Data sources and searches

The following databases for health sciences were systematically searched from inception until 28 March 2017: MEDLINE; CINAHL; Cochrane Library; and EMBASE. Our search terms combined the population with the intervention. The medical subject headings used were ‘Diabetes Mellitus, Type 1’, ‘Insulin Dependent Diabetes Mellitus’, ‘Insulin Deficiency’, ‘Diet, Carbohydrate-Restricted’, ‘Ketogenic Diet’, ‘Low Carbohydrate Diet’, ‘Diet, Low Carbohydrate’, and ‘Diet, Ketogenic’. The complete search strategy for Ovid MEDLINE is shown in S1 Table. The search was restricted to human studies published in English. Citations and abstracts of all papers retrieved from these searches were downloaded into Endnote reference management software (Endnote X7.7.1, Thomson Reuters 2016). Reference lists of included studies were also searched.

Study selection

Two reviewers (JT & KR) independently screened titles and abstracts of all retrieved records for obvious exclusions. The reviewers (JT & KR) then independently assessed the remaining papers based on full text, applying pre-specified eligibility criteria for included studies. Disagreements were resolved by consensus through adjudication with a third independent researcher. Studies included in the review had to be primary research studies of interventions or exposures including controlled trials, cohort-type studies and case-control trials. Case-series and case-reports with a methods section could also be included. In the case of multiple reports from the same study, we used the most complete or recently reported data. The studies had to quantitatively measure the effects of a dietary intervention or observed mean intake below the AMDR for carbohydrate (i.e., <45% total energy as dietary carbohydrate) in adults and/or children with type 1 diabetes that were otherwise healthy. The minimum duration of the intervention or exposure period was two weeks and the study had to adequately report on at least one of the following outcomes; HbA1c, severe hypoglycaemia, total daily insulin dose, body mass index (BMI), quality of life and mean daily blood glucose.

Data extraction and quality assessment

Data extraction was carried out by JT using a piloted customized data extraction form adapted from the Cochrane Collaboration [15] (S2 Table). For studies investigating different levels of carbohydrate restriction, the lowest reported or prescribed level of dietary carbohydrate intake was considered the intervention and the highest level was considered the comparator. One reviewer (JT) contacted authors for relevant data that were missing from the included papers (S3 Table).

Risk of bias assessments were conducted for methodological quality of each included study using the critical appraisal tool most appropriate for its design. For randomised controlled trials, the Cochrane Collaboration’s Risk of Bias tool for randomised studies was used [16]. This assesses bias as ‘low risk’, ‘high risk’ or ‘unclear risk’ across seven domains. For specificity, we separated blinding of participants and blinding of personnel into two separate domains. For pre-post intervention studies, the National Institute of Health’s quality assessment tool for before-after studies with no control group was used [17]. This tool evaluates potential flaws in study methods or implementation using twelve closed questions. The ratings (yes/no/other) on the different items are then used by reviewers to assess overall risk of bias as ‘good’ (low risk of bias), ‘fair’ (susceptible to some bias) or ‘poor’ (significant risk of bias). For case-series and case-reports, we used the critical appraisal checklists from the Joanna Briggs Institute [18]. These checklists are a series of 8 to 10 closed questions (yes/ no/unclear/not applicable) which help form an overall appraisal for each study assessed. For standardisation, we used this assessment to classify studies as ‘low risk’, ‘high risk’ or ‘unclear risk’ of bias. All risk of bias assessments were performed at the study level. However, when information was specifically related to outcome measures (e.g., ‘blinding of outcome assessment’) judgement was made according to our primary outcome, HbA1c. If HbA1c was not measured, the next reported secondary outcome was used. If a decision could not be reached on bias assessments, an additional investigator made a decision.

The GRADE approach was used to assess the quality of the body of evidence at the outcome level for our primary outcome, HbA1c [19]. This involved consideration of within-study risk of bias, consistency, directness of evidence, precision of effect estimates and risk of publication bias. This approach specifies four levels of quality; ‘high', 'moderate', 'low' and 'very low’.

