The Effectiveness of Pharmacological and Non-Pharmacological Interventions for Improving Glycaemic Control in Adults with Severe Mental Illness: A Systematic Review and Meta-Analysis

People with severe mental illness (SMI) have reduced life expectancy compared with the general population, which can be explained partly by their increased risk of diabetes. We conducted a meta-analysis to determine the clinical effectiveness of pharmacological and non-pharmacological interventions for improving glycaemic control in people with SMI (PROSPERO registration: CRD42015015558). A systematic literature search was performed on 30/10/2015 to identify randomised controlled trials (RCTs) in adults with SMI, with or without a diagnosis of diabetes that measured fasting blood glucose or glycated haemoglobin (HbA1c). Screening and data extraction were carried out independently by two reviewers. We used random effects meta-analysis to estimate effectiveness, and subgroup analysis and univariate meta-regression to explore heterogeneity. The Cochrane Collaboration’s tool was used to assess risk of bias. We found 54 eligible RCTs in 4,392 adults (40 pharmacological, 13 behavioural, one mixed intervention). Data for meta-analysis were available from 48 RCTs (n = 4052). Both pharmacological (mean difference (MD), -0.11mmol/L; 95% confidence interval (CI), [-0.19, -0.02], p = 0.02, n = 2536) and behavioural interventions (MD, -0.28mmol//L; 95% CI, [-0.43, -0.12], p<0.001, n = 956) were effective in lowering fasting glucose, but not HbA1c (pharmacological MD, -0.03%; 95% CI, [-0.12, 0.06], p = 0.52, n = 1515; behavioural MD, 0.18%; 95% CI, [-0.07, 0.42], p = 0.16, n = 140) compared with usual care or placebo. In subgroup analysis of pharmacological interventions, metformin and antipsychotic switching strategies improved HbA1c. Behavioural interventions of longer duration and those including repeated physical activity had greater effects on fasting glucose than those without these characteristics. Baseline levels of fasting glucose explained some of the heterogeneity in behavioural interventions but not in pharmacological interventions. Although the strength of the evidence is limited by inadequate trial design and reporting and significant heterogeneity, there is some evidence that behavioural interventions, antipsychotic switching, and metformin can lead to clinically important improvements in glycaemic measurements in adults with SMI.


Introduction
People with severe mental illness (SMI) (schizophrenia and other illnesses characterised by psychosis) have a lower life expectancy compared with the general population by around 15 to 20 years [1]. A higher prevalence of comorbid conditions (e.g. diabetes and cardiovascular disease) and poorer management of physical health contribute to this health inequality [2]. Around 13% of people with SMI have diabetes compared with 6% of the general population, and the difference is increasing [3]. As diabetes interventions are scaled up for the general population, these inequalities may increase further. This is because generic interventions are unlikely to be suitable for people with SMI due to the complex combination of psychological, social and financial barriers they face in managing their health [4].
Although there are more than 40 published systematic reviews of studies targeting physical health in people with SMI, these have focused mainly on anthropological outcomes [5][6][7][8], with few investigating diabetes prevention and treatment [9,10]. It is well-established that modest improvements in glycated haemoglobin (HbA 1c ) and blood glucose levels can avoid onset of diabetes and have a significant impact on preventing diabetic complications in the general population [11]. A few reviews have investigated the effect of pharmacological [6,12] and behavioural [7,8,13] interventions on these glycaemic measurements in people with SMI. An older review investigated both pharmacological and behavioural interventions [14]. However in all of these, glycaemic effects were examined as a secondary outcome only. This makes it difficult to determine which interventions are effective for improving glycaemic control in people with SMI. The aim of this systematic review and meta-analysis is to identify pharmacological and behavioural interventions for improving diabetes outcomes that have been tested in the adult SMI population, and to determine their effectiveness in lowering HbA 1c and fasting blood glucose [15].

