Intervention targets for reducing mortality between mid-adolescence and mid-adulthood: a protocol for a machine-learning facilitated systematic umbrella review

Abstract

Introduction

A rise in premature mortality—defined here as death during the most productive years of life, between adolescence and middle adulthood (15–60 years)—is contributing to stalling life expectancy in high-income countries. Causes of mortality vary, but often include substance misuse, suicide, unintentional injury and non-communicable disease. The development of evidence-informed policy frameworks to guide new approaches to prevention require knowledge of early targets for intervention, and interactions between higher level drivers. Here, we aim to: (1) identify systematic reviews with or without meta-analyses focused on intervention targets for premature mortality (in which intervention targets are causes of mortality that can, at least hypothetically, be modified to reduce risk); (2) evaluate the review quality and risk of bias; (3) compare and evaluate each review’s, and their relevant primary studies, findings to identify existing evidence gaps.

Methods and analysis

In May 2023, we searched electronic databases (MEDLINE, PubMed, Embase, Cochrane Library) for peer-reviewed papers published in the English language in the 12 years from 2012 to 2023 that examined intervention targets for mortality. Screening will narrow these papers to focus on systematic reviews with or without meta-analyses, and their primary papers. Our outcome is death between ages 15 and 60 years; with potential intervention targets measured prior to death. A MeaSurement Tool to Assess systematic Reviews (AMSTAR 2) will be used to assess quality and risk of bias within included systematic reviews. Results will be synthesised narratively due to anticipated heterogeneity between reviews and between primary studies contained within included reviews.

Ethics and dissemination

This review will synthesise findings from published systematic reviews and meta-analyses, and their primary reviewed studies, meaning ethics committee approval is not required. Our findings will inform cross-cohort consortium development, be published in a peer-reviewed journal, and be presented at national and international conferences.

PROSPERO registration number

CRD42022355861.

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• This review will be reported in line with best practice guidelines (Preferred Reporting Items for Systematic Reviews and Meta-Analyses statement) and is registered with PROSPERO.

• The comprehensive search strategy was designed a priori with a medical librarian and the review will be supplemented by machine learning to expedite screening and extraction across multiple reviewers.

• Only studies published in the English language and published in the 12 years from 2012 to 2023 will be included in this review.

• Data synthesis will be narrative, rather than quantitative.

Introduction

Longevity is a key indicator of the health of populations, but life expectancy is beginning to decline in many settings, suggesting reductions in overall population health. In 2019, across high-income countries (HICs), the Global Burden of Disease (GBD) study estimated life expectancy at 81.3 years (95% uncertainty interval (UI) 81.2 to 81.4), with healthy life expectancy 67.4 years (95% UI 64.2 to 70.4).¹

While this life expectancy in HICs remains higher than life expectancy across low/middle-income countries (LMICs),¹ and represents a commendable advance over the last century,^{2 3} since the mid-2000s gains in lifespan have stalled in HICs, for both males and females.^{1 4} For example, the annual rate of increase in life expectancy at birth among males in HICs has reduced from an average of 0.29 years in 2001–2006 to 0.21 years in 2011–2016, and from 0.21 years to 0.12 years among females during the same period⁴; with the stagnation greater in countries such as USA,⁵ England and Wales (eg, +0.02 to +0.07 years in 2011–2016).⁴ Moreover, Ho and Hendi’s recent demographic analyses showed that life expectancy gains have not only stalled, but that male life expectancy between 0 and 65 years reversed in nine of the 18 HICs examined (and in four countries for women) during 2014 and 2015.⁵ These declines in life expectancy among men 0–65 years were most pronounced in the USA, Italy, UK, France and Australia.⁵ Such reductions can signal the deterioration of a population’s health profile, which may be driven by structural determinants (eg, inequality, geographic location/deprivation) and behavioural and environmental factors arising from worsening socioeconomic and social circumstances.

Stalling life expectancy is partly attributable to the rise in mortality across the productive years of late adolescence and early-to-middle adulthood.^{1 4–8} For example, in the USA between 2010 and 2021 all-cause age-specific mortality rates among children and adolescents aged 1–19 years rose from 25.82/100 000 to 29.93/100 000 (increasing from 32.97/100 000 to 38.65/100 000 in males 1–19 years).⁹ Across a similar time period between 2010 and 2019 all-cause mortality rates in the USA among men aged 35–44 years rose from 212/100 000 to 257/100 000, and in women from 128/100 000 to 143/100 000.¹⁰ Across all HICs, the GBD study has estimated that the probability of dying from any cause between ages 15 and 60 years is unacceptably high, especially for men, at 10% (14% in USA, 8% in UK, Australia, New Zealand and Canada).¹ The GBD probability estimates for dying between years 15 and 60 are also unacceptably high in LMICs. For example, 17% for men in Latin America and the Caribbean, and 14%–19% for men and women in South Asia.¹ These higher estimates in LMICs may reflect differences in the drivers and causes of death (eg, higher rates of communicable disease deaths) between HICs and LMICs.^{11 12}