Data synthesis and analysis

To summarise the effects of low-carbohydrate diets on type 1 diabetes outcomes in controlled trials, we extracted mean outcome values for the intervention and control groups at baseline and follow-up. For other studies with only an intervention group or for trials where only one participant group was relevant to our study, we extracted mean outcome values for the intervention or observed group at baseline and follow-up. Standard deviations and/or standard errors, sample sizes, follow-up time and published levels of significance (i.e., P-values) were also taken. If no P-value was published and raw outcome data were available, the P-value was calculated using the R Statistical Language (R version 3.4.0, R Foundation for Statistical Computing 2017). Results were considered statistically significant if P < 0.05 and non-significant if P ≥ 0.05. Standard errors were converted to standard deviations by SD = SE x √n.

A meta-analysis was not able to be conducted due to obvious heterogeneity and we used text and tabular format to summarise the outcome data of all low-carbohydrate diets. Studies were classified into one of three groups, determined by the dietary carbohydrate amount of the intervention prescription where a g/day or %TEI value was available. Where no specific prescription was available and compliance to the intervention was reported, studies were classified according to the reported dietary intake data of participants. Very Low Carbohydrate Ketogenic Diet (VLCKD) studies are those in which the intervention is less than 50 grams of total carbohydrate per day. True Low Carbohydrate Diet (TLCD) studies are those in which the intervention is 50–130 grams of total carbohydrate per day. False Low Carbohydrate Diet (FLCD) studies are those in which the intervention is below the AMDR for carbohydrate (i.e., <45% total energy) but does not meet the ADA criteria for a low-carbohydrate diet (i.e., <130 grams of total carbohydrate per day or <26% total energy as carbohydrate).

Literature search results

The database search identified 2724 possibly relevant studies that were screened by titles and abstracts (Fig 1). A further 2645 records were excluded with 79 full-text articles subsequently assessed for eligibility. Six articles were excluded because the intervention was less than two weeks, 10 studies were excluded because the study design was a review, letter or cross-sectional, seven studies were excluded because no type 1 diabetes mellitus sub-group were analysed, 27 studies were excluded because there was no low-carbohydrate diet intervention, 12 studies were excluded because inadequate dietary data were reported, seven studies were excluded because there were inadequate measurement and/or reporting of outcome data, and one study was excluded because an updated version of the report/data exists. Eleven additional records were identified through searching the reference lists of included studies. A total of nine studies were eligible and included for this review. A full list of excluded studies with reasons is provided (S4 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. PRISMA flow chart.

https://doi.org/10.1371/journal.pone.0194987.g001

Study characteristics

Of the nine included studies, there were two randomised controlled trials [10, 20], four pre-post intervention studies [8, 21–23] two retrospective case-series [24–25], and one case-report [26]. These studies are heterogeneous with respect to study design, sample size, intervention, comparator, outcomes assessed and study quality (Table 1). Year of publication ranged from 1980 to 2016 and the number of participants ranged from 1 to 48. The duration of exposure to a low-carbohydrate diet ranged from two weeks to five years. The two controlled trials compared a low-carbohydrate diet (intervention) to a higher-carbohydrate diet (comparator) using either a crossover [20] or parallel [10] design. All other studies compared a low-carbohydrate diet (intervention) to baseline or usual diet (comparator).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Characteristics of included studies.

https://doi.org/10.1371/journal.pone.0194987.t001

Three studies were classified as FLCD[20, 21, 23], four studies were classified as TLCD [8, 10, 22, 26], and two studies were classified as VLCKD [24–25] (S5 Table). Seven studies [8, 10, 20–21, 24–26] attempted to control for the confounding effect of insulin therapy and two studies [22–23] did not (Table 1).

