We searched MEDLINE (www.ovid.com/product-details.901.html), EMBASE (www.ovid.com/product-details.903.html) and CINAHL (www.ebscohost.com/nursing/products/cinahl-databases/cinahl-complete) databases (all accessed 6 February 2020) from January 1980 to December 2019, using the following combined free text and medical subject heading search strategy: (Diabetes OR Diabetes Mellitus OR ‘Diabetes Mellitus, Type 2’) AND (Mortality OR ‘Cause of death’ OR ‘Child mortality’ OR Death OR Trends). We limited the search to articles published in English (Electronic supplementary material [ESM] Table 1). The reference lists of identified articles were also screened and we undertook a Google search to identify any relevant government reports.

To be included in this review, studies needed to report all-cause mortality rates in populations with diabetes in two or more separate time periods such that a trend could be calculated. Studies only reporting death attributable specifically to diabetes were excluded. A detailed explanation of inclusion and exclusion criteria is provided in the ESM Methods.

Studies that were derived from nationally representative data were prioritised over studies that were based on regional or selected groups. Where there were multiple publications using the same data source, we chose the publication that covered the longest time. Each title and abstract was screened by at least two authors (D. J. Magliano, D. N. Koye, E. L. M. Barr, J. L. Harding, J. W. Sacre, J. Wang, L. Chen, M. Tabesh, J. E. Shaw and R. M. Islam) and discrepancies were resolved by discussion.

We categorised the definition of diabetes that was reported in the original publications into one of five groups: self-report of physician-diagnosed diabetes, linkage to medication or reimbursement registries, measurement of blood glucose or HbA_1c, clinical diagnosis by physicians or from International Classification of Disease codes, or administrative algorithm derived.

Death was determined by linkage to the national death registries within the respective country.

Extracted data included study and population characteristics, crude or standardised mortality rate (per 1000 person-years) or cumulative mortality (% per year) during each time period or annual mortality rate (per 1000 people) among people with diabetes and those without diabetes (if available) (including number of deaths and the corresponding number of people with and without diabetes, respectively), and the reported pattern of the trend (increasing, decreasing or unchanged mortality [i.e. no statistically significant (p < 0.05) evidence of changing mortality rates]) over the study’s observation period (Table 1 and ESM Table 2). Each study’s observation period was the time interval between the first and last calendar years that mortality rates were reported. If mortality rates were reported for a multi-year time period, instead of by individual calendar year, the mortality rate was attributed to the calendar year that corresponded to the midpoint of the time period.

From each study, we extracted data from each subpopulation reported, such that a study that reported mortality trends in men and women separately contributed two populations (ESM Table 2). To avoid double counting, when the data were reported in the total population and by sex and ethnic groups, we included data only once and prioritised ethnic-specific data over sex-specific and total population data. We classified studies as two major ethnic categories, i.e. Europid and non-Europid populations, based on the predominant ethnicity of the country where they were conducted. Studies from the indigenous populations were included in the non-Europid category, irrespective of the country where they were conducted.

Age-specific mortality rates for each time period and the reported pattern of mortality trends by age group were also extracted where available (ESM Table 2) and then categorised into four age groups (<40, 40–54, 55–69 and ≥70 years). In studies where numbers of deaths and numbers with diabetes were reported by smaller age groups, we re-calculated mortality rates across our specified larger age groups. Where there were multiple age groups within any of our broader age groups, but insufficient information to combine the data into a new category, only data from the one age group that was most representative of the broader age group were used.

Data were extracted independently by pairs of authors (L. Chen and R. M. Islam or J. Wang) and were checked for any errors (D. J. Magliano). Data reported only in graphs were extracted using DigitizeIt software [11]. The study quality of all included studies was assessed by two authors (L. Chen and T. R. Hird) using a modified Newcastle–Ottawa scale for assessing the risk of bias of cohort studies [12] (details in the ESM Methods). Any disagreements were discussed between two authors (L. Chen and D. J. Magliano), and another author (J. E. Shaw) gave a final judgement if no consensus could be reached.

The primary analysis was to investigate the proportions of populations reporting increasing, decreasing and unchanged mortality rates in people with diabetes within comparable calendar times. We allocated the reported trend of each population to the midpoint of each study’s observation period, and then assigned this trend to one of four pre-specified calendar periods: 1970–1979, 1980–1989, 1990–1999 and from 2000 onwards.

