A Critical Systematic Review of Modelling Approaches and Methodologies used in Hyperlipidaemia Economic Evaluations

Abstract

Background and Objectives

Health economic modelling integrates evidence from multiple sources and relies on transparency to support reimbursement decisions. Hyperlipidaemia is a major contributor to cardiovascular disease and is routinely evaluated within health technology assessment frameworks. This systematic review examines health economic models of hyperlipidaemia, evaluates the methodological approaches used in the model development and identifies opportunities to improve model quality and transparency.

Methods

A systematic literature search was conducted in MEDLINE and Embase between 1987 and 2025 to identify hyperlipidaemia health economic models. Screening, data extraction and quality assessment were performed manually, and artificial intelligence (AI) software was used for additional checking, this was followed by conflict resolution. Findings are presented through narrative synthesis. This study was registered in the International Prospective Register of Systematic Reviews (PROSPERO; CRD420251043922). The review assessed key aspects of model structure including model type, type of hyperlipidaemia, population, hyperlipidaemia-related events, model outcomes, time horizon, software used and discounting approach. We assessed methodological quality through the Philips checklist, with a focus on model structure, data usage, the way in which studies addressed uncertainty through sensitivity analysis and model validation.

Results

A total of 154 unique model-based economic evaluations were identified, comprising 132 Markov models, 9 microsimulation models and 1 discrete event simulation. Most economic evaluations explored general hypercholesterolaemia in 138 studies, followed by familial hypercholesterolaemia in 23 studies, while lipoprotein(a) was investigated in two studies. Primary prevention was examined in 89 models, secondary prevention in 50 models and a combination of both in 15 evaluations. Overall methodological quality was assessed as high for models’ structure and data usage; however, it was moderate for model consistency.

Conclusions

Hyperlipidaemia models generally had transparent assumptions and a justified structure. However, current models often lack systematic data-selection practices to identify the most appropriate evidence for evaluation. Extensive uncertainty analysis and model validation were frequently absent in the assessed models. To support decision-making, model results should be displayed in an open-source format with publicly available code. Furthermore, patient values are rarely incorporated in current modelling practices, representing missed opportunities for patient-centred care.

Supplementary Information

The online version contains supplementary material available at 10.1007/s40273-026-01631-2.

Key Points for Decision Makers

Health economic models are informed by evidence from multiple data sources. However, systematic selection of appropriate clinical and epidemiological evidence is not always applied. Robust model results rely on appropriately and accurately selected inputs relevant for the population the model is intended to serve.

Despite best-practice guidelines, hyperlipidaemia models are rarely validated or do not adequately assess uncertainty in key data assumptions, which may undermine model accuracy and influence decision-making.

Model development should involve collaboration between patients and researchers and prioritise open-source models with publicly available code.

Introduction

Hyperlipidaemia is a chronic and progressive condition characterised by elevated levels of lipid fractions such as low-density lipoprotein cholesterol (LDL-C), triglycerides and lipoprotein(a) [lp(a)], arising from a complex interaction of genetic predisposition, lifestyle factors and ageing [1]. The condition typically remains asymptomatic for prolonged periods, while cumulative lipid exposure contributes to the development of atherosclerotic plaque. Clinical manifestations often occur later in life through cardiovascular events such as myocardial infarction or stroke, which substantially affect survival, quality of life and healthcare use [2]. Both ischaemic heart disease (IHD) and stroke are leading causes of mortality globally [3]. Hyperlipidaemia is a major modifiable risk factor and one of the primary contributors to IHD and stroke-related mortality worldwide [3]. Elevated LDL-C is estimated to account for 40.5 deaths per 100,000 people globally [3]. The long latent phase of disease, together with the presence of multiple co-existing risk factors and treatment pathways, creates important challenges for modelling long-term disease trajectories and evaluating prevention strategies.

Pharmaceutical and lifestyle interventions are essential for the management of lipid abnormalities and are commonly evaluated through health economic modelling prior to reimbursement decision. Health economic modelling is a core component of the health technology assessment (HTA) process [4]. Clinical trials underpin estimates of intervention effectiveness; however, because preventive cardiovascular interventions produce benefits beyond observed trial follow-up, economic models must extrapolate treatment effects over longer time horizons. To support appropriate interpretation of results by decision-makers, transparency and accuracy of model inputs and assumptions are essential.

Hyperlipidaemia is a major modifiable risk factor for cardiovascular disease (CVD), with elevated LDL-C contributing substantially to premature morbidity, mortality and long-term healthcare costs despite the availability of effective therapies. Conditions such as familial hypercholesterolaemia and elevated lp(a) remain underdiagnosed and undertreated, reflecting persistent gaps in prevention and management [5]. Evaluating interventions in this disease area requires economic models capable of projecting lifetime cardiovascular risk, treatment effects and population health impacts under uncertainty. These characteristics make hyperlipidaemia an important context for examining methodological approaches used in model-based economic evaluations.

Several earlier systematic reviews [6–13] examined broader cardiovascular disease (CVD) or hypercholesterolaemia in general; summarised results rather than modelling methods, focused on single drugs, drug classes or screening interventions with restrictive outcomes; and often used short search windows. This systematic review evaluates the structure, methods and quality of health economic models for hyperlipidaemia, including uncertainty analysis and validation practices [14]. Finally, it identifies priorities for future model development and key areas for improvement in their reporting and transparency.

Methods

Registration and Guidance

The systematic review was conducted and reported in accordance with the Preferred Reporting Items for Systematic Review and Meta-analyses (PRISMA) 2020 statement [15]. An internal protocol was developed prior to commencement, guided by criteria described in the editorial prompt for a collection of systematic reviews of modelling approaches [14]. This review was registered with the International Prospective Register of Systematic Reviews (PROSPERO; registration number: CRD420251043922).

Information Sources and Search Strategy

A literature search was conducted on the 6 May 2025 in Ovid MEDLINE (R) and Ovid Embase and Embase Classic to identify relevant studies published between 31 August 1987 and 6 May 2025. The search timeline reflects the approval date of the first hydroxymethylglutaryl (HMG)–CoA reductase inhibitor, lovastatin, by the US Food and Drug Administration. The database selection in this review followed the search strategy used in a previously published systematic review of modelling methodology in diabetes [16]. The search was limited to studies published in English. A complete database search strategy is detailed in the electronic supplementary material (ESM; ESM Workbook Sheet 3: Search strategy). As the focus of this review was on the development of health economic models rather than cost-effectiveness outcomes, the review was structured in accordance with recommendations from the International Society for Pharmacoeconomics and Outcomes Research (ISPOR) good practices taskforce report [17]. To identify additional relevant literature, the reference lists were screened using self-developed software that extracts references from the PDF files into Microsoft Excel and applies the original search strategy. For unreadable PDF files, manual hand searching was performed.

Study Selection

The review included studies that performed a model-based economic evaluation of treatments for hyperlipidaemia with or without a screening component, in primary or secondary prevention settings. Studies in which the population consisted exclusively of patients with other comorbidities including diabetes, cancer or human immunodeficiency virus were excluded. Eligible interventions included any screening strategies, pharmacological or non-pharmacological treatments. Studies were included if they explicitly reported the use of a decision-analytic modelling approach relevant to hyperlipidaemia (e.g. Markov models, microsimulation, decision trees) or presented model-based outcomes such as costs, cost-effectiveness estimates or long-term projections of cardiovascular events. Detailed reporting of model inputs, validation methods or time horizon was not required for study inclusion to capture the breadth of modelling applications. No comparator was chosen for this review. The systematic review assessed key model characteristics, including model structure, data sources, hyperlipidaemia-related events, treatment effects, model outcomes, validation techniques, uncertainty handling, data extrapolation and reported limitations. Models were excluded if they were used exclusively for budget impact analysis to maintain alignment with a previous systematic review [16]. Economic evaluations that did not employ a modelling approach or only included a within trial analysis were excluded. Models used solely for disease risk prediction, without an economic evaluation component, were excluded. Studies that incorporated prediction models, such as the Coronary Heart Disease (CHD) Policy Model [18], within an economic evaluation framework were included [19–21], such prediction-only models were consulted solely for contextual understanding. Further details on the eligibility criteria are provided in the ESM (ESM Workbook sheet 4: Eligibility criteria).

All references identified from the database searches, conducted using a predefined search strategy, were downloaded and managed in EndNote 21 [22], where duplicates were removed. Titles and abstracts were independently screened against the eligibility criteria using Covidence [23] by K.V. and T.B.A. Discrepancies were resolved through discussion with a third reviewer (Z.A.). In addition, K.V. conducted artificial intelligence (AI)-assisted screening using ASReview [24, 25], which applies machine-learning algorithms. The AI tool was used as an additional reviewer to verify screening decisions and minimise potential bias arising from human judgement. Screening was stopped once the ASReview recommended threshold of 93 (10%) titles and abstracts were consecutively classified as irrelevant. Following consensus-based screening and rechecking of excluded records, the AI model was iteratively trained to confirm inclusion decisions. Titles and abstracts meeting the eligibility criteria proceeded to full-text screening. Full texts were reviewed by K.V., with a random 20% sample independently reviewed by T.B.A. and Z.A. for quality assurance. Any discrepancies were resolved through discussion with a reviewer (Z.A.), and eligible studies were progressed to data extraction.