Eligibility criteria
We included randomised controlled trials (RCTs) of interventions to improve diabetes outcomes for adults (aged 18 years and over) with SMI. We defined SMI as schizophrenia, bipolar disorder, psychosis or other non-organic psychotic disorders, including schizoaffective disorder and severe depression. To be included, studies had to measure at least one of the following outcomes: i) in people without diabetes at baseline: incidence of diabetes, HbA 1c or fasting glucose; and ii) in people with diabetes at baseline: HbA 1c , fasting glucose, weight, body mass index (BMI), or diabetic complications.
We restricted studies to those published in peer reviewed journals and the English language.
The protocol for the review has been published on the International Prospective Register of Systematic Reviews (PROSPERO), registration number CRD42015015558 [15]. We carried out the review in accordance with the PRISMA guidelines (see S1 PRISMA Checklist).

Search strategy
The search strategy comprised three concepts: 'diabetes', 'SMI', and 'RCTs or systematic reviews'. An example of the strategy is provided in the supporting information (see S1 Appendix).
Literature searches were performed in CINAHL (EBSCO); Embase Classic+Embase (Ovid); PsycINFO (Ovid); Ovid Medline; PubMed; Cochrane Database of Systematic Reviews (Wiley) and Central Register of Controlled Trials; Database of Abstracts of Reviews of Effect (Wiley); and Conference Proceedings Citation Index (Thomson Reuters). We also searched three trial registries (ClinicalTrials.gov, International Clinical Trials Registry Platform (WHO), ISRCTN registry). Searches were performed on 12/12/2014, and updated on 30/10/2015 (except for trial registries).

Study selection
Search results were managed in EndNote version 7 software. Citations and abstracts were screened to exclude studies that did not meet the selection criteria. References of relevant reviews identified during the screening process were also searched. Relevant full-text articles were retrieved and assessed for eligibility; missing data to help assess eligibility were sought from corresponding authors.

Data extraction and synthesis
Study characteristics and data for meta-analysis were extracted into a tailored and piloted data collection form [15]. Multiple reports from the same study were linked and missing data were requested from study authors. The Cochrane Collaboration tool was used to assess risk of bias [16]. All stages of study selection and data extraction were conducted independently by two reviewers, with discrepancies resolved through discussion and where consensus could not be reached, arbitration by a third reviewer.
Due to the heterogeneity of diabetes interventions, we categorised interventions as pharmacological, non-pharmacological or mixed (interventions combining medication with a nonpharmacological approach) [15]. Pharmacological interventions were further sub-grouped into categories: i) diabetes medications (including metformin, sulphonylureas, insulin and thiazolidinediones); ii) weight loss treatments (including antiparkinsonian, anticonvulsant and antidepressant medications thought to promote weight loss, as well as anti-obesity drugs and appetite suppressants); iii) combinations of weight loss and diabetes medications; iv) switching antipsychotic medication; and v) an 'other' category.
Non-pharmacological interventions were categorised as behavioural (targeting a change in an individual's behaviour) or organisational (targeting a change in the environment or organisation of care).
We planned to explore effectiveness of interventions in prevention of diabetes. However, many studies did not distinguish between people with and without diabetes at baseline. Of the studies that excluded people with diabetes at baseline, none measured incidence of diabetes or reported data that would enable us to estimate this. We therefore pooled the results across all studies for glycaemic control, using outcome data for HbA 1c and fasting glucose.
We analysed pharmacological and non-pharmacological interventions separately, and because we expected significant heterogeneity between studies, we used random-effects meta-analysis and assessed for heterogeneity using the I-squared statistic. To allow combining of post-intervention and change scores for outcomes, and since outcomes were reported consistently, we calculated the unstandardised difference in means (MD) [16].
To assess effects across key intervention characteristics, we conducted subgroup analyses for pharmacological interventions by type of drug category; and for behavioural interventions by duration (short ( 6 months) or long (>6 months)), and whether or not interventions included repeated physical activity. We also conducted univariate random effects meta-regression using intervention duration as a continuous variable (number of weeks). Both duration and physical activity have been identified as key components of effective diabetes interventions in the general population [17].
To explore potential differential effects in people with and without diabetes, we conducted separate subgroup analyses, for i) studies excluding participants with diabetes, and ii) those that only included people with diabetes and SMI or did not specify diabetes status. We also conducted univariate random effects meta-regression using mean HbA 1c or fasting glucose at baseline to explore whether or not this explained some of the heterogeneity among studies [18].
To investigate possible baseline imbalance observed during data extraction, we repeated the main meta-analyses using mean difference at baseline [19]. We explored the impact of study quality and heterogeneity by undertaking sensitivity analyses, using 'leave-one-out' analyses to test if single studies had a disproportionate effect on the results. We used the trim-and-fill method and inspection of funnel plots to investigate publication and small study bias [20]. The trim-and-fill analysis adjusts for any funnel plot asymmetry and provides an effect size estimate that takes account of observed publication bias.
Comprehensive Meta-Analysis (CMA) version 2 software was used for all statistical analyses.