Between HICs, there is variation in drivers of increasing mortality rates among adolescents and young adults.^{5 11} Much of the literature has focused on patterns in the USA compared with other HICs. Results suggest that declines in life expectancy in the USA have been driven by deaths in young and middle-aged adults from external causes (eg, opioid overdose).⁵ By contrast, in other HICs, reduced life expectancy has been attributed to deaths among those aged over 65 years from causes including rising incidence of suicide, diseases of ageing (eg, Alzheimer’s disease and other nervous system diseases) and non-communicable diseases (NCDs; eg, respiratory and cardiovascular diseases, diabetes).^{5 7 11} NCD mortality between ages 15 and 74 years and/or years life lost to NCDs has been frequently linked with obesity, hypertension and/or tobacco smoking.^{11 13–15}

Because of these emergent patterns, rising death rates throughout the USA have been referred to as the epidemic of despair^{8 16 17} with deaths among young and middle-aged adults in the USA commonly arising from alcohol misuse, suicides, violence, drug overdoses and unintentional injury (eg, drowning, falls, road crashes).^{8 18} While it has been suggested that deaths of despair are unique to the USA,^{8 18 19} there is growing recognition that these patterns of deaths (eg, substance use deaths) are now rising throughout other HICs, although at a slower pace.^{10 11 20} This emergent epidemic is primarily occurring within young working-age adolescents and young to middle-aged adults in the general, rather than clinical, population, who are participating in health-compromising behaviours potentially caused by exposure to early-life and/or present deteriorating social and economic conditions (eg, austerity).17 19 21–23

Internationally, literature on mortality and premature mortality in the general population has grown extensively over the past decade, and subsequently has been synthesised in various systematic reviews and meta-analyses. These reviews are variable in their focus covering specific intervention targets (ie, causes of mortality that are hypothetically modifiable), all-cause mortality or multiple intervention targets for specific causes of death (eg, drug overdose, suicide).^24–28 Commonly studied intervention targets for cause-specific death in the young-to-middle aged adult population include obesity, antisocial behaviour, substance use, psychological factors, personality, stressful life events/conflict, housing and exposure to violence.

To progress developmentally informed policy, we align this review with a life course framework, acknowledging that life course patterns/trajectories towards early mortality are shaped by experiences from birth, and emerge within the context of multiple, interconnected systems over time.^29–31 Evidence shows that interventions have maximum effect when they are implemented early in the life course and are multifaceted and target all bioecological levels of development, including the microsystem (eg, depressive symptoms, abuse), mesosystem and exosystem (eg, neighbourhood deprivation, housing stability, employment) and the highest level macrosystem (eg, education systems, economic policy and systems, geographic location).^{32 33} As such, in this review, we aim to identify evidence gaps for intervention targets in each of these levels/systems at a population-level across the life course. Following the principle of equifinality,³⁴ we expect that some drivers across these levels will be common across multiple mortality endpoints, and thereby potentially represent particularly efficient opportunities for preventive intervention.

To build an integrated life course framework that can be used to guide investment in the prevention of mortality during the adolescent and early adulthood years, there exists an unmet need to summarise recent systematic reviews and meta-analyses examining intervention targets (including structural and social determinants) for all-cause, and cause-specific, mortality between adolescence and middle adulthood. In line with GBD calculated prevalence rates and past research,^{1 35 36} and congruent with our review objectives, we define mortality during adolescence and early/middle adulthood as living to the age of 15 but dying before the age of 61. While death at 60 years is earlier in the life course than WHO, Organisation for Economic Co-operation and Development and Eurostat definitions of premature mortality,^37–39 deaths after 60 years are more likely attributable to diseases of ageing, which are outside the focus of this review.

Our resulting review has two overarching objectives to: (1) comprehensively synthesise the strength of evidence on intervention targets for reducing mortality between adolescence and middle adulthood in the general population, and (2) use findings from objective 1 to inform the development of a consortium which will harmonise data across existing life course studies with a view to identifying population-level intervention targets.

Accordingly, this review aims to: (1) identify and synthesise published systematic reviews (with or without meta-analyses) focused on intervention targets for mortality between adolescence and middle adulthood; (2) evaluate the review quality and risk of bias; and (3) summarise, compare and evaluate the review, and their relevant primary study, results to identify evidence gaps and to build an integrative evidence base to inform intervention targets to prevent death before middle adulthood.

Methods and analysis

We report this protocol in accordance with the Preferred Reporting Items for Systematic Review and Meta-Analysis Protocols statement.⁴⁰ Any changes to the protocol will be documented via PROSPERO, should they arise. The review will be conducted by combining traditional review methods (ie, bibliographic literature searches) with a screening process expedited by a machine learning algorithm (see the Selection of review studies section).