Data Extraction

Data were extracted from the included studies using a data collection form adapted from a previously published systematic review of health economic models [16]. Responses were recorded using a self-developed Google Forms template. Initial data extraction was performed by K.V. followed by a secondary verification using large language-model-based (LLM) AI tools. A structured extraction template and the corresponding article PDFs were input into GPT-5 [26] and Claude2 (Sonnet 4.5) [27] to generate additional extraction outputs, with each AI model applied to 50% of the included studies. All extracted data were manually compared, and any discrepancies were resolved through a consensus process detailed in the ESM (ESM Workbook sheet 5: AI).

Quality Assessment

Methodological quality was assessed using the Philips checklist [28], which is specifically developed for the critical appraisal of decision-analytic models [17]. Alternative frameworks, including the Consolidated Health Economic Evaluation Reporting Standards (CHEERS) [29], Technical Verification (TECH-VER) [30] and the Assessment of the Validation Status of Health Economic Decision Models (AdViSHE) [31] were considered but were not selected as primary tools because they focus mainly on reporting quality or model validation rather than detailed evaluation of model structure and assumptions. This approach was also consistent with previous systematic reviews of economic modelling in chronic diseases such as diabetes [16]. Identified methodological limitations were interpreted in relation to their potential impact on model validity and decision-making in cardiovascular prevention. A separate risk-of-bias assessment was not undertaken, as the Philips checklist captures key domains addressed in validated risk of bias tools, including the Bias in Economic Evaluation (ECOBIAS) checklist [32]. Further details are provided in the ESM (ESM Workbook Sheet 8: Quality assessment form).

All articles were evaluated by K.V., and H.K. independently reviewed a random sample of 10 articles as a quality check. Results were compared and a consensus was reached where required with a third reviewer (Z.A.). Articles were subsequently reviewed using the same LLM AI approach applied during data extraction. Recommendations from the International Society for Pharmacoeconomics and Outcomes Research and the Society for Medical Decision Making (ISPOR-SMDM) on modelling good research practices, including model parameter estimation and uncertainty [33, 34], were used to guide answers to specific sections of the Philips checklist.

Checklist scoring followed a structured approach. A score of 1 was assigned when a question was answered completely, 0.75 when a question was answered with missing details, 0.25 when a question was not answered but justified and 0 when a question was not answered without any justification. For irrelevant questions, a “not applicable” score was assigned, and the question was excluded from the overall total score. A theoretical maximum score of 58 points out of 60 questions is possible, as two checklist items (17 and 58) are negatively worded. Studies were classified as high quality if ≥ 75 to 100% of the Philips checklist items were answered completely, medium quality for ≥ 50 to < 75% and low quality if < 50%. Percentages were calculated by summing total points for each evaluation and dividing by the relevant section total, (i.e., 23 for structure, 32 for data and 5 for consistency), excluding items that were not applicable. Mean scores were summarised by study and by year using linear regression. The impact of the Philips checklist over time was examined using interrupted time series analysis with study weighted least squares regression. All analyses and figures were generated using Microsoft Excel 2021 and Python 3.14

Data Synthesis

Models for hypercholesterolaemia, familial hypercholesterolaemia and lipoprotein(a) were synthesised together, as the differences between these conditions did not affect model structure, uncertainty analysis or validation methodology. Extracted data were synthesised using an established framework comprising; development of a theoretical model of how interventions work, for whom and why, preliminary synthesis of findings; exploration of relationships with the data; and assessments of the synthesis robustness [35].

Justification for the Use of Artificial Intelligence

AI was used in this systematic review to support, but not replace, manual processes. A protocol for the use of AI including prompts, auditing procedures and results can be found in the ESM (ESM Workbook sheet 5: AI). All screening and extraction steps were completed manually prior to AI use, and AI outputs were subsequently compared with human decisions. ASReview [24] was used as a third reviewer to verify the completeness of title and abstract screening and did not replace manual screening or consensus-based discussion. The use of LLM AI was informed by findings from a systematic review examining generative AI across the systematic review process [36]. The selection of Claude2 (Sonnet 4.5) [27] and GPT-5 [37] was supported by evidence indicating close alignment between these models and human judgement [36].

Results

Search Strategy

The systematic literature search identified 154 model-based economic evaluations published between 1987 and 2025 [19–21, 38–188]. The study selection process is summarised in the PRISMA flow diagram (Fig. 1). The initial database search identified 1568 records, including 518 from MEDLINE and 1050 from Embase. Citation searching uncovered an additional 5297 references. After removal of 404 duplicates, 1164 titles and abstracts were screened. Of these, 866 records were excluded, and 298 studies proceeded to full text review. Following full text screening, 135 studies met the eligibility criteria, while 163 were excluded. An additional 19 studies were identified through citation searching, resulting in a total of 154 included studies.

Fig. 1.

Preferred Reporting Items for Systematic Review and Meta-analyses (PRISMA) diagram

Model Structure

The main structural characteristics of included studies are summarised in Table 1, with detailed study-level information provided in the ESM (ESM Workbook sheet 7: Data extraction responses). Of the 154 included studies, most employed Markov modelling approaches (130 studies 84%) [19–21, 38, 41, 42, 44, 45, 47–49, 51–53, 55–64, 66–72, 74–89, 91–98, 100, 101, 103–108, 110, 112–114, 116–121, 123–146, 148–159, 161–168, 170–175, 177, 180–183, 185, 186], followed by microsimulation modelling in 10 (7%) studies [99, 109, 115, 160, 169, 178, 179, 184, 187, 188], and one model was a discrete event simulation [73]. Most evaluations focused on general hypercholesterolaemia (138 studies 90%) [19–21, 38–82, 84–105, 107–111, 113–124, 126–133, 136–144, 147–161, 163, 167–172, 174, 175, 177–184, 187, 188]. Familial hypercholesterolaemia was examined in 23 (15%) studies [83, 106, 112, 118, 125, 128, 134–136, 139, 140, 145, 146, 156, 162, 164–166, 176–178, 185, 186]. Lipoprotein(a) was modelled in two studies [173, 178].

Table 1.

Summary of most common characteristics identified in hyperlipidaemia economic models

Type of hyperlipidaemia	Number	Percentage	Models featured
General hypercholesterolemia	138	90%	[19–21, 38–82, 84–105, 107–111, 113–124, 126–133, 136–144, 147–161, 163, 167–172, 174, 175, 177–184, 187, 188]
Familial hypercholesterolemia	23	15%	[83, 106, 112, 118, 125, 128, 134–136, 139, 140, 145, 146, 156, 162, 164–166, 176–178, 185, 186]
Lipoprotein(a)	2	1%	[173, 178]

Cohort	Number	Percentage	Models featured
Primary prevention	89	58%	[19–21, 40, 43, 44, 48–52, 55–59, 68–70, 72, 74–76, 80–84, 87, 90–94, 96, 97, 100–103, 106–117, 119–122, 125, 130, 132, 134–137, 139, 141, 144–146, 152, 155, 160, 162–167, 169, 176, 178–188]
Secondary prevention	50	32%	[38, 39, 42, 45, 47, 53, 54, 61–64, 66, 67, 73, 77–79, 85, 86, 88, 89, 95, 104, 105, 123, 124, 126, 127, 129, 133, 138, 142, 143, 147–151, 153, 154, 157–159, 161, 170–175]
Primary and secondary prevention	15	10%	[41, 46, 60, 65, 71, 98, 99, 118, 128, 131, 140, 156, 166, 168, 177]

Simulation method	Number	Percentage	Models featured
Cohort	143	93%	[19–21, 38, 39, 41–45, 47–53, 55–72, 74–101, 103–114, 116–146, 148–159, 161–177, 180–186]
Patient level	8	5%	[54, 73, 115, 147, 160, 178, 179, 187, 188]
Unknown	3	2%	[40, 46, 102]

Type of model	Number	Percentage	Models featured
Markov	130	84%	[19–21, 38, 41, 42, 44, 45, 47–49, 51–53, 55–64, 66–72, 74–89, 91–98, 100, 101, 103–108, 110, 112–114, 116–121, 123–146, 148–159, 161–168, 170–175, 177, 180–183, 185, 186]
Microsimulation	10	7%	[99, 109, 115, 160, 169, 178, 179, 184, 187, 188]
Discrete event	1	1%	[73]