Results
A total of 3,721 citations were identified by database searches, and a further 27 articles from the reference lists of systematic reviews. After removing duplicates, 2,278 records were screened for relevance by title and abstract, and 197 full text articles retrieved. Of these, 104 did not meet the selection criteria and were excluded. The remaining 93 articles described 73 studies. Nineteen of these were ongoing studies (see S1 Table).
A total of 54 studies were included in the systematic review . Six of these studies did not provide usable data for the meta-analysis [26,32,42,44,46,51]. Study characteristics are summarised in Table 1.

Participants
There were a total of 4,392 participants; 2,315 were assigned to intervention and 2,077 to control arms. Participants were mainly drawn from mental health outpatient and inpatient settings (n = 47 studies). One study recruited from supported housing schemes [24], and one from residential care facilities and day programmes [29]. Five studies did not report setting [47,57,68,69,71].
All but three studies [48,57,71] included both men and women, although overall women were under-represented (41%). Mean age ranged from 25 to 53 years, with a mean age across studies of 43 years. Ethnicity was poorly recorded, but varied significantly due to the range of countries included.    Eighteen studies recruited participants with schizophrenia; 20 with schizophrenia and schizoaffective disorder; two with bipolar disorder; and 14 with various SMIs. The majority of studies included clinically stable participants who had been diagnosed for several years.
Inclusion of participants with diabetes varied. Only one study specifically recruited people with type 2 diabetes [29]. Twenty-three studies excluded participants with diabetes (one excluded type 1 diabetes [59]). The remaining studies did not specify this in eligibility criteria. Mean HbA 1c at baseline ranged from 4.1% to 7.4% (n = 26 studies); 13 studies reported a mean in the American Diabetes Association (ADA) pre-diabetes category (5.7-6.4%) and two in the diabetes range (!6.5%). Mean fasting glucose ranged from 4.3 to 6.8mmol/L (n = 43 studies); 14 studies reported a baseline mean in the ADA high risk category (5.6-6.9mmol/L) [75].
Thirty studies targeted overweight participants or those who had experienced significant weight gain. Mean BMI at baseline ranged from 20.2 to 41.9Kg/m 2 (n = 49 studies); 41 studies reported a mean BMI of over 25Kg/m 2 .

Interventions
Of the 54 RCTs identified, 40 assessed a pharmacological, 13 a non-pharmacological and one a mixed intervention.
Among the non-pharmacological studies, six compared an intervention with usual care [21,22,25,27,30,32], and three provided basic information or advice to controls at baseline [23,26,29]. In one study, the intervention was also given to the control group after week 12 of 24 weeks [28]. Two studies included an active control arm [24,31]. One trial did not describe the control intervention [33].
The mixed intervention study included four arms: metformin, metformin plus a lifestyle intervention, lifestyle plus placebo, and a control arm receiving placebo alone [34].
Pharmacological interventions. In total, 23 different medications from 15 categories of drug were evaluated (see Table 2).
Intervention duration varied from four weeks to 12 months, being 6 months or less in most studies.
Non-pharmacological interventions. All 14 non-pharmacological interventions targeted change in individual behaviour rather than organisation of care. Interventions were variously described as lifestyle interventions, weight loss programmes and physical exercise programmes; however, there was considerable overlap between these categories. In total, eight interventions included regular exercise sessions [23-25, 27, 30, 31, 33, 34] and three restricted calorie intake [28,33,34]. All but one intervention [31] included dietary recommendations, and all but two [31,33] employed educational and behavioural strategies promoting a healthier lifestyle.
Staff delivering interventions varied, but the majority were mental health staff. No intervention specifically included carers of participants, although in one, carers were invited to join a session [21].

