Background
The prevalence of diabetes, particularly type 2 diabetes, is rapidly increasing worldwide, particularly in low- and middle-income countries (LMICs). Around 537 million adults globally lived with diabetes in 2021, an increase of 70 million since 2019 [
1]. The International Diabetes Federation (IDF) estimated that the number will increase by 46%, i.e., 783 million adults, by 2045 [
1]. However, this projection may be underestimated considering the unintended effect of control strategies associated with multiple waves of Coronavirus Disease 2019 (COVID-19) on diabetes risk distribution, for example, increased stress and low physical activity levels due to lockdown [
2]. Although 79% of persons with diabetes live in LMICs, total health expenditure on diabetes remains lower in LMICs than in high-income countries (HICs) because of resource constraints and lower priority given to chronic diseases [
1,
3].
Additionally, diabetes accounts for a large proportion of morbidity and mortality in LMICs. IDF indicates that the Western Pacific and South-East countries experienced the highest diabetes fatalities between 2010 and 2019 [
1]. Diabetes is one of the leading causes of death in LMICs among people aged 20–99 years, with the majority of deaths occurring in people under 60 years of age [
1]. According to the IDF Atlas, Africa recorded more than a 10% increase in diabetes deaths between 2010 and 2019. Compared with deaths from other territories, most people who died in Africa were below 60 years old. South Africa and the Democratic Republic of Congo were the most affected by diabetes mortalities [
1].
Studies have reported the potential of non-pharmacological (NP) population-based strategies to reduce the diabetes burden [
4], but most of these studies are large-scale randomized control trials (RCTs) and cohort studies conducted in high-income countries, and transferability issues may limit the application of the evidence in LMICs. NP diabetes interventions are public health community-based strategies targeted at persons with or without diabetes to control or prevent the disease. They include fiscal policies, legislation in – but not limited to – trade and agriculture, health promotion activities or changes to physical environments that can influence modifiable risk factors, e.g., physical activity and dietary patterns. RCTs and cohort studies can generate evidence of the economic and health effects of NP diabetes interventions for policy decision-making [
5], but their application can be challenging from an ethical, implementation time and cost perspective. Ethical issues, including – but not limited to – delaying life-saving intervention in human subjects, could constitute a breach of duty of care; consequently, ethics committees may deny such studies approval. In LMICs, large-scale RCTs and cohort studies may be challenging to fund.
Decision analytical models (DAMs) overcome these challenges by using mathematical and logical relationships to abstract real-world phenomena for investigation and simulation. DAMs are most useful where the lack of data due to a rare event, legal circumstances, time, technical, ethics and funds prevent real-live studies. LMICs can benefit from their use to prioritise interventions and policies as they enable comparatively cheap, convenient, and risk-free experiments that are impossible in reality [
5]. DAMs can offer an opportunity to perform “what-if” scenarios to predict and explore NP diabetes interventions in LMICs before these interventions are implemented, saving policymakers the cost and time that would have been invested in trial and error.
Despite the benefits of DAMs, their application to diabetes in LMICs is limited. Most DAMs for type 2 diabetes are built among high-income populations and then generalized to LMICs, which could be misleading considering that the difference in culture, ethnicity and health system capacities influence diabetes control. For instance, diabetes develops 10 years earlier and at lower body weight in Africans and Asians than in Europeans [
6]. There is a need to consolidate the literature on DAMs application, particularly their methodology, to diabetes in LMICs to identify gaps in their adoption and advance their use.
Whereas studies have appraised the application of DAMs to study diabetes interventions [
7‐
11], few have focused on NP diabetes interventions in LMICs. Our study adds to this research. Mukonda, Cleary and Lesosky’s review on computer simulation models for type 2 diabetes in LMICs, like our present study, reports on model-based economic evaluations to support decisions in type 2 diabetes care, to assess their quality and validity [
11]. However, unlike Mukonda and colleagues, we focus on DAMs for NP interventions and consider both type 1 and 2 diabetes. NP interventions are important for reducing diabetes risk and improving its outcomes [
3], providing benefits to both diabetic and non-diabetic populations. Our review focuses on NP interventions as they are understudied in the literature compared to pharmacological interventions [
11]. Moreover, the methods (and outcomes reported) used to investigate NP interventions can be very different from those used to examine pharmacological interventions as the latter concentrates on biology, physiology, and medicines compared to NP intervention, for which local context, including culture, health system and infrastructure is so important [
3,
11]. As a result, it is often more reasonable to transfer efficacy estimates of pharmaceutical interventions than NP ones from RCTs in one setting to another. Thus, the methods for conceptualizing and developing models and the types of models that are fit for purpose differ for NP interventions. DAMs are also employed for scaling up pharmacological interventions’ use to understand the effectiveness of distributing these interventions in a population. The modelling focus and assumptions differ compared to NP ones. By focusing on NP interventions, we maintain a more homogenous set of studies and gain insight into this understudied area.