Outcome measure	Number	Percentage	Models featured
QALY	135	88%	[19–21, 44, 45, 47, 48, 51, 53, 55, 56, 58–63, 66–69, 71–88, 90–99, 101–159, 161–180, 182–188]
LYG	64	42%	[38, 39, 41–43, 46–48, 52, 53, 60, 62–64, 69, 71–75, 79, 81, 84, 86, 88, 89, 94, 95, 99, 104, 106, 108, 110, 111, 113, 118, 120–122, 128, 133, 135, 136, 144, 146, 148, 150, 152, 156, 157, 162–168, 173, 178, 179, 183, 184, 186, 187]
ICER (cost/LYG)	25	16%	[39, 41, 42, 46–48, 51–53, 63, 64, 71, 73, 74, 79, 89, 104, 106, 108, 110, 135, 144, 146, 165, 187]
ICER (cost/QALY)	117	76%	[20, 21, 45, 47–50, 53, 55, 56, 58–60, 62, 63, 67–70, 72–74, 76, 79–88, 91, 92, 94, 95, 97–99, 102–108, 110, 112–134, 137, 139–159, 161–165, 167–173, 175, 177–179, 181–188]

Hyperlipidaemia Related Event	Number	Percentage	Models featured
Non-fatal MI	115	75%	[19–21, 38, 39, 42–45, 48, 49, 51, 52, 55, 56, 58–60, 62–67, 69–73, 75–77, 79–91, 94–96, 98, 99, 104, 105, 107–117, 119–121, 123–126, 129, 132–134, 136–138, 140–145, 147–150, 153–161, 163, 164, 168, 170, 172–182, 184, 185, 187, 188]
Non-fatal stroke	100	65%	[21, 42, 44, 48, 49, 51, 56, 59, 62, 65–68, 71–73, 75, 77–82, 84, 86, 87, 89–91, 93, 95–100, 102, 104, 107–110, 112, 114–116, 118–134, 136–141, 143–145, 147, 149, 151, 153–161, 163, 164, 168–172, 174, 176–178, 180, 181, 184, 187, 188]
Fatal CHD	105	68%	[20, 21, 39, 41, 43, 44, 46–49, 51–53, 56–64, 66, 68, 71, 73–75, 78, 80–82, 85, 86, 88, 90, 92–96, 100–108, 116, 118, 121–128, 130–132, 134, 135, 137–140, 144–148, 150–157, 159–162, 164–173, 175, 177–180, 183, 186–188]
Fatal MI	29	19%	[38, 47–49, 55, 66, 70, 72, 84, 87, 98, 99, 109, 110, 113, 115, 117, 119, 133, 141, 143, 147, 149, 158, 174, 179, 181, 182, 184]
Fatal stroke	28	18%	[47–49, 66, 68, 72, 84, 87, 95, 98, 99, 109, 110, 115, 118, 119, 122, 128, 131–133, 141, 143, 149, 158, 174, 180, 184]

Time horizon	Number	Percentage	Models featured
10	17	11%	[19, 21, 38, 39, 42, 57, 61, 68, 82, 90, 92, 106, 107, 137, 149, 180, 183]
Lifetime	99	64%	[41, 43–49, 55, 58–60, 62, 63, 66, 67, 70–72, 75, 77–79, 83–88, 91, 93–95, 97, 99, 100, 102–104, 108–118, 120, 122–129, 132, 134–136, 139, 140, 143–148, 155–159, 161–167, 170, 171, 173–179, 181, 182, 184–188]

Discount rate	Number	Percentage	Models featured
3%	75	49%	[19–21, 41, 45, 47–49, 55, 57, 59, 62–64, 68, 71, 73, 75, 76, 80, 82, 84, 86–88, 93, 96, 97, 99–106, 108, 111, 112, 114, 115, 118–122, 124, 127–129, 136, 138, 139, 141–144, 147, 149, 151, 155, 157, 158, 161, 164, 168, 169, 174–177, 181, 183, 184, 187]
5%	41	24%	[38, 39, 42, 44, 46, 47, 52, 58, 61, 69, 70, 72, 74, 78, 79, 89, 91, 106, 108, 113, 116, 133, 146, 148, 150, 152–154, 156, 162, 166, 167, 170–174, 178, 182, 186, 188]

Software	Number	Percentage	Models featured
Microsoft Excel	50	29%	[19, 50, 54, 58, 59, 61, 66, 67, 70, 72, 74, 97, 100, 101, 104, 106–108, 112, 117, 118, 121, 124, 128, 132–134, 136–138, 140, 143, 144, 146–149, 156, 157, 161, 162, 164–168, 173, 174, 177, 186]
TreeAge	39	22%	[48, 52, 53, 55, 61, 62, 64, 68, 71, 75, 80–82, 87, 91, 93, 102, 103, 105, 109, 114, 116, 119, 123, 129–131, 141, 142, 145, 151, 153, 154, 158, 169, 170, 172, 182, 184]
Unknown	51	29%	[20, 21, 38–43, 47, 49, 51, 56, 57, 60, 63, 73, 76–79, 83–86, 88–90, 92, 94, 96, 98, 99, 110, 111, 113, 120, 122, 125, 126, 135, 139, 150, 152, 155, 159, 163, 171, 175, 176, 181, 185]

Sensitivity analysis	Number	Percentage	Models featured
Deterministic sensitivity analysis	137	89%	[19–21, 38–42, 44–55, 57–62, 64–69, 71–84, 86–96, 98, 99, 102–106, 108, 110–116, 118–120, 122–131, 133–139, 141–159, 161–179, 181, 182, 184–188]
Probabilistic sensitivity analysis	118	77%	[19–21, 40, 50, 51, 53–56, 59, 62, 63, 65–67, 69–84, 86–89, 91–95, 97–101, 103, 104, 106–108, 110–112, 114–116, 118–121, 123, 124, 126, 128–132, 134, 136–138, 141–146, 148–159, 161–172, 174–179, 181, 182, 184–188]
Non-parametric boot strapping	11	7%	[75, 77–79, 85, 86, 98, 147, 175, 183, 185]

QALY quality adjusted life year, LYG life years gained, ICER incremental cost effectiveness ratio, CHD coronary heart disease

Primary prevention strategies were evaluated in 89 (58%) studies [19–21, 40, 43, 44, 48–52, 55–59, 68–70, 72, 74–76, 80–84, 87, 90–94, 96, 97, 100–103, 106–117, 119–122, 125, 130, 132, 134–137, 139, 141, 144–146, 152, 155, 160, 162–165, 167, 169, 176, 178–188]. Screening-based interventions were evaluated in 21 studies (14% of all included studies) [48, 52, 80, 93, 106, 107, 112, 125, 134, 135, 146, 162, 164–166, 169, 176, 180, 181, 185, 186]. These screening strategies were typically linked to downstream pharmacological treatment decisions and therefore required modelling approaches capable of capturing both diagnostic pathways and long-term disease outcomes. Twelve of these screening-based studies (57%) [52, 80, 93, 106, 112, 125, 134, 162, 164, 176, 185, 186] explicitly incorporated decision tree or decision-analytic structures to model screening performance (e.g. detection rates, uptake, sensitivity and specificity) prior to transition into state-transition models for long-term cost-effectiveness analysis. In terms of interventions, 48 (31%) primary prevention studies explored pharmacological treatment [19–21, 40, 50, 55, 56, 58, 59, 68–70, 72, 74–76, 81–84, 87, 91, 92, 94, 96, 100, 103, 108, 111, 113, 114, 116, 120–122, 130, 132, 136, 139, 144, 145, 152, 155, 160, 163, 178, 181, 187] and risk score reclassification was modelled in 11 (7%) studies [19, 40, 44, 97, 115–117, 119, 167, 184, 188]. These studies used relatively simple modelling frameworks and did not apply more advanced decision-analytic techniques.

Secondary prevention interventions were examined in 50 (32%) studies [38, 39, 42, 45, 47, 53, 54, 61–64, 66, 67, 73, 77–79, 85, 86, 88, 89, 95, 104, 105, 123, 124, 126, 127, 129, 133, 138, 142, 143, 147–151, 153, 154, 157–159, 161, 170–175], while 15 (10%) studies modelled both primary and secondary prevention strategies [41, 46, 60, 65, 71, 98, 99, 118, 128, 131, 140, 156, 166, 168, 177].

Model time horizon and annual discount rate were the most variable structural elements across included studies. A lifetime time horizon was most common, reported in 99 (64%) studies [41, 43–49, 55, 58–60, 62, 63, 66, 67, 70–72, 75, 77–79, 83–88, 91, 93–95, 97, 99, 100, 102–104, 108–118, 120, 122–129, 132, 134–136, 139, 140, 143–148, 155–159, 161–167, 170, 171, 173–179, 181, 182, 184–188]. Shorter time horizons were used in a subset of evaluations, most commonly 10-year horizons, reflecting limitations of the underlying risk equations, and were reported in 17 (11%) studies [19, 21, 38, 39, 42, 57, 61, 68, 82, 90, 92, 106, 107, 137, 149, 180, 183]. The shortest extrapolation period was 1 year, observed in a single study [50].