Insulin
Diabetes treatment used to regulate carbohydrate and fat metabolism in the body.

Metformin
A biguanide used to prevent and treat Type 2 diabetes by increasing insulin sensitivity and reducing the amount of glucose produced and released by the liver.
Pioglitazone, Rosiglitazone Thiazolidinediones help to regulate glucose and fat metabolism by improving insulin sensitivity, allowing insulin to work more effectively. Rosiglitazone has been withdrawn from the EU. Amantadine A dopamine agonist approved to treat extrapyramidal side effects and parkinsonian, with potential to decrease prolactin (plays a role in metabolism) or to decrease appetite.

Orlistat
An anti-obesity drug which acts on the gastro-intestinal tract by reducing absorption of dietary fat.

Reboxetine
An anti-depressant, believed to promote weight loss by inhibiting serotonin re-uptake and by doing so regulate eating behaviour and appetite control.
Sibutramine A centrally acting appetite suppressant which has been withdrawn from the UK and other countries.
Topiramate, Zonisamide Anticonvulsant (epileptic) medication believed to promote weight loss by stimulating energy expenditure and decreasing body fat stores (by inhibiting carbonic anhydrase).

Weight loss and diabetes combination n = 2 [38, 56]
Amantadine + metformin + zonisamide, Metformin + amantadine + zonisamide Treatment algorithms of anti-diabetic, anti-parkinsonian and anti-epileptic medications to allow patients to switch between treatments depending on clinical response.
Metformin + sibutramine Adding an appetite suppressant to an anti-diabetic may enhance weight loss potential. Aripiprazole, Quetiapine, Ziprasidone Switching to or adding an atypical antipsychotic associated with fewer metabolic side effects is hypothesised to alleviate weight gain and metabolic abnormalities caused by the more commonly used antipsychotics like olanzapine and clozapine.
Olanzapine orally disintegrating The orally disintegrating form of olanzapine is argued to induce fewer metabolic side effects than the standard tablet.
Dehydroepiandrosterone (DHEA) A steroid hormone with systemic anti-atherosclerotic properties which help to increase insulin sensitivity and prevent development of metabolic syndrome components.

Fluvoxamine
An anti-depressant used in combination with clozapine could help to reduce clozapine dose thereby alleviating APM induced weight gain and metabolic side effects.
Melatonin, Ramelteon Hypnotics used to treat insomnia which act on the circadian rhythm (normal sleep-wake cycle) and are believed to be important metabolic regulators.

Memantine
Used to treat dementia, memantine has an antidepressant like and mood stabilizing effect and is believed to reduce binge eating episodes and weight.

(Continued)
Intervention duration varied from 12 weeks to 18 months, with the majority being between 4 and 6 months. Group sessions were provided in 12 interventions, 4 of which also included individual sessions or follow-up calls. Sessions varied from 30 minutes to 2 hours in length, with frequency ranging from 3-times weekly to once a month.

Outcomes
Our primary outcomes of interest were HbA 1c and fasting glucose.
Nineteen pharmacological studies measured both of these outcomes. A further five measured HbA 1c (one did not provide data) [42], and 16 measured fasting glucose (one did not provide data [51] and one provided dichotomous data that were not useable in meta-analysis [46]).
Three behavioural studies measured both HbA 1c and fasting glucose [24,29,30]. One of these did not provide data for HbA 1c [30], and one reported log transformed data for fasting glucose which were not useable in meta-analysis [29]. One study measured HbA 1c only [25], and the other nine studies measured fasting glucose only (one did not provide data [32] and one provided dichotomous data that were not useable [26]).
The mixed intervention study only measured fasting glucose. All studies measured HbA 1c and/ or fasting glucose at the end of the intervention period. Details of the primary outcomes and follow-up period for each study are shown in Table 1.

Risk of bias
The risk of bias assessment for each study is provided in S2 Table. Only one study was assessed as low risk across all domains [74]. Reporting of trial design was limited in many studies. Attrition was a particular problem for behavioural interventions and also for antipsychotic switching trials, many of which reported higher discontinuation rates in the intervention compared to control groups.