Where a test of significance of trends was not reported or when a study’s observation period was longer than 10 years (ESM Table 3), we used joinpoint regression to identify the years or time periods where linear trends changed significantly either in direction or in magnitude, and calculated an annual per cent change (APC) for each segment identified and the average annual per cent change (AAPC) over the study’s whole observation period with the weights equal to the length of each segment [13, 14] (details in the ESM Methods). For studies reporting mortality rates in only two time periods, we performed log-linear regression to determine the difference in mortality rates between the two time periods and to calculate the APC in mortality rates. We classified the mortality trend as ‘increasing’ if the point estimate of the APC was positive and this value was statistically significant based on the 95% CI of the APC, i.e. the 95% CI excluded the null value of zero. Conversely, a trend was defined as ‘decreasing’ if the point estimate of the APC was negative and its 95% CI excluded the null value of zero. When the 95% CI of the APC included the null value of zero, we classified the trend as without statistical significance, or ‘unchanged’. If two different trends (either direction or magnitude) were detected over two different segments/parts of the study period in one population, we considered these as two separate populations.

We examined the variance in the APC in mortality rates in people with diabetes using random-effects meta-analysis weighted by the inverse of the variance of the APC, and we used the I² statistic to estimate the percentage of variability caused by between-population heterogeneity [15]. We conducted subgroup analyses among populations with data on APC in mortality rates according to the following variables: predominant ethnicity (Europid, non-Europid); sex (men, women and total); year of study midpoint (1970–1989, 1990–1999 and 2000–2016); number of time periods with mortality rates reported (<4 and ≥4 time periods); whether the studies were of high quality (greater than 6 out of a maximum of 11 points) vs low quality (a score of 6 or below) on the Newcastle–Ottawa scale; and whether the data were sourced from an administrative or health insurance or registry database vs from a survey database. Random-effects metaregression analyses were used to test differences across subgroups. Subgroup analyses were also conducted by restricting to those studies based on administrative or health insurance or registry database.

We examined whether people with diabetes had a greater or lesser reduction in mortality rates than those without diabetes over each study’s whole observation period among studies with available data for comparison. If a statistical test of difference in mortality trends between diabetes and non-diabetes was not reported in the study, we calculated the z score using the point estimate and its variance for the difference in the APCs in mortality rates between those with and without diabetes [16].

To test whether use of multiple populations per study was driving our primary results, we conducted a sensitivity analysis in which we examined the patterns of mortality trends among the total population with diabetes in each study. We also investigated the patterns of mortality trends in people with diabetes after excluding studies with quality scores in the lowest quartile. Analyses were conducted using Stata software version 15.1 (Stata Corporation, College Station, TX, USA) and the Joinpoint Regression Program, version 4.7.0.0 (Statistical Methodology and Applications Branch, Surveillance Research Program, National Cancer Institute, Bethesda, MD, USA).

A total of 30,295 abstracts published between 1 January 1980 and 31 December 2019 were identified. Among these, 89 papers described mortality trends among people with diabetes, of which 36 met all inclusion criteria (ESM Fig. 1). Of the 36 articles, two reported crude mortality rates by age groups and sex and age-and sex-adjusted mortality rates, respectively, but were based on the same data source and thus were treated as one study [17, 18]. Another article reported mortality trends using data from two countries (Canada and the UK) [6], and was treated as one study with two data sources. Therefore, 35 studies (36 data sources) from 17 countries/regions reporting mortality trends since 1970 contributed to the analysis; all were from high-income countries (Table 1 and ESM Table 2). Only 17% (6/35) of studies provided data on predominantly non-Europid populations [8, 19–23], including two (6%) in indigenous people [20, 21].

The most common data source was administrative data (47% of the 36 sources), followed by registry data (28%), health surveys linked to a mortality database (17%) and health insurance data (8%); 64% (23/36) of the data sources were based on national data. These data sources employed various definitions of diabetes, i.e. clinical diagnosis (33%), diabetes medication (via linkage to medication registries) (8%), an algorithm (incorporating at least two different criteria) (42%), measurement of blood glucose or HbA_1c (3%) and self-report (11%). The sample sizes of the populations were greater than 10,000 per time period in 75% (27/36) of the data sources and greater than 150,000 per time period in 50% of the sources. Eleven studies (31%) restricted the population to type 2 diabetes while the remainder included all people with diabetes. Annual and standardised mortality rates were reported in 69% (24/35) and 71% (25/35) of the studies, respectively.