Specifically, our review investigates the methodological application of DAMs investigating NP diabetes interventions in LMICs with three objectives:
1.
To investigate how DAMs have been applied to investigate NP policy interventions in LMICs,
2.
To assess what gaps exist in the modelling procedures according to the Consolidated Health Economic Evaluation Reporting Standards (CHEERS) checklist.
3.
and how to advance their adoption to diabetes research in LMICs.
We focus on four types of DAMs: systems dynamic, agent-based models, microsimulations, discrete event simulations and Markov models and a hybrid of any of them, hereafter referred to as DAMs, in NP diabetes intervention. Multiple factors determine the choice of DAMs in a study, e.g., study problem and goal, system processes, data availability and aggregation level, and analysis unit. Each approach has specific assumptions and structure, detailed in Additional file
1.
Discussion
The chronic nature of type 1 and 2 diabetes, high cost and implementation challenges of large-scale cohort studies necessitate the use of decision analytical modelling to extrapolate evidence of short-term empirical studies to predict costs and health benefits over a long-term period. This study reviewed the application of DAMs, specifically Markov cohort models, system dynamics, discrete event, microsimulation, and agent-based models, to study NP diabetes intervention in LMICs. We applied the CHEERS checklist to appraise seventeen studies identified through a systematic search.
Overall, there were more studies conducted in Asian sub-populations, particularly China, and most studies reported data as a limitation. The dearth of studies exclusively from LMICs is somewhat indicative of the scarcity of the use of DAMs in research in the region. Healthcare data sources exist in LMICs, e.g., national demographic and health surveys and WHO’s studies on Global Aging and Adult Health. However, the levels of detail and quality are often insufficient. Consequently, confidence in the validity of outcomes/results from models using such data sources could be questioned. This is daunting given the expected escalation of the diabetes burden, and it highlights the need for more structures for data collection/generation in LMICs.
In the absence of sufficient and quality data in LMICs, calibration techniques are used to adjust parameters or even existing models with different population characteristics, e.g., models from HICs. The objective of (re)calibration is to enhance the representativeness or predictability of a model which otherwise might under- or over-estimate model outputs. Calibration approaches reported in the reviewed studies include Markov Chain Monte Carlo —a standard Bayesian posterior computation approach [
33] and adjusting the Risk Equations for Complications of Type 2 Diabetes (RECODe) with local data. RECODe was designed for Americans in particular; however, the equation can be adjusted with local data to represent context-specific epidemiology.
The International Society for Pharmacoeconomics and Outcomes Research have also produced protocols for the adaption of an existing validated model to a specific context/ population [
34]. During such an adaption process, it is essential that baseline characteristics of the modelled cohort are adjusted to reflect local epidemiological data and possibly the transition probabilities between states. Also crucial for the adaption are context-specific cost, health state preferences, and utility data. The requirement of context-specific data brings us back to the challenge of reliable/robust diabetes data in LIMCs and how the situation limits the application of DAMs for diabetes in the region. Modellers could obtain model data from various sources: ongoing clinical trials or existing data through systematic review and meta-analysis, routine data collection, expert opinion, and observational studies. Among these data sources, researchers have suggested using observational research data on type 2 diabetes in LMICs despite concerns about selection bias as compared to RCTs, highlighting how such metrics, including health-related quality of life, are seldom altered by selection bias [
35]. As a result, observational studies, which are easier to conduct in LMICs, can fill in the data gaps required for the implementation of DAMs.
Model validation is another procedure to improve models’ fitness for purpose, especially when adapting parameters or existing models developed from other populations. Nearly half of the reviewed publications do not report how models were validated, and few discussed how applicable their model is in other LMICs. Some of the reviewed studies mentioned using data from the United Kingdom Prospective Diabetes Study (UKPDS) Outcomes Model and RECODe to estimate transition probabilities and risk equations for predicting diabetes complications. It is worth noting that both were developed for the United Kingdom and United States populations, respectively. Even though these models had undergone external validation, it is unclear if the models had also been validated for the LMIC populations of interest.