All outcomes

Results for all primary and secondary outcomes are presented in Table 2. Effect sizes were not calculated because raw outcome data were not available for all studies and most outcomes were inconsistently reported. In one study [8], outcome data was provided for both all participants (I_A) and adherent participants only (I_B) as shown in Table 2. Results for our primary outcome (HbA1c) were available from eight of nine studies reviewed. Results for secondary outcomes of interest were inconsistently reported. Two studies reported the effect of a low-carbohydrate diet on frequency of severe hypoglycaemia [23, 26], five studies reported total daily insulin [8, 10, 22, 24, 26], three studies reported BMI [8, 10, 25], and one study reported mean daily blood glucose [10]. No studies adequately reported change(s) in quality of life. Only results from the most frequently reported outcomes are described below.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Effect of intervention and comparator diets on type 1 diabetes management outcomes (primary and secondary).

https://doi.org/10.1371/journal.pone.0194987.t002

HbA1c

Eight studies investigated the effect of a low-carbohydrate diet on glycaemic control using HbA1c [8, 10, 20–24, 26] (Table 2). One study [26] reported a follow-up value for HbA1c but did not provide baseline data so was not included for this outcome. Four studies [10, 20, 22–23] reported non-significant changes in HbA1c with a low-carbohydrate diet and three studies [8, 21, 24] reported statistically significant reductions (P < 0.05). Of the two studies that compared a low-carbohydrate diet to a higher-carbohydrate diet [10, 20], neither showed a significant difference between groups at follow-up.

Total daily insulin

Of the five studies that reported daily insulin usage [8, 10, 22, 24, 26], two TLCD studies [10, 22] demonstrated statistically significant reductions in total daily insulin within carbohydrate restriction groups (P < 0.05) (Table 2) with one study [10] also reporting a statistically significant difference between the low-carbohydrate group and high-carbohydrate comparator (P < 0.05). Levels of significance could not be calculated or obtained in three studies due to inadequate sample size [26] and lack of raw participant data [8, 24].

Risk of bias assessments

Full assessments of the two RCTs reviewed with support for judgements are included in S6–S8 Tables. Risk of bias in one [20] was rated ‘low’ in four domains and ‘unclear’ in four domains. The other controlled trial [10] was rated ‘low’ in seven domains and ‘unclear’ in one domain. Of the pre-post intervention studies (S9–S13 Tables), two were rated as ‘fair’ quality [8, 21] and two were rated as ‘poor’[22–23]. The poor-quality studies did not attempt to control for the confounding influence of insulin therapy on HbA1c. One case-series [24] had an overall risk of bias of ‘high’ and the other case-series [25] had a ‘low’ risk of bias (S14 Table). Both reports had clear criteria for inclusion, valid methods for identification of type 1 diabetes, clear outcome results of cases and appropriate statistical analyses. The case-report [26] had an overall appraisal of ‘low’ risk of bias (S15 Table).

GRADE assessment

The quality of evidence could not be upgraded from the established level of confidence because the statistical analyses required for such categories (i.e., large magnitude of effect, dose-response gradient, effect of plausible residual confounding) were not performed. The five categories used for downgrading the quality of evidence were assessed and included in Table 3. Categories were assigned a rating of zero if the appropriate statistical analyses required to confidently downgrade the evidence based on the criteria were not able to be performed, or if it could be appropriately justified that the evidence should not be downgraded for that category.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Summary of findings table (GRADE) for primary outcome (HbA1c).

https://doi.org/10.1371/journal.pone.0194987.t003

Clinical significance of results

The United Kingdom Prospective Diabetes Study highlighted the importance of lowering HbA1c to reduce the risk of micro and macrovascular complications in patients with diabetes [1]. Further, Juutilainen and colleagues (2008) reported that in type 1 diabetes patients, a 1% rise in HbA1c increased individual risk of cardiovascular mortality by 52.5% [27]. Cardiovascular diseases are the main cause of morbidity and mortality in diabetes, affecting about 55% of patients, compared to 2–4% of people without diabetes [28]. Our review identified statistically significant improvements in glycaemic control in 3 of 8 studies (n > 1) that reported a change in mean HbA1c with a low-carbohydrate diet [8, 21, 24]. In addition, the importance of non-significant changes in HbA1c (i.e., maintaining glucose levels) reported in the other 5 studies [10, 20, 22, 23] is worth noting considering the natural progression for diabetes is toward poorer glycaemic control, preventing HbA1c from rising could be considered a successful outcome.