We separated the data into populations based on ethnicity and sex, and performed joinpoint or log-linear regression in those populations without reported mortality trends or with study periods longer than 10 years, generating 69 separate populations with a single trend over the whole or part of a study’s observation period (Table 2). Overall, all-cause mortality rates in people with diabetes declined in 43% (3/7), 53% (10/19) and 74% (32/43) of the populations studied in 1970–1989, 1990–1999 and 2000–2016, respectively (Fig. 1).

ESM Fig. 2 shows the APCs in mortality rates by sex and the midpoint of the study period among the 67 separate populations (APC data unavailable in two populations). We found that 81% (54/67) of the populations had a negative change in mortality rates (the point estimate of the APC less than zero), of which the 95% CI of the APC was less than zero in 42/54 (ESM Fig. 3). The proportion of populations with small sample sizes (i.e. <10,000 per time period) was similar between those with unchanged and those with significantly decreasing mortality rates (47% [9/19] vs 27% [12/45], p = 0.11). There was significant heterogeneity in the APCs in mortality rates among the 67 populations (I² = 95.1%, p < 0.001; ESM Fig. 2) and across different subgroups (I² = 37.1~97.4%; ESM Figs 4 and 5).

Figure 1 summarises the findings from Tables 2 and 3, stratifying by predominant ethnicity, pre-specified age groups and sex within comparable calendar periods. The proportion of populations with declining mortality rates was higher in predominantly Europid populations than in non-Europid populations in 1990–1999 (75% [9/12] vs 14% [1/7], p = 0.01), but was not statistically different in 2000–2016 (78% [28/36] vs 57% [4/7], p = 0.25).

Among two US studies reporting mortality trends among different ethnic groups with diabetes, an earlier study showed significant declining mortality rates in predominantly Europids, while mortality rates were unchanged among US black, Hispanic and Asian people in 1994–2001 [19]; another recent US study reported declining mortality rates in Europids and black people, but not in other US ethnic groups from 1988 to 2015 [8]. Another study did not show that trends differed between Canadian indigenous and non-indigenous populations with diabetes in 1995–2007 [21].

Of the 14 studies reporting age-specific mortality trends, six described trends in the period earlier than 2000 [6, 19, 24–27] and no clear patterns of trends were observed during this period (Table 3, Fig. 1 and 2). In 2000–2016, the proportion of populations reporting declining mortality rates was lower in those aged <40 years than in those aged 55–69 years (33% vs 88%, p = 0.02) and in those aged ≥70 years (33% vs 76%, p = 0.02).

Among studies that directly compared mortality trends among different age groups, a Scottish study reported reductions of similar magnitudes in all age groups [28]. In contrast, greater reductions in mortality rates were observed in younger adults in a study from Canada [18], and in 65–74 years olds compared with those aged ≥75 years in an Italian population [29].

A total of 15 studies provided data on mortality trends by sex, mainly in 2000–2016 (Table 2, Fig. 1). Similar proportions of populations with declining mortality rates were observed for men (75%) and women (69%) in 2000–2016. Among six studies specifically comparing men and women, four showed greater mortality declines in men [8, 29–31], while the remaining two reported no significant difference [18, 32].

A total of 44 populations either reported a difference in mortality trends between people with and without diabetes or had data available comparing the difference in the APCs in mortality rates between these two groups in each study. Of these, 17% (1/6), 60% (6/10) and 71% (20/28) showed that the reduction in mortality rates in diabetes was at least as great among those with diabetes as among those without diabetes in 1970–1989, 1990–1999 and 2000–2016, respectively (ESM Table 4).

Figure 3 shows the APCs in mortality rates by diabetes status in the 33 populations that had data available for calculating the APC separately for those with and without diabetes. In 1970–1989, 1990–1999 and 2000–2016, 25% (1/4), 60% (6/10) and 58% (11/19) of populations had a more negative or similar APC estimate in diabetes compared with non-diabetes (ESM Table 5).

Trends in the excess risk of mortality from diabetes, as measured by mortality rate ratio or standardised mortality ratio, were compared between men and women in two studies, both showing similar declines in men and women over time [6, 32]. One study reported that the mortality rate ratios became lower over time for all age groups except for the group aged <44 years in the Canadian and UK cohorts [6].