Annual discount rates were applied to both costs and quality-adjusted life years (QALYs). A discount rate of 3% was most frequently used, reported in 75 (49%) studies [19–21, 41, 45, 47–49, 55, 57, 59, 62–64, 68, 71, 73, 75, 76, 80, 82, 84, 86–88, 93, 96, 97, 99–106, 108, 111, 112, 114, 115, 118–122, 124, 127–129, 136, 138, 139, 141–144, 147, 149, 151, 155, 157, 158, 161, 164, 168, 169, 174–177, 181, 183, 184, 187]. A discount rate of 5% was applied in 41 (27%) studies [38, 39, 42, 44, 46, 47, 52, 58, 61, 69, 70, 72, 74, 78, 79, 89, 91, 106, 108, 113, 116, 133, 146, 148, 150, 152–154, 156, 162, 166, 167, 170–174, 178, 182, 186, 188]. No discounting was reported in 12 (8%) studies [47, 50, 51, 54, 65, 77, 109, 111, 117, 140, 160, 180]. The highest discount rate identified was 7.2%, applied in one study based on local Iranian economic evaluation guidelines [111, 189].

The most frequently simulated events in hyperlipidaemia economic models were non-fatal myocardial infarction, included in 115 (75%) studies [19–21, 38, 39, 42–45, 48, 49, 51, 52, 55, 56, 58–60, 62–67, 69–73, 75–77, 79–91, 94–96, 98, 99, 104, 105, 107–117, 119–121, 123–126, 129, 132–134, 136–138, 140–145, 147–150, 153–161, 163, 164, 168, 170, 172–182, 184, 185, 187, 188], followed by non-fatal stroke in 100 (65%) studies [21, 42, 44, 48, 49, 51, 56, 59, 62, 65–68, 71–73, 75, 77–82, 84, 86, 87, 89–91, 93, 95–100, 102, 104, 107–110, 112, 114–116, 118–134, 136–141, 143–145, 147, 149, 151, 153–161, 163, 164, 168–172, 174, 176–178, 180, 181, 184, 187, 188].

Fatal cardiovascular outcomes were less commonly modelled. Fatal myocardial infarction was included in 29 (19%) studies [38, 47–49, 55, 66, 70, 72, 84, 87, 98, 99, 109, 110, 113, 115, 117, 119, 133, 141, 143, 147, 149, 158, 174, 179, 181, 182, 184], fatal stroke in 28 studies (18%) [47–49, 66, 68, 72, 84, 87, 95, 98, 99, 109, 110, 115, 118, 119, 122, 128, 131–133, 141, 143, 149, 158, 174, 180, 184] and fatal coronary heart disease in 105 (68%) studies [20, 21, 39, 41, 43, 44, 46–49, 51–53, 56–64, 66, 68, 71, 73–75, 78, 80–82, 85, 86, 88, 90, 92–96, 100–108, 116, 118, 121–128, 130–132, 134, 135, 137–140, 144–148, 150–157, 159–162, 164–173, 175, 177–180, 183, 186–188]. Of the 105 fatal coronary heart disease health states seven (5%) studies are inferred from the model’s structure to be events exclusively of fatal myocardial infarction [39, 60, 64, 85, 88, 103, 148, 150].

Additional outcomes were modelled less frequently. Diabetes was included in five (3%) studies [115, 119, 178, 187, 188], adverse events in 14 (9%) studies [20, 21, 59, 68, 76, 87, 92, 93, 107, 115, 119, 152, 153, 169], transient ischaemic attack in 14 (9%) studies [49, 56, 66, 67, 73, 77, 79, 86, 90, 118, 121, 125, 134, 137], congestive heart failure in 12 (8%) studies [42, 49, 56, 77, 86, 118, 121, 126–128, 145, 181] and coronary artery bypass grafting in five (3%) studies [49, 52, 84, 115, 119]. Post-event health states were incorporated in 12 (8%) studies typically lasting up to 2 years following an event to capture the acute burden of disease [41, 53, 55, 61, 71, 92, 103, 121, 144, 155, 161, 171].

Economic evaluations predominately reported economic outcomes in terms of QALYs gained, which were included in 135 (88%) studies [19–21, 44, 45, 47, 48, 51, 53, 55, 56, 58–63, 66–69, 71–88, 90–99, 101–159, 161–180, 182–188]. Life years gained (LYG) were reported in 64 (42%) studies [38, 39, 41–43, 46–48, 52, 53, 60, 62–64, 69, 71–75, 79, 81, 84, 86, 88, 89, 94, 95, 99, 104, 106, 108, 110, 111, 113, 118, 120–122, 128, 133, 135, 136, 144, 146, 148, 150, 152, 156, 157, 162–168, 173, 178, 179, 183, 184, 186, 187]. Less commonly reported outcomes included number needed to treat (NNT), which was presented in eight (5%) studies [19, 89, 117, 143, 152, 159, 168, 183], return on investment (ROI) in two studies [129, 132] and disability-adjusted life years (DALYs) in two studies [21, 100]. The cost/QALY was reported in 117 (76%) studies [20, 21, 45, 47–50, 53, 55, 56, 58–60, 62, 63, 67–70, 72–74, 76, 79–88, 91, 92, 94, 95, 97–99, 102–108, 110, 112–134, 137, 139–159, 161–165, 167–173, 175, 177–179, 181–188], and the cost/life year gained was reported in 25 (16%) studies [39, 41, 42, 46–48, 51–53, 63, 64, 71, 73, 74, 79, 89, 104, 106, 108, 110, 135, 144, 146, 165, 187].

Models were most commonly developed using Microsoft Excel, reported in 50 (29%) studies [19, 50, 54, 58, 59, 61, 66, 67, 70, 72, 74, 97, 100, 101, 104, 106–108, 112, 117, 118, 121, 124, 128, 132–134, 136–138, 140, 143, 144, 146–149, 156, 157, 161, 162, 164–168, 173, 174, 177, 186] followed by TreeAge software, used in 39 (22%) studies [48, 52, 53, 55, 61, 62, 64, 68, 71, 75, 80–82, 87, 91, 93, 102, 103, 105, 109, 114, 116, 119, 123, 129–131, 141, 142, 145, 151, 153, 154, 158, 169, 170, 172, 182, 184]. Advanced statistical software was used less frequently, including STATA in six studies [137, 178–180, 187, 188], R in six studies [103, 127, 160, 162, 180, 183] and SAS in four studies [45, 147, 157, 160]. In 51 (29%) studies, software used for model development was not reported [20, 21, 38–43, 47, 49, 51, 56, 57, 60, 63, 73, 76–79, 83–86, 88–90, 92, 94, 96, 98, 99, 110, 111, 113, 120, 122, 125, 126, 135, 139, 150, 152, 155, 159, 163, 171, 175, 176, 181, 185].

Uncertainty Analyses

Uncertainty was assessed using both the data extraction framework and the Philips’ checklists, with results summarised in Table 1 and Fig. 2. Methodological uncertainty was explored in 130 (84%) studies [19–21, 38–42, 44–50, 52–69, 71–88, 90–94, 96, 99, 102–105, 107, 110, 112–116, 118–121, 123–129, 131–139, 142–158, 162–175, 177–179, 181, 182, 184–188] and parameter uncertainty was analysed in 113 (73%) studies [19–21, 39, 48, 52, 53, 56, 57, 59, 61, 65–67, 69, 70, 72–88, 91–96, 98–101, 103, 104, 107, 108, 110–112, 114, 116, 118–120, 122–124, 126–132, 134, 136–139, 141–144, 146, 148–159, 161–179, 181, 182, 184–188], most commonly through variation of discount rates and model time horizons.

Fig. 2.

Philips checklist uncertainty reporting over time

Deterministic sensitivity analysis was the most frequently applied approach, reported in 137 (89%) studies [19–21, 38–42, 44–55, 57–62, 64–69, 71–84, 86–96, 98, 99, 102–106, 108, 110–116, 118–120, 122–131, 133–139, 141–159, 161–179, 181, 182, 184–188], while probabilistic sensitivity analysis was performed in 118 (77%) studies [19–21, 39, 40, 50, 53–56, 59, 62, 63, 65–67, 69–84, 86–89, 91–95, 97–101, 103, 104, 106–112, 114–116, 118–120, 122–124, 126–132, 134, 136–138, 141–146, 148–159, 161–172, 174–179, 181, 182, 184–188]. The boot-strapping method was used in 11 (7%) studies [75, 77–79, 85, 86, 98, 147, 175, 183, 185].