In this review, all five studies reporting total daily insulin showed clinically significant reductions with a low-carbohydrate diet [8, 10, 22, 24, 26]. The excessive use of insulin that is often required to achieve glycaemic control in type 1 diabetes increases susceptibility to severe hypoglycaemia and may lead to some measure of hyperinsulinemia [29]. Hyperinsulinemia is associated with; excessive weight gain [30], development of the metabolic syndrome [31], inflammation and atherosclerosis [32], Alzheimer’s Disease [33] and cancer [34]. Findings of the present review suggest that low-carbohydrate intakes may assist in reducing or preventing hyperinsulinemia in type 1 diabetes by decreasing the absolute amount of insulin required for tight glycaemic control.

Insulin therapy is also a major confounder in all studies attempting to examine the effect of a lifestyle intervention on HbA1c in type 1 diabetes. The process of randomisation attempts to control for the potential differences in insulin therapy protocols within a study population. However, in non-randomised studies, controlling for insulin can be difficult. In the current review, insulin protocols of participants are presented in Table 1 as they were reported in the study. We classed these as ‘controlled’ if an attempt was made to standardise the insulin protocol of participants or ‘not controlled’ if researchers did not intervene or flexible protocols were promoted. These classifications contributed to the overall risk of bias, and studies that did not control for insulin [22–23] could not receive low-level judgements (‘low’ risk of bias or ‘good’ and ‘fair’ quality).

Strengths of present review

A major strength of the present review is the inclusion of a wide variety of study designs to capture all the available evidence and effectively serve as a library of all published data on low-carbohydrate diets in type 1 diabetes management to date. Standard systematic reviews that set out to evaluate the effect(s) of an intervention tend to exclude evidence based on factors such as small sample size (e.g., n = 1), lack of data (e.g., reported dietary intake) and study design (e.g., non-randomised). This approach is useful in informing public policy and national dietary guidelines. However, this method of review fails to collect data that could be particularly useful in areas of specialised practice where conclusive evidence is limited.

To assist with the interpretation of this review, we performed rigorous risk of bias and quality assessments of all included papers at the study and outcome levels. Use of the GRADE criteria addressed the inherent bias that could have been introduced with the inclusion of low levels of evidence, such as case-reports [35], which was not necessarily identified in the study-level risk of bias assessments (Table 3). Multiple study-level risk of bias assessments were performed when the review authors considered the original appraisal tool to be potentially insufficient in assessing the quality of a paper. For example, two studies were experimental in design, yet the low-carbohydrate diets we assigned as interventions were more accurately, observed outcomes (i.e., exposures) [21,23]. The original appraisal tool applied is specific to intervention studies [17], and the additional appraisal tool is specific to studies of exposures. The Cochrane Collaboration’s tool for assessing risk of bias in non-randomised studies of exposures (ROBINS-E) [36] was used and though ROBINS-E is currently under development, its application is cited in existing review protocols [37]. Results were considered consistent with the original assessments (S16 and S17 Tables).

Limitations of review

Several limitations of our study should be acknowledged. Our search only focused on four online databases. As a result, studies from other relevant journals or grey literature may have been missed. The decision to include studies ‘from inception’ meant that many corresponding authors could not be contacted to retrieve data. For the same reason, methods of outcome measurements may not have been as accurate or reliable as methods used in more recent studies. In addition, we were unable to calculate meaningful effect sizes or conduct a meta-analyses due to the inclusion of multiple study designs, small sample sizes and reports with inadequate reporting of raw participant data and/or standard deviations. It should also be acknowledged that HbA1c is generally considered a three-month average and studies with a follow-up of less than three months may not have been sufficient to detect a true effect. Nevertheless, inadequate follow-up periods were addressed in the risk of bias assessments.

Some limitations also arose from the inclusion of studies with missing or inadequate reported dietary data of participants. A major potential flaw being that the current review did not explicitly exclude studies where dietary carbohydrate was increased from a very low baseline intake to a level that remained below 45% total energy at follow-up. However, the studies without baseline dietary data included in this review provided enough information to confidently assume that carbohydrate was decreased during the intervention or observation period. Chantelau et. al. (1982) [21] described the diet of participants as it were prior to the study in the text of the report, while three studies [8, 24, 25] explicitly identified the intervention as a “carbohydrate-restricted” diet (S5 Table).