When the patterns of mortality trends were examined based on individual studies, declining mortality rates in people with diabetes was observed in 82% (23/28) of studies in 2000–2016, which was similar to the 74% (32/43) of separate ethnic-specific or sex-specific populations showing decreasing trends in this time period.

The median quality score was 8 out of 11 (interquartile range: 6.5–9) (ESM Table 6). We repeated the analyses reported in Fig. 1 after excluding studies with quality scores in the lowest quartile (i.e. a score of 6 or below), and observed similar patterns in the 26 studies that were deemed high quality, with 77% (30/39) of populations showing declining mortality trends in people with diabetes in 2000–2016.

A key strength of this work is its systematic approach and robust methodology, though several potential limitations deserve consideration. First, there was variation among studies in the definition of diabetes and not all studies differentiated between type 2 and type 1 diabetes. Since less than 10% of all cases of diabetes are due to type 1 diabetes, our findings should only be generalised to type 2 diabetes. Second, we used joinpoint regression to detect changes in trends over a long study period, and when more than one segment was detected, we reported the trend separately for each different segment. Joinpoint regression uses a robust permutation test to detect the number of segments and has been previously shown to yield an acceptable type I error [14], which means that the probability of ‘false-positive’ detection of multiple segments when none exist is unlikely to be beyond the type I error used (5% in our analysis). However, the power of the test for detecting segments is a function of several factors, namely the magnitude of change in trends between segments, the number of measured time-points and the size of the underlying population, which means that there is a possibility we may have missed more subtle changes in trends in studies conducted in smaller populations with fewer follow ups. However, we noted that the 19 populations with unchanged mortality rates had a similar distribution of sample size to the other 45 populations reporting decreasing trends. Moreover, as 75% of the data sources had sample sizes greater than 10,000 in each time period, the likelihood of misclassification would be small. Third, as the number of deaths among younger adults with diabetes, particularly with type 2 diabetes, was small (reflected by the wide CI of the APC in mortality rates in age group <40 years), it is possible that there was inadequate power to detect real mortality trends in some younger populations. Fourth, the use of multiple populations (ethnic-specific or sex-specific) per study may lead to bias. However, our sensitivity analysis conducted among the total population with diabetes in each study demonstrated similar patterns of mortality change compared with the results based on separate ethnic-specific or sex-specific populations. Fifth, no estimates of mortality trends in diabetes have been reported from low-or middle-income countries. Among the limited studies in non-Europid populations, we found that 86% (6/7) and 43% (3/7) of the populations studied in 1990–1999 and 2000–2016 did not have statistical evidence for a declining mortality, including indigenous populations [20, 21] and US ethnic groups (black, Asian and Hispanic) [8, 19]. This suggests that people with diabetes living in low- and middle-income countries, which share a similar demographic background to the non-Europid populations included in this review, may not have the declining mortality trends found here. Sixth, this review focused on published data and is, therefore, subject to potential publication bias, in which only increasing or only decreasing mortality rates could be perceived as worth publishing. However, since there is a reasonable spread of increasing, decreasing and unchanged mortality trends, it would appear that such a publication bias is not a major issue. Seventh, despite excluding closed cohort studies from our review, the selection bias related to diabetes duration may not be completely removed even in open cohort studies, because surviving participants who were recruited during earlier periods of the study might have a longer duration of diabetes than those recruited from later periods. Eighth, lack of available data precluded analyses relating to cause-specific mortality, diabetes duration or age at diabetes onset, and we were not able to source unit record data by age and sex on all populations studied; therefore, it is not possible to standardise the mortality rates from each individual study to a specific external population. Ninth, the Newcastle–Ottawa scale for assessing the quality of cohort studies was not designed for studies of trends, and may not perform ideally [50]. It is therefore possible that the sensitivity analysis that excluded the studies with scores in the bottom quartile did not correctly identify all of the lowest-quality studies. Tenth, considering the significant heterogeneity observed among the 67 separate ethnic-specific or sex-specific populations and across different subgroups, we did not report the pooled estimates of annual mortality change in people with diabetes or without diabetes. It is also worth noting that since we report multiple trends from within a particular population, the overall heterogeneity shown by our trend estimates represents both within- and between-population heterogeneity and it will be larger than the between-population heterogeneity alone. Finally, although we did not formally search the grey literature, a Google search identified only one additional article, which was from Russia and was of low quality [51].