When deterministic analyses were undertaken, 62 (40%) of these studies [19, 21, 38, 45, 48, 52, 53, 59, 73, 74, 76–78, 82–84, 86, 87, 91, 93, 96, 100, 103, 105, 110, 112, 118, 119, 128, 134, 136, 137, 142–144, 147–150, 153, 154, 156, 158, 159, 162, 163, 166, 168–174, 176–179, 182, 185, 186, 188] clearly stated and justified the parameter ranges. Among studies performing probabilistic sensitivity analysis, 71 (46%) studies reported justified distributional assumptions [19, 20, 39, 50, 53, 55, 59, 63, 65, 67, 69–71, 73–79, 81–84, 86–88, 93, 95, 97–100, 103, 109, 110, 112, 115, 116, 118, 119, 123, 126–129, 134, 136, 141, 143, 144, 146, 148–150, 153, 154, 156, 158, 161, 162, 166, 169, 171, 178, 179, 181, 182, 184–186]. Gamma and uniform distributions were most used for cost parameters, while beta distributions were applied to utilities and transition probabilities.

Risk Extrapolation and Structural Uncertainty

The risk of hyperlipidaemia-related events was explicitly modelled in 140 (91%) studies [19–21, 38, 39, 41–52, 54–69, 71–118, 120–132, 134–147, 149–153, 155–159, 161, 162, 164, 166–170, 173–179, 181–185, 187, 188]. Risk estimates were most commonly derived from individual literature sources specific to the target population, reported in 57 (37%) studies [39, 45–49, 52, 54, 55, 61–63, 65, 73, 75, 77–79, 81, 84, 86, 87, 91, 95, 97, 103–106, 109–111, 122, 123, 126, 127, 129–132, 135, 140–142, 146, 147, 152, 155, 156, 158, 162, 166, 168, 175, 177, 181, 184]. Evidence specifically from clinical-trials-informed risk of disease in 58 (38%) studies [45–50, 54, 55, 61, 62, 65–67, 69, 73, 75, 78, 83, 84, 86, 87, 97, 103–105, 108–111, 118, 122–124, 126–131, 134, 140, 145, 147, 150, 151, 153, 155–159, 161, 168, 170, 177, 181, 182, 185].

Regression-based cardiovascular risk equations derived from observational cohort data were used in 55 (36%) studies to predict future cardiovascular risk (typically over 10 years) in economic modelling [19–21, 38–44, 51, 56–58, 60, 63, 64, 68, 69, 71, 72, 74, 76, 79–82, 85, 88, 90, 93, 94, 96, 98–102, 107, 112, 113, 115, 117, 120, 121, 125, 136, 137, 139, 142, 144, 164, 169, 176, 183]. The Framingham risk equations [190] were used in 37 (24%) studies [19, 20, 38, 41–44, 51, 56–58, 60, 64, 68, 71, 76, 80, 82, 85, 88, 90, 93, 94, 96, 98–101, 112, 113, 115, 120, 121, 144, 164, 169, 176]. Multiple versions of the Framingham equations were applied across studies, with early applications using data from 1988 [38, 190] and more recent applications using data from 2009 [176, 191]. Alternative risk equations were used less frequently including, the Pooled Cohort Equation (PCE) [192] used in one study [183], the Systematic Coronary Risk Evaluation model (SCORE) [193] in four studies [98, 101, 107, 136] and Wu et al. [194] in two studies [72, 74]. A summary of risk equation sources is provided in Table 2.

Table 2.

Source of risk equations in hyperlipidaemia economic models

Source of Risk Equations	Number	Percentage	Models Featured
Literature studies	57	37%	[39, 45–49, 52, 54, 55, 61–63, 65, 73, 75, 77–79, 81, 84, 86, 87, 91, 95, 97, 103–106, 109–111, 122, 123, 126, 127, 129–132, 135, 140–142, 146, 147, 152, 155, 156, 158, 162, 166, 168, 175, 177, 181, 184]
Framingham	37	24%	[19, 20, 38, 41–44, 51, 56–58, 60, 64, 68, 71, 76, 80, 82, 85, 88, 90, 93, 94, 96, 98–101, 112, 113, 115, 120, 121, 144, 164, 169, 176]
CTTC meta-analysis	4	3%	[66, 67, 108, 128]
Other Risk equations	10	6%	[21, 72, 74, 102, 117, 125, 136, 137, 139, 183]
Mendelian Randomization	6	4%	[167, 173, 178, 179, 187, 188]
Other meta-analysis	6	4%	[89, 92, 116, 138, 149, 174]
CTTC and Literature	12	8%	[118, 124, 134, 145, 150, 151, 153, 157, 159, 161, 170, 185]
Literature studies and meta-analysis	6	4%	[50, 59, 83, 114, 143, 182]
Risk equation and Literature	1	1%	[69]
Meta-analysis and Risk equations	1	1%	[107]
Unknown	14	9%	[40, 53, 70, 119, 133, 148, 154, 160, 163, 165, 171, 172, 180, 186]

CTTC Cholesterol Treatment Trialists’ Collaboration

The clinical factors used in Framingham [190] were age, total cholesterol, high-density lipoprotein cholesterol, systolic blood pressure, smoking status and diabetes. SCORE-2 [193] differentiates further modelling sex and the PCE [192] models treatment for systolic blood pressure and race. Modelling of more than five risk factors was reported in 36 (23%) studies [19–21, 51, 58, 60, 63, 64, 69, 72, 74, 76, 80–82, 85, 88, 93, 94, 98–102, 107, 112, 113, 115, 117, 121, 136, 137, 139, 144, 164, 169]. Furthermore, Framingham and the PCE equations are derived through Cox proportional hazards models, while SCORE2 was derived using Fine and Gray competing risk-adjusted models. These risk equations are commonly calibrated through Harrell’s C index or Chi-squared tests as a measure of goodness of model fit within a broader population and was reported as reasonable in SCORE-2 and very good in the PCE.

Clinical trial data were commonly extrapolated to a lifetime horizon for modelling in economic evaluations in 46 (30%) studies [45–49, 55, 62, 66, 67, 75, 78, 83, 84, 86, 87, 97, 103–105, 108–111, 118, 122–124, 126–129, 134, 140, 145, 147, 155–159, 161, 170, 177, 181, 182, 185]. However, risk estimates from regression equations were extrapolated to lifetime horizons in 27 (18%) studies [41, 43, 44, 58, 60, 63, 71, 72, 79, 85, 88, 93, 94, 99, 100, 102, 112, 113, 115, 117, 120, 125, 136, 139, 144, 164, 176].

Risk of event occurrence was based on LDL-C levels derived at a single time point in 149 (97%) studies [19–21, 38–166, 168–177, 180–186]. However, the cumulative burden of cholesterol and the subsequent risk of a first cardiovascular event was examined in seven (5%) studies [146, 165, 167, 178, 179, 187, 188].

Studies that used risk equations were able to quantify risk of recurrent events through applying risk multipliers in nine (6%) studies [21, 56, 58, 71, 93, 98, 107, 117, 136], incorporating tunnel states to represent event history in ten (7%) studies [19, 51, 69, 80, 85, 99, 100, 113, 120, 121] or through microsimulation approaches to track individual disease trajectories in one study [115]. A change in event risk after a first event was not quantified in 19 (12%) studies [20, 38–40, 42, 43, 57, 60, 64, 68, 72, 74, 88, 90, 94, 96, 137, 142, 144, 183]. For studies that used clinical trials for risk estimations, 27 (18%) studies applied risk multipliers after an incident event [45, 47, 48, 55, 73, 78, 86, 87, 97, 103, 104, 108, 110, 111, 118, 123, 124, 126, 128, 134, 150, 159, 168, 177, 181, 182, 185], 4 (3%) studies incorporated tunnel states to represent event history [66, 67, 69, 75] and 21 (14%) studies did not account for a change in risk [46, 49, 50, 54, 61, 62, 65, 83, 84, 109, 122, 130, 131, 134, 145, 155–158, 161, 170].

Structural uncertainty was explored in 27 (18%) studies [20, 21, 72, 77, 83, 84, 91, 93, 110, 120, 121, 124, 127, 128, 133, 144, 146–148, 152–154, 158, 162, 166, 184, 187]. To assess the impact of alternative structural assumptions or survival modelling approaches, 18 (12%) studies conducted sensitivity analyses [20, 21, 72, 77, 83, 84, 91, 93, 110, 120, 121, 127, 133, 144, 147, 153, 162, 184], while 9 (6%) studies performed scenario analyses [124, 128, 146, 148, 152, 154, 158, 166, 187].