In addition, compliance to the intervention could only be assessed in three studies [8, 10, 20] where both a rigid carbohydrate prescription and adequate reported dietary data of participants were available. Anderson et al. (1991) [20] exhibited excellent compliance among participants, while Krebs et al. (2016) [10] describes a population that exceeded the carbohydrate prescription by more than 20% (+28 g/day) and still fit the classification of the intended intervention. Nielsen et al. (2012) [8] presented results separately depending on the level of adherence to the intervention of participants, yet no reported dietary data was provided.

Recommendations for further research

With such limited data available for this review, more high-quality studies are necessary to further inform practitioners of patients with type 1 diabetes on the effect(s) of reducing dietary carbohydrate. Data presented here suggests a particular focus on TLCD and VLCKD designed interventions may provide positive steps forward for reductions in HbA1c. Future research should more adequately consider the potential consequences of interventions that aim to reduce HbA1c in this population. Interventions should only be considered effective if HbA1c can be reduced toward target levels without increasing severe hypoglycaemic episodes, total daily insulin, BMI (≥25 kg/m²), mean blood glucose, and/or negatively affecting quality of life. Therefore, outcomes of this review should form the minimum set of outcomes that are reported in future type 1 diabetes research. None of the studies in this review measured all six outcomes and only four studies [8, 10, 22, 24] reported complete measurements for both HbA1c and insulin dose.

Another important consideration for future studies is whether to substitute carbohydrate with fat or protein. Vernon et al. (2003) [25] and O’Neill et al. (2003) [24] specifically substituted carbohydrate with fat, while Bernstein (1980) [26] and Ireland et al. (1992) [22] increased protein intake. Nielsen et al. (2012) [8] appeared to increase both fat and protein in relative proportions. This review is unable to draw any conclusion for potential differences in effect and more primary studies are necessary.

Conclusion

Type 1 diabetes is a chronic disease with severe complications for its mismanagement. To strengthen patient-centered care and improve individual capacity for problem solving and self-management, health professionals should be equipped with the appropriate evidence base to present multiple management strategies to their patients. Dietary strategies can serve as effective adjuncts to pharmaceutical therapy in the treatment of various metabolic diseases.

This systematic review presents all available evidence for low-carbohydrate diets in the management of type 1 diabetes mellitus. The existing body of evidence is limited and more primary studies evaluating the short and long-term effects of low-carbohydrate diets on type 1 diabetes management outcomes are necessary to support its use in practice.

Assistance

Thank you to Associate Professor Anna Rangan for assisting during the protocol stage and adjudication of differences between JT and KR during screening.

Thank you to Prof. Bero (Chair of Medicines Use and Health Outcomes, University of Sydney) for assisting in the selection of risk of bias / critical appraisal tools and offering ongoing support for this part of the review process.

Thank you to Monica Cooper (Academic Liaison Librarian, University of Sydney) for assisting with the online database searches.

Named contact (Guarantor)

Dr Kieron Rooney, kieron.rooney@sydney.edu.au (+612 9351 9135), Discipline of Exercise and Sport Science, Faculty of Health Sciences, University of Sydney. 75 East Street, Lidcombe NSW 2141. Australia.

Funding sources/sponsors

None.

Conflicts of interest

Jessica Turton completed an internship (2016) at a private practice that supports the use of low-carbohydrate diets.

Dr. Ron Raab is the President of Insulin for Life Australia (https://www.insulinforlife.org.au) and past Vice-President of the International Diabetes Federation.

Dr. Kieron Rooney has given talks for "Low Carb Down Under" on the biochemistry of low carbohydrate diets and has been a collaborator on primary research investigating the effect of lower carbohydrate diets for weight loss. A full disclosure of previous funding and published research is available at http://sydney.edu.au/health-sciences/about/people/profiles/kieron.rooney.php.

These competing interests do not alter our adherence to PLOS ONE policies on sharing data and materials.