Effectiveness of interventions
For HbA 1c , six of 28 (five pharmacological and one behavioural), and for fasting glucose, nine of 48 studies (five pharmacological, three behavioural and one mixed intervention) showed improvement in the intervention group compared to the control interventions. The remainder reported no difference between groups (see Table 1).

Meta-analysis
In the 48 trials included, there were a total of 4,052 participants; 2,150 were assigned to intervention and 1,902 to control arms. Pharmacological interventions. For pharmacological interventions, we pooled data from 22 studies for HbA 1c (n = 1515) and 34 for fasting glucose (n = 2536) (see Fig 2).
For fasting glucose there was a small but statistically significant improvement of -0.11 mmol/L (95% CI, [-0.19, -0.02]; p = 0.02) for the intervention group compared to controls. Again, there was heterogeneity (I 2 = 57%). Investigation of baseline imbalance (see S1 Fig) showed that the control group had slightly lower levels of fasting glucose (MD = 0.07mmol/L; Improving Glycaemic Control in Adults with Severe Mental Illness 95% CI, [0.01, 0.14]; p = 0.03), a difference that while statistically significant, was very small, and if anything would lead to underestimation of the overall effect size.
Subgroup analyses of studies that excluded participants with diabetes at baseline, and those that did not, showed that pharmacological interventions were effective in lowering HbA 1c only in the mixed population (MD = -0.11%; 95% CI, [-0.21, -0.01]; p = 0.04; I 2 = 59%). For fasting glucose, neither group showed a statistically significant improvement compared to controls ( Table 3). The meta-regression found no association between baseline HbA 1c or fasting glucose levels and effect size (see S3 Fig).
To explore this further, we repeated the subgroup analysis for certain categories of pharmacological interventions: diabetes medication, weight loss medication and antipsychotic switching for fasting glucose; and diabetes medication for HbA 1c . No group showed statistically significant improvements compared to controls (Table 3). We observed larger effect sizes in studies that did not exclude diabetes at baseline for the diabetes medication and weight loss medication categories, but similar effects for antipsychotic switching (Table 3). We did not have sufficient data to examine the remaining categories.

Sensitivity analyses
Leave-one-out analyses showed that no single study had a disproportionate effect on each of the main meta-analyses. However, funnel plots showed some asymmetry (see Fig 4), suggesting potential publication bias for both the behavioural and pharmacological literature. The trim-and-fill analysis suggests there is some evidence of missing studies (shown as black on the funnel plots in Fig 4).The adjusted effect sizes, accounting for publication bias are presented in Table 4. Publication bias adjusted effect sizes suggest that pharmacological interventions reduce both HbA 1c and fasting glucose, and behavioural interventions are effective in reducing fasting glucose but not HbA 1c .

Summary of evidence
Overall, compared to usual care, both pharmacological and behavioural interventions improved fasting glucose levels, but not HbA 1c in people with SMI, with behavioural interventions showing a larger difference compared with pharmacological interventions. However, after adjusting for publication bias, there was some evidence that pharmacological interventions may also improve HbA 1c . Subgroup analyses showed improvements in HbA 1c for antipsychotic switching and metformin; and in fasting glucose for metformin. For behavioural interventions, those that included regular physical activity were more effective in lowering fasting glucose than those that did not. Subgroup analysis and meta-regression showed that interventions of longer duration resulted in greater improvements in fasting glucose compared to Improving Glycaemic Control in Adults with Severe Mental Illness usual care, and this may help to explain why the small number of studies measuring HbA 1c did not show an improvement, as only one of these was greater than 6 months in duration.  Some categories of pharmacological interventions (diabetes and weight loss medications), appeared to have a smaller effect on lowering glycaemic measurements in studies that excluded people with diabetes at baseline compared to the effect observed in studies that did not. However, it was not possible to investigate this robustly because of limited data, and the metaregression of all pharmacological interventions showed no association between baseline levels of HbA 1c or fasting glucose and effect size. For behavioural interventions, studies that included participants with higher baseline glucose levels appeared to be more effective in a meta-regression, although the subgroup analysis showed no difference between studies that excluded those with diabetes compared to those that did not.
In common with these previous reviews, we found the improvements reported in HbA 1c and fasting glucose were modest. However, there was considerable heterogeneity in results. Differences in effect sizes and direction of effect between studies made it difficult to assess the overall effectiveness of interventions. Several studies showed a reduction in fasting glucose and at the same time an increase in HbA 1c or vice versa [37-39, 43, 54]. These results are difficult to explain because logically one would expect a corresponding change, particularly in longer duration studies which would take account of the time required to alter HbA 1c . However, there are a number of trials that demonstrate that although HbA 1c and fasting glucose are well correlated, they do not always respond in similar ways [76].
For people with SMI, this relationship may be complicated further by the metabolic side effects of anti-psychotic medication, which will work against interventions designed to improve glycaemic control [77]. For example, in several of the pharmacological and behavioural intervention studies, fasting glucose or HbA 1c increased in both the intervention and control groups, but with a smaller increase in the intervention group [24,31,43,49,56]. Through subgroup analysis and meta-regression, we have been able to identify certain intervention and population characteristics that may explain some of the differences in effect between studies, and identify particular interventions that show the most promise. However, these findings should be viewed within the context of methodologically limited trials, and for the antipsychotic switching and behavioural interventions, substantial dropout in follow-up.