Tarride and colleagues [
36] have identified some drawbacks to using UKPDS data: 1) the majority of UKPDS participants were of European descent and considering that diabetes progresses differently in different ethnic groups, there is a challenge with generalizing data to varied ethnic groups. 2) Technology and treatment regimens have advanced since the study was conducted (1977–1997). Such developments, e.g., new medicines and disease incidence, could affect the transition between diabetes states, affecting the applicability of the model. 3) Differences in culture, lifestyle, demographics, health systems and available health technology means that the rate of diabetes progression in a United Kingdom population is different from LMIC populations, affecting the application of the model to LMICs. These issues highlight the need for extensive model validation and reporting on the validation process to increase model representativeness, transparency and confidence in model processes and outcomes, consequently increasing model adoption. Researchers could adopt different tools, e.g., CHEERs checklist, International Decision Support Initiative (iDSI) Reference Case for Economic Evaluation [
37], Diabetes Modelling Input Checklist [
38] and the Overview, Design concept and Details protocol [
39] (in ABMs) to improve model reporting and transparency.
Although most of the reviewed studies used a Markov cohort model or microsimulation (Table
2), it is unclear what modelling approach is appropriate as their appropriateness depends on multiple factors, including the study question/objective. The assumption of “memoryless” property, meaning that transition to another state is independent of the previous states, is a fundamental characteristic of Markov modelling but also a drawback [
5] as this would be a simplification for diabetes where progression to another diabetes state is dependent on the previous state. Given the substantial patient variability and interconnected risk in diabetes, complexity is almost unavoidable [
9]. Another challenge with using Markov cohort models in a cost-effectiveness analysis is the likelihood of skewed estimates of the incremental cost-effectiveness ratio resulting from “uncaptured” patient heterogeneity [
40].
Unlike Markov cohort models, microsimulation models individual characteristics through time, thereby representing patient variability and overcoming the Markov assumption. Individual level characteristics can affect how they move through the model and can also be utilized to modify the likelihood of future occurrences. Details of DAM approaches are available in Additional file
1. Microsimulation allows for modellers to assign risk of events or complications, taking into consideration a patient’s unique demographic characteristics, risk factors, and event history and is not constrained to “mutually exclusive” health states as is the case in Markov-based models. However, the challenge with microsimulation compared with a Markov cohort model is the heavy computation and data required [
9]. In addition to data challenges, LMICs currently lack the technical capacity for building heavy computational models. Collaborations between international and local academic and research institutions could help increase modelling expertise in LMICs. Knowledge exchange initiatives, such as the Mount Hood Diabetes Challenge Network [
38], which facilitates the sharing of ideas and skills among diabetes simulation modellers through workshops, hackathons, and conferences, can build capacity in LMICs for advancing DAMs.
Additionally, incorporating evidence of intervention cost and effect in DAMs can increase the use of such models as decision support in LMICs. Economic evaluation methods that use DAM estimate the potential lifetime impact of investments, which means policymakers can account for or justify their choices. Having such evidence can contribute to policymakers prioritizing diabetes interventions. The CHEERS protocol recommends reporting perspectives of economic evaluations and a lifetime horizon for tracking cost and effect. It is encouraging to note that all reviewed studies reported on the perspectives of their economic analysis, giving readers a context within which cost is tracked.
Study limitations and directions for future research
The study’s main limitation is that research that would have added to the discussion may have been left out due to exclusion criteria, such as eliminating published material that is not presented in English or studies on pharmacological interventions. We eliminated the latter due to their difference in modelling methodologies, assumptions and focus compared to NP intervention studies. By collecting data relevant to modelling, important insights into the effectiveness and cost-effectiveness of the interventions being simulated are not reported. Further research should thoroughly analyze DAMs in connection to the decision issue at hand and the selection of cost-effective NP interventions. Second, although we judge the CHEERS checklist appropriate for this review, the Diabetes Modelling Input Checklist designed for improving transparency in reporting input data of model-based economic evaluation and the IDSI Reference case are equally suitable and should be explored by future studies. Nevertheless, given the large overlap with both checklists, we believe nothing significant has been overlooked that may have altered the findings of this analysis.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.