Quality Assessment

Quality assessment results are summarised in Figs. 3, 4, 5, 6, 7, with study-level responses provided in the ESM (ESM Workbook sheet 9: Quality assessment responses). Using the Philips checklist, section-specific and overall scores are shown in Fig. 3, with each point representing an individual evaluation. Overall checklist scores increased over time, averaging 37 points in early studies to 44 points in recent evaluations. Overall and section-level scores, presented in Fig. 4, indicate increased adherence to best-practice modelling principles [28].

Fig. 3.

Philips checklist scores over time

Fig. 4.

Philips checklist scores based on section averages

Fig. 5.

Philips checklist average domain scores over time

Fig. 6.

Philips checklist overall score interrupted time series

Fig. 7.

Philips checklist responses by question

Overall, 117 (76%) studies were rated with a high score [21, 45, 48, 50, 53–56, 58, 59, 61, 66, 67, 69, 71–89, 91–101, 103, 104, 106, 108–116, 118–130, 133, 134, 136–139, 142–144, 146–150, 152–159, 161–167, 169–179, 181–188], 33 (21%) studies were rated with a medium score [19, 38–40, 42–44, 46, 47, 49, 51, 52, 54, 57, 60, 62–65, 68, 70, 90, 102, 105, 107, 131, 132, 135, 140, 141, 145, 151, 168], and four (3%) studies were rated with a low score [41, 117, 160, 180].

In the structural domain, 141 (92%) studies achieved a high score [19–21, 40, 43–48, 50–59, 61–64, 66–89, 91–104, 106–116, 118–146, 148–159, 161–179, 181–188], while 13 (8%) studies received a medium score [38, 39, 41, 42, 49, 60, 65, 90, 105, 117, 147, 160, 180].

For the data domain, 86 (56%) studies were rated with a high score [20, 21, 53, 55, 58, 61, 66, 67, 69, 72–77, 79, 80, 83, 84, 86–89, 91, 93, 95–97, 99, 101, 103, 104, 106, 107, 110, 111, 113–116, 118, 119, 121–123, 127–130, 133, 134, 137, 138, 142–144, 146–150, 152–158, 162, 163, 165–167, 170–172, 174, 175, 177–179, 181, 182, 184, 187, 188], 61 (40%) studies were rated with a medium score [19, 38–40, 42–48, 50, 52, 54, 56, 57, 59, 60, 62–65, 68, 70, 71, 78, 81, 82, 85, 90, 92, 94, 98, 100, 102, 105, 108, 109, 112, 120, 124–126, 131, 132, 135, 136, 139, 141, 145, 151, 159, 161, 164, 168, 169, 173, 176, 183, 185, 186], and seven studies (5%) were rated with a low score [41, 49, 51, 117, 140, 160, 180].

For the consistency section, 29 (19%) studies were rated with a high score [19–21, 55, 58, 73, 76, 80, 84, 91, 100, 110, 115, 120, 134, 137, 139, 143–145, 147, 148, 158, 167, 170, 178, 181, 184, 187], 79 (51%) studies were rated with a medium score [40, 42, 45, 47–49, 52–54, 59–61, 64, 69, 77–79, 81–83, 85–87, 89, 92, 93, 96–99, 103, 105, 106, 109, 111, 113, 114, 117, 119, 122, 125, 127–129, 131–133, 135, 136, 138, 141, 142, 146, 149, 151, 153–157, 159–161, 163, 164, 168, 169, 171–177, 179, 180, 182, 183, 185], and 46 (30%) studies were rated with a low score [38, 39, 41, 43, 44, 46, 50, 51, 56, 57, 62, 63, 65–68, 70–72, 74, 75, 88, 90, 94, 95, 101, 102, 104, 107, 108, 112, 116, 118, 121, 123, 124, 126, 130, 140, 150, 152, 162, 165, 166, 186, 188].

No study achieved a perfect score of 58 points. Seven studies achieved the highest overall scores, exceeding 90% [137, 146, 148, 153, 158, 166, 187]. Trends in mean annual scores, with 95% confidence intervals, are shown in Fig. 5.

The introduction of the Philips checklist in 2006 coincided with higher overall model quality scores. Interrupted time series analysis (Fig. 6) showed an average increase of approximately 2 points in overall scores after 2006. Scores remained relatively stable in subsequent years. The number of low scoring outliers (scores ≤ 30) decreased from four studies in the initial 18 evaluations [41, 42, 49, 51] to six studies in the 136 evaluations published after 2006 [60, 90, 105, 117, 160, 180] Further details are provided in the ESM (ESM Workbook sheet 10: Interrupted time series).

Several Philips checklist items were frequently inadequately addressed, particularly those related to data selection, uncertainty analysis and validation Fig. 7. As the Philips checklist is a general assessment tool, some items were commonly rated as not applicable. In particular, the item assessing continuing treatment effects after treatment cessation was consistently marked irrelevant, reflecting the lifelong nature of hyperlipidaemia treatment. Comparable items are not included in other checklists such as CHEERS [29], TECH-VER [30] and AdViSHE [31], indicating potential scope for adaptation of the Philips checklist for chronic conditions.

Systematic data selection methods were adequately reported in 29 (19%) studies [20, 43, 55, 58, 66, 67, 72, 89, 91, 95, 107, 111, 115, 116, 125, 129, 130, 134, 138, 143, 144, 148, 155, 165, 166, 172, 175, 187, 188]. Structural uncertainty was analysed in 27 (18%) studies [20, 21, 72, 77, 83, 84, 91, 93, 110, 120, 121, 124, 127, 128, 133, 144, 146–148, 152–154, 158, 162, 166, 184, 187], and heterogeneity was explored in 47 (31%) studies [20, 40, 45, 46, 54, 60, 61, 63–65, 67, 69, 78, 80, 83, 96, 102, 103, 109, 113, 114, 122, 128, 130, 133, 139, 145–148, 150, 153, 154, 157–159, 162, 163, 166, 167, 169–171, 177, 179, 181, 186]. Comprehensive uncertainty assessment across stochastic, structural, parameter and heterogeneity domains was observed in 12 (8%) studies, all published after 2010 [20, 83, 128, 146–148, 153, 154, 158, 162, 166, 181].

Model Validation

Model validation encompassing internal, external or cross-validation was reported in 122 (79%) studies [19–21, 40, 42, 45, 47–49, 51–53, 55, 57–61, 64, 66, 67, 69, 72, 73, 76–89, 91–93, 96–101, 103, 105–107, 109–111, 113–117, 119–139, 141–148, 151–165, 167–178, 180–184, 186, 187]. External validation was undertaken in 22 (14%) studies [19, 20, 51, 55, 58, 73, 76, 80, 84, 91, 115, 120, 123, 139, 145, 148, 158, 162, 167, 178, 181, 184] comparing model outputs with independent real-world evidence (RWE) or clinical trial data not used in model development. Cross-validation was reported in 117 (76%) studies [19–21, 40, 42, 45, 47–49, 52, 53, 55, 57–61, 64, 66, 67, 69, 72, 76–89, 91–93, 96–101, 103, 105–107, 109–111, 113–117, 119–122, 124–139, 141–148, 151–161, 163–165, 167–177, 180–184, 186, 187], in which results were compared with outputs from other models. Internal validation was reported in 19 (12%) studies [20, 21, 73, 80, 100, 110, 115, 134, 137, 143, 144, 148, 149, 158, 170, 178, 179, 184, 187], where model logic or code was independently checked. Two evaluations publicly released model code enabling full model reconstruction [178, 187].

Checking with Artificial Intelligence

After the ASReview stopping threshold of 10% was reached, 241 articles proceeded to full text screening. During data extraction and quality assessment, LLM outputs most closely aligned with human assessments for questions related to model structure. In contrast, responses concerning model validation, uncertainty analysis and risk extrapolation frequently differed from human-valued answers, even when prompts explicitly referenced best-practice guidelines [33, 34]. Compared with human assessments, Claude 2 generated 1520 differing responses and GPT-5 generated 1667 differing responses. Differences in study inclusion following title and abstract screening as well as discrepancies in quality assessment responses, are summarised in the ESM (ESM Workbook Sheet 5: AI).

Discussion

Findings and Best Practice

Health economic models of hyperlipidaemia published since 1987 have applied a range of disease simulation approaches informed by economic evaluation methods and biological knowledge. Representing clinical complexity in hyperlipidaemia modelling requires consideration of heterogeneous risk profiles, age-dependent disease progression, recurrent cardiovascular events, competing mortality risks and time-varying treatment effects. While cohort-based Markov models remain widely used due to their transparency and relative simplicity, they often rely on simplifying assumptions such as population homogeneity and constant transition probabilities. These assumptions may limit their ability to fully reflect cumulative exposure to lipid levels and dynamic changes in cardiovascular risk over the life course.