Limitations
Although we included a larger number of studies compared with previous reviews, a limitation of our findings relates to the quantity and quality of evidence included, and the substantial risk of potential bias identified in included studies. We were also unable to fully explore differential effects between those with and without diabetes, or to compare our findings to evidence in the general population because of the lack of data to measure onset of diabetes in those without diabetes, and diabetic complications in those with diabetes. Previous reviews have also commented on the paucity and poor quality of evidence in this area [10,14]. Strengths of our review include a published protocol, robust search, independent screening and data extraction by at least two reviewers, and the use of appropriate meta-analytic methods to explore the results.

Implications for clinical practice
These results suggest that antipsychotic switching strategies, metformin, sustained behavioural interventions, and behavioural interventions that include regular physical activity offer the greatest potential to improve glycaemic control in the SMI population. Whilst the effect sizes were modest, such improvements in glycaemic control can help to avoid onset of diabetes and attenuate diabetes complications [11], therefore, the small differences reported in key subgroups may still be clinically significant. Also, combining pharmacological and behavioural strategies may incrementally (or perhaps even synergistically) increase effectiveness [34]. However, the effect sizes observed were modest when compared to the general population [18,78], suggesting that tailored interventions which address the specific challenges faced by people with SMI are needed.
In real world settings, the SMI population will face challenges in adhering to new medications or engaging in sustained behavioural interventions involving attendance at regular group sessions [23,79]. These challenges will likely be compounded when implementing multifaceted interventions. Moreover, we need to reflect carefully before pursuing adjunctive pharmacological therapies in a population for whom polypharmacy is already problematic; and the potential acceptability of switching from an antipsychotic medication providing clinical stability to one which may help to improve physical health, but for which the efficacy in preventing relapse in mental illness is uncertain. These considerations, along with the sparse evidence base, mean that recommendations for clinical practice remain limited. Nonetheless, this review does provide some evidence to support current practice of providing lifestyle interventions and switching to antipsychotics with a better metabolic profile in people with SMI.

Conclusions
Improving diabetes outcomes in SMI is a global priority, but the evidence-base to guide clinical practice is limited. Despite the challenges described above, a number of pharmacological and behavioural approaches warrant further exploration. Metformin is already a well-established treatment in diabetes [17]. Its use alongside antipsychotic prescriptions to prevent diabetes merits further investigation. Switching of antipsychotic medication is also common in clinical practice. Research is needed to understand which antipsychotics offer the greatest potential benefit, and to optimise dosage and timing of such interventions in order to reduce glycaemic burden, whilst maintaining clinical stability for people with SMI.
Behavioural interventions show perhaps more promise than pharmacological strategies, but little is known about the behaviour change techniques that might be most effective for people with SMI and diabetes. This is a key area for research, if we are to avoid ever-increasing inequalities in health and access to healthcare, as diabetes management becomes increasingly predicated on self-management. Future research should focus on the design of appropriate interventions, and test the potential acceptability and feasibility of delivering them in a real world setting, before establishing effectiveness in a trial evaluation.