminished physical and emotional well-being (54). However, the neuroendocrine, proinflammatory, and neurodegenerative underpinnings of DSPN could also lead to cardiovascular disease through increased oxidative stress and levels of advanced glycation end products (55, 56). DSPN is also associated with balance impairment that could lead to falls and injuries (56). DSPN is a leading factor of diabetic foot ulcers and amputations, both associated with increased mortality through infection and chronic inflammation (57).

Furthermore, diabetic pain per se (through similar mechanisms as DSPN itself) could be associated with an even further increased risk of mortality, as suggested by the observation of Lapin et al. on increased mortality only in those patients with painful DSPN but not in those with DSPN without pain compared to DSPN free controls (27). This hypothesis seems to be further strengthened by a meta-analysis that suggests an increased risk of mortality in people with widespread body pain irrespective of its origin even after adjustment for some mortality risk factors (58).

Alternatively, it is also possible that DSPN is a marker of other diseases that increase mortality. Indeed, microvascular diabetes complications (especially DSPN and CAN) show remarkable clustering (59). Thus, it is possible that the association between DSPN and mortality could be mediated through CAN.

Our meta-analysis was based on a predefined protocol standardizing the inclusion of studies and also the testing of our hypotheses. Overall, our current analysis is based on a sample with over 150 thousand participants with sufficiently high rates of mortality and DSPN to provide stable estimates. The main outcome (all-cause mortality) is a hard outcome that is unlikely to be imprecisely reported. The observed effect size is clinically relevant and statistically significant. The similar results in our main and sensitivity analyses further confirm the observed strong associations. Furthermore, our tests for influential studies and publication bias argues against major effects of individual studies or selective publication.

As the quality of any meta-analysis is mostly determined by the quality of the included studies, the main limitations of our report are mostly related to the quality of the included studies. First, a large proportion of the included studies include patients from tertiary care centers, from high-income countries with Caucasian origin. All of these factors limit the external validity of our findings. Furthermore, although some of the studies reported on multiple adjusted associations between DSPN and all-cause mortality, the role of unmeasured confounding cannot be excluded. The included studies had limited data on important mortality risk factors, such as different medications, laboratory parameters, proportion of participants with risk factors at target, or other comorbidities with increased mortality. Furthermore, the differences in the definition of DSPN could have biased our results. Given that we included only cohort studies, bias related to lost to follow-up cannot be excluded. While all-cause mortality is an easily obtainable outcome, it would be of interest to see the associations of DSPN with different causes of death.