More flexible modelling approaches, including individual-level microsimulation and multi-state modelling frameworks, allow explicit representation of patient heterogeneity, evolving comorbidity patterns and recurrent events [99, 109, 115, 160, 169, 178, 179, 184, 187, 188]. Such approaches may therefore provide a more clinically realistic depiction of disease progression and treatment pathways. However, their application is frequently constrained by data limitations, computational requirements and uncertainty in long-term risk estimation. Strengthening methodological approaches to better align model structures with underlying disease processes remains an important area for future research.

Robust model structure is reliant on clearly defined well-justified assumptions regarding time horizon, discounting, input parameters, scope, cycle length, interventions and causal relationships between disease states. Quality assessment using the Philips checklist indicated that, while assumptions were generally reported in a transparent way, justification and comprehensive sensitivity analyses were frequently limited.

Economic evaluations of interventions for hyperlipidaemia frequently combine screening and treatment components within a single modelling framework. Screening strategies typically require decision-tree models to represent patient identification pathways and risk reclassification, whereas pharmacological interventions are commonly evaluated using long-term state-transition models to capture disease progression and recurrent cardiovascular events. Failure to distinguish these modelling requirements may obscure important methodological differences. For example, Marquina et al. [162] demonstrated the need for additional decision-analytic components by modelling detection rates, test uptake, specificity and sensitivity of genetic screening before subsequent evaluation in a Markov framework. Future studies should explicitly justify modelling choices on the basis of intervention mechanisms and relevant time horizons.

Modelling hyperlipidaemia presents important methodological challenges, particularly in representing lifetime exposure to lipid risk factors, treatment persistence and changes in cardiovascular risk following incident events. Across included studies, important differences were observed between models informed by clinical trial event rates (59 studies) and those based on epidemiological risk equations (55 studies). While trial-based models often rely on shorter time horizons with fewer assumptions regarding risk translation, epidemiological models typically require additional assumptions to project long-term risk. In 55 studies, risk estimates were derived from epidemiological risk equations (e.g. Framingham, SCORE2 and related equations [190, 192, 193]) and translated within modelling frameworks by transforming predicted multi-year risk estimates into cycle-specific transition probabilities. Typically, modelled populations are first characterised using baseline risk factor profiles, after which predicted 5 or 10-year cardiovascular risk is calculated. These risk estimates are then converted into annual probabilities, enabling incorporation into Markov or microsimulation structures. While this approach is widely used (55 studies in this review), the methodological assumptions underlying this translation, including constant risk, independence of competing risks and the trajectory of risk factor effects over time, are often not explicitly addressed in economic models.

Cardiovascular risk equations differ substantially in their derivation populations, included risk factors, calibration methods and outcome definitions and may therefore have limited applicability for evaluating contemporary prevention strategies or emerging lipid-lowering therapies.

While Framingham [190] risk equations include six risk factors, SCORE2 [193] incorporates seven, and the PCE [192] includes nine. More recent cardiovascular risk equations not applied in the models identified in this review, such as the Predicting Risk of Cardiovascular Disease Events (PREVENT) [195] equation, incorporate up to 14 risk factors. Similarly, the New Zealand PREDICT-1° [196] equation, and by extension AusCVDRisk [197], include up to 16 risk factors. Substantial variation in the number and type of risk factors included across cardiovascular risk equations may contribute to differences in predicted risk and modelled cost-effectiveness results. More rigorous appraisal of evidence inputs and transparent justification of modelling choices are therefore essential to improve the credibility and policy relevance of hyperlipidaemia economic evaluations.

Approaches to modelling recurrent events, comorbidities and competing mortality varied substantially within models using epidemiological risk equations or clinical trial estimates, with 36 (23%) studies applying risk multipliers, 14 (9%) studies using tunnel states and one study using individual-level microsimulation to better reflect heterogeneous disease trajectories. Overall, many models rely on simplified representations of risk that do not fully capture the cumulative biological impact of prolonged exposure to lipid risk factors [178], which may affect the structural validity of long-term projections. Greater methodological development is therefore needed to better represent lifetime cardiovascular risk dynamics in prevention modelling [146, 165, 167, 178, 179, 187, 188].

Recent modelling approaches using microsimulation have integrated cumulative cholesterol exposure to determine lifetime cardiovascular risk [178, 179, 187, 188], representing a distinct departure from traditional single time point LDL-C measurements. These approaches quantify “cholesterol years” enabling more patient-specific risk assessment and informing timing of preventative interventions [198]. For example, Ademi et al. emphasise the importance of early treatment in childhood for individuals with heterozygous familial hypercholesterolaemia [146, 165], whereas Morton et al. [178] quantify the impact of delayed initiation by comparing interventions starting at ages 40 and 60 years, with transparent reporting of model structure, event trajectories and economic outcomes.

The credibility of hyperlipidaemia economic models depends critically on the quality, relevance and time horizon of their input evidence. Many studies rely on short-term clinical trial data or cardiovascular risk equations derived from observational cohorts that estimate medium-term (5–10-year) risk [205]. While informative, these evidence sources are not designed to capture cumulative exposure to lipid risk factors, long-term treatment adherence or lifetime disease progression, and therefore require extrapolation beyond observed data. Several challenges identified in this review reflect limitations in the type of underlying evidence (e.g. short-term clinical trials or risk estimates derived from observational studies), rather than limitations in modelling methodology alone.

The introduction of the Philips checklist [28] in 2006 coincided with higher quality scores in subsequently published models. A review by Philips et al. [199] describes ongoing improvements in modelling practices in the year preceding the formal introduction of the Philips checklist. Several additional checklists have since been developed, including CHEERS [29], TECH-VER [30] and AdViSHE [31]. However, these tools differ in scope and depth from the Philips checklist. After 2006, quality scores appeared to stabilise, with minimal sustained increase observed. Systematic approaches to data selection with justification were not consistently embedded in published hyperlipidaemia economic models. Existing methodological guidance emphasises the importance of using systematic literature review to inform model structure and parameterisation, particularly for costs, utilities and transition probabilities [34, 200–203]. Transparent reporting of data sources, and selection rationale remains critical to ensure model credibility and relevance for decision-making [200, 202]. Ribeiro et al. [116] exemplify best practice by deriving treatment effects via systematic review, meta-analysis and Bayesian mixed-treatment comparison with transparent parameter selection [204].

Many models adhere to established best-practice recommendations by clearly defining uncertainty, selecting parameter estimates from evidence-based sources and applying standard statistical approaches for point and interval estimation [34]. However, relatively few economic evaluations explicitly assessed structural uncertainty [20, 21, 72, 77, 83, 84, 91, 93, 110, 120, 121, 124, 127, 128, 133, 144, 146–148, 152–154, 158, 162, 166, 184, 187] and heterogeneity [20, 40, 45, 46, 54, 60, 61, 63–65, 67, 69, 78, 80, 83, 96, 102, 103, 109, 113, 114, 122, 128, 130, 133, 139, 145–148, 150, 153, 154, 157–159, 162, 163, 166, 167, 169–171, 177, 179, 181, 186]. Structural uncertainty and heterogeneity were infrequently assessed across health economic models, reflecting the additional analytical requirements associated with re-specifying model structure and conducting subgroup analyses [34, 205]. In contrast, Marquina et al. [166] demonstrate this by explicitly addressing parameter and structural uncertainty and heterogeneity through systematic variation of key assumptions and age-stratified subgroup analyses, with transparent reporting of statistical methods where published estimates were unavailable.

Model validation was limited across most economic evaluations. Internal validation of models was explicitly reported in only a small subset of studies [20, 21, 73, 80, 100, 110, 115, 134, 137, 143, 144, 148, 149, 158, 170, 178, 179, 184, 187]. While models are used in reimbursement decision-making, independent scrutiny of model validation remains uncommon [200]. Clear specification of model structure and validation procedures should therefore be considered a minimum standard, regardless of whether full code is publicly available [200, 206, 207]. An open-source approach, as demonstrated by Morton et al. [178, 187] enables reconstruction of the model and represents a high standard of transparency, accountability and transferability [34]. No other included evaluations adopted a comparable level of openness.

Cross-validation was the most commonly reported validation approach across economic evaluations [19–21, 40, 42, 45, 47–49, 52, 53, 55, 57–61, 64, 66, 67, 69, 72, 76–89, 91–93, 96–101, 103, 105–107, 109–111, 113–117, 119–122, 124–139, 141–148, 151–161, 163–165, 167–177, 180–184, 186, 187]. However, in most studies, cross-validation was limited to comparisons of final cost-effectiveness outcomes, rather than interrogation of underlying model structure and assumptions. Meaningful cross-validation requires transparent reporting of model logic and implementation, which remains uncommon, particularly where models are proprietary [34, 202]. Xi et al. [171] provide a robust example by explicitly comparing their evaluation with that of Liang et al. [154] detailing differences in structural design, the modelling of LDL-C–outcome relationships and the resulting implications for risks, utilities, costs and long-term outcomes.

External validation using clinical trial data or RWE was reported in few economic evaluations [19, 20, 51, 55, 58, 73, 76, 80, 84, 91, 115, 120, 123, 139, 145, 148, 158, 162, 167, 178, 181, 184]. Although models are often structured to mirror trial-based outcomes, heterogeneity in trial design, populations and endpoints limits the feasibility of direct validation [34]. As a minimum standard, economic models should benchmark outcomes against independent data sources within a comparable scope and clearly account for discrepancies between observed and modelled results [34, 200]. Pandya et al. [115] provide a robust example, calibrating model inputs using the Framingham Offspring and atherosclerosis risk in communities (ARIC) study [208–210] and subsequently validating outputs against (NHANES III) [211] mortality data. Importantly, differences between modelled and observed outcomes were explicitly reported both before and after calibration, demonstrating a high level of transparency. Consistent with best-practice guidelines, modelling studies should apply systematic data selection processes to support transparent comparison between modelled outputs and external evidence, documenting data sources, calibration or simulation approaches, discrepancies and their implications for results [33, 34, 200–202]. Where such validation is not feasible, limitations should be clearly acknowledged alongside opportunities for further refinement [34].

Limitations identified in this review arise not only from modelling approaches but also from the nature of the underlying clinical and epidemiological evidence. Short trial follow-up, reliance on risk prediction equations and the need to extrapolate long-term outcomes introduce uncertainty into estimates of cost-effectiveness. In addition, substantial variability exists in how models represent patient characteristics, risk factor progression and disease trajectories. While stronger clinical and epidemiological evidence can improve the representation of clinical heterogeneity, rigorous adherence to best-practice principles in model structure, evidence selection, validation, uncertainty analysis and transparent reporting remains essential to ensure credible and policy-relevant findings [33, 34, 199, 200].

Recommendations for Model Improvement

Several opportunities exist to strengthen economic modelling practice in hyperlipidaemia. Despite availability of best-practice guidance to conduct economic modelling, overall model quality has plateaued since 2006, underscoring the need for renewed, coordinated efforts across researchers, journals and decision-makers. One potential approach is the development of a collaborative modelling network, analogous to the Mount Hood Diabetes Challenge Network [212] or the Cancer Intervention and Surveillance Modeling Network (CISNET) [213], in which models are iteratively refined, compared and validated in an open and structured environment. Such collaboration could be supported through dedicated workshops, conference-based workshops, or moderated forums linking published models to facilitate transparent comparison of assumptions, structure and results. Importantly, this framework would enable broader engagement beyond researchers alone.

Second, emerging methodological developments in HTA, including the use of RWE and advanced evidence synthesis approaches, may help address key limitations in hyperlipidaemia modelling and expand intervention applicability [5, 214, 215]. RWE from registries, electronic health records and biobanks can inform long-term disease burden, treatment adherence, safety outcomes and population heterogeneity where randomised trial evidence is limited. Such approaches are particularly relevant for evaluating preventive strategies and novel therapies requiring lifetime projections. However, challenges related to bias, confounding, transportability and data governance continue to influence the extent to which RWE is incorporated in HTA decision-making [223]. Causal inference approaches such as Mendelian randomisation may provide complementary estimates of cumulative lifetime effects of risk factor modification, particularly in CVD and cancer [146, 165, 167, 178, 179, 187, 188, 216]. These developments highlight the need for continued methodological innovation in prevention modelling.

Models identified in this review primarily evaluated outcomes using standard economic metrics such as QALYs, life-years gained and incremental cost-effectiveness ratios. QALYs are typically derived from patient-reported preference-based instruments, most commonly the European Quality of Life Five-Dimension, Five-Level Version (EQ-5D), although other generic measures [e.g. Short Form Six Dimensions (SF-6D)] or mapping approaches may also be used [217]. However, utility data are not consistently collected within lipid-lowering clinical trials. Consequently, many modelling studies rely on utility estimates obtained from external sources, including observational cohort studies or prior economic evaluations in comparable populations [218]. While this approach is methodologically accepted, it introduces uncertainty related to transferability across populations, treatment contexts and disease severity. Furthermore, most models use average health-state utility values, which may not adequately reflect patient-level consequences of prevention strategies, including treatment burden, adherence challenges, disutility related to screening or long-term medication use and variation in patient preferences. These limitations suggest that future modelling work could benefit from improved integration of patient-reported data collected prospectively within trials and registries, as well as methodological advances to better represent patient-centred outcomes within long-term decision-analytic frameworks. An example of this approach is provided by the FH Europe Foundation and the Lipoprotein(a) International Task Force [219]. Embedding such participatory approaches within collaborative modelling initiatives may improve model relevance, transparency and credibility for health technology assessment and reimbursement decision-making.

Fourth, economic models should be routinely self-assessed against validated checklists. While Philips was applied in this systematic review, several other tools, including the CHEERS [29], TECH-VER [30] and AdViSHE [31] are also available, and any departures from checklist recommendations should be explicitly justified. The outcomes of such assessments should be made publicly available, either within publications or as supplementary material, to support transparency and comparability across models. While many journals require adherence to reporting standards such as CHEERS [220–222], this review highlights the need for more targeted methodological guidance for disease-specific modelling contexts. Development of consensus recommendations, like ISPOR-SMDM task force reports, could support improved transparency, structural validity and consistency in model-based economic evaluations of cardiovascular prevention.

Finally, the Philips checklist [28] highlights three core principles; essential for high methodological quality: identification, sourcing and validation of assumptions, which underpin high-quality economic evaluation. Embedding these principles more explicitly within checklist items would support systematic documentation of assumptions, evidence sources and validation steps. These practices align with the requirements of health technology assessment processes globally, including the Pharmaceutical Benefits Advisory Committee (PBAC) in Australia [223, 224], the National Institute for Health and Care Excellence (NICE) in the UK [225] and the European reimbursement framework [226].

Limitations

This systematic review has several limitations. First, inclusion was restricted to English-language journal articles; economic models reported only in conference abstracts or non-English publications were excluded. To mitigate this, a broad search window spanning 2025–1987 was applied. MEDLINE and Embase were searched to maintain methodological consistency with a prior systematic review [16]. Other databases (e.g. PubMed, Scopus, Google Scholar, NHS-EED, EconLit) were not searched, which may have excluded potentially relevant studies. Studies that reported the use of a modelling approach were included; however, it may have introduced selection bias by excluding studies with limited methodological reporting. Data extraction and quality assessment using the Philips checklist involve elements of subjective judgment, and some checklist items may not fully capture the nuances of all modelling approaches. Finally, although the Philips checklist [28] was selected, relevant elements from other reporting and validation frameworks (i.e. CHEERS, TECH-VER, AdViSHE) [29–31] were not explicitly evaluated.

Conclusions

Economic models for hyperlipidaemia draw on evidence from clinical trials, RWE and well-established biological principles. However, across some evaluations, practices pertaining to data selection, uncertainty analysis and model validation are not aligned with best-practice guidelines from ISPOR and HTA bodies. Although improving transparency in reporting model assumptions is essential for interpretation and reproducibility, transparency alone does not resolve structural challenges in representing disease complexity. Future modelling studies should prioritise methodological approaches that better capture cumulative risk exposure, heterogeneous treatment pathways and long-term disease progression, supported by improved longitudinal casual evidence.

Input from patients and representative groups on clinically relevant model design and analyses can support reimbursement and policy decisions which continue to rely predominantly on aggregate outcomes. To support high-quality reimbursement decisions in resource-constrained settings, economic models should adapt transparent and well-justified structures and be informed by systematic selection of high-quality data. All assumptions and inputs should be identified, sourced and validated, with model code and documentation made openly available wherever possible. Decision-makers should prioritise models that rigorously address uncertainty and demonstrate validity using data relevant to the populations they are intended to inform.

Supplementary Information

Below is the link to the electronic supplementary material.

Funding

Open Access funding enabled and organized by CAUL and its Member Institutions. This study was funded by the National Health and Medical Research Council Ideas Grants Application ID: 2012582

Declarations

Conflict of Interest

Authors have no conflicts of interest to disclose.

Availability of data and material

All data are included in this article and its electronic supplementary materials and in the relevant references. Any additional data are available on request from authors.

Code availability

Not applicable.

Author contributions

K.V.: conceptualisation and design, literature search and selection, quality assessment of included studies, data extraction, statistical analysis and interpretation of data, drafted and review of the manuscript. H.K.: quality assessment of included studies, statistical analysis and interpretation of data and review of the manuscript. T.B.A.: literature search and selection, review and revision of the manuscript. S.R.: data interpretation and review of the manuscript. C.B.: interpretation of data and review of the manuscript. Z.A.: conceptualisation and design, literature search and selection, statistical analysis, interpretation of data and writing and review the manuscript and supervision.

Ethics approval

The study did not require ethics approval.

Consent to participate

Not applicable.

Consent for publication

Not applicable.