Search methods for identification of studies

The search strategy has been developed in consultation with a university research librarian at The Royal Children’s Hospital in Melbourne, Australia. Using variants and combinations of search terms for intervention targets (ie, hypothetically modifiable causes, risk factors or antecedents defined below) for mortality between ages 15 and 60 years, in May 2023 we searched three health, medical and biomedical databases of peer-reviewed research: MEDLINE (Ovid), Embase and PubMed. We supplemented these searches by searching for published Cochrane reviews via The Cochrane Library. The MEDLINE (Ovid) search is presented in table 1. The search strategy included a combination of subject headings (MeSH terms), thesaurus terms and keywords. It included various examples of known intervention targets (eg, socioeconomic factors, depression), but also generic terms (eg, risk or risks) to capture lesser-known factors. We made small modifications to the MEDLINE (Ovid) strategy for the other databases (see online supplemental file). To avoid missing review papers, the electronic search strategies are not limited to reviews and meta-analyses, which will only be selected once a thorough electronic literature search has been executed. As well as reviews/meta-analyses identified via database searches, to identify additional review references we will: (1) search the reference lists of identified reviews, (2) search publications lists of identified authors and (3) seek key references from the last 12 years by emailing international experts in the field.

Table 1.

MEDLINE search strategy

1	(adolescen* or teen* or adult* or man or men or woman or women or middle-aged or youth or youths).af.
2	(risk or risks).tw,kf,hw.
3	td.fs.
4	2 or 3
5	(premature-death* or premature-mortalit* or premature-casualt* or early-death* or early-mortalit* or early-casualt* or adult*-mortalit*-risk or adolescen*-mortalit*-risk or teen*-mortalit*-risk or youth-mortalit*-risk).tw,kf.
6	mortality, premature/
7	5 or 6
8	“Cause of Death”/
9	Mental Health/
10	Student Dropouts/
11	exp Socioeconomic Factors/
12	exp body fat distribution/ or body mass index/ or body size/ or body weight/ or exp overweight/ or waist circumference/ or skinfold thickness/ or waist-hip ratio/
13	exp Substance-Related Disorders/
14	exp Alcohol Drinking/
15	accidents/ or accidents, home/ or accidents, occupational/ or accidents, traffic/ or drowning/
16	exp Mental Disorders/
17	affective symptoms/ or depression/ or self-injurious behavior/ or suicide/ or suicide, completed/
18	Cholesterol/
19	Hyperglycemia/
20	exp Smoking/
21	exp Poisoning/
22	Anxiety/
23	blood pressure/ or heart rate/ or respiratory rate/
24	exp Heart Diseases/mo [Mortality)
25	adverse childhood experiences/ or exp bullying/ or exp crime/ or exp dangerous behavior/ or divorce/ or exp homicide/ or juvenile delinquency/ or exp parental death/ or exp violence/
26	Prisons/
27	(smoking or obesity or blood-pressure or substance-abuse or alcohol or drug or crime* or criminal* or prison* or incarcerat* or mental-health or mental-illness* or unwanted-pregnan* or unplanned-pregnan* or socioeconomic* or socio-economic* or income or violence or abuse* or relationship* or separat* or divorc* or ((singl* or solo) and parent*) or school-dropout* or school-leaver* or ((internal* or external*) and (problem* or disorder*)) or bullying or leave-home or left-home or homeless* or ((parent* or mother* or father*) and (death or mortality)) or adverse-childhood or adversity or health-behavior* or health-behaviour* or unhealthy-life* or unhealthy-behavior* or unhealthy-behaviour*).tw,kf.
28	8 or 9 or 10 or 11 or 12 or 13 or 14 or 15 or 16 or 17 or 18 or 19 or 20 or 21 or 22 or 23 or 24 or 25 or 26 or 27
29	1 and 4 and 7 and 28
30	limit 29 to (english language and yr=“2012 -Current”)
31	limit 30 to (case reports or comment or editorial or guideline or letter or practice guideline)
32	30 not 31

Eligibility criteria

As shown in table 2, the eligibility criteria for inclusion are based on the PICO framework (Population or Problem, Intervention or Exposure, Comparison or Control, Outcome(s)).⁴¹

Table 2.

Summary of review inclusion and exclusion criteria

	Inclusion criteria	Exclusion criteria
Study or publication type	Peer-reviewed systematic reviews of primarily quantitative data with or without meta-analysis (or similar statistics).40 44 Reviewed literature identified using a systematic structured search of bibliographic databases. Ideally, the review has been conducted and reported according to Preferred Reporting Items for Systematic Reviews and Meta-Analyses54 and/or the methods used are explicit, prespecified, transparent and reproducible.40	Reviews excluded if they: (i) are not peer-reviewed, or (ii) do not detail systematic and prespecified protocols, or (iii) are commentaries rather than critiques.44 Case reports, comments, editorials, guidelines, letters or practice guidelines.
Population	Individuals included in the studies within the reviews are primarily involved in a population-representative or community cohort (note: if the review is not primarily focused on free-living community or population cohort studies, then the community/population cohort studies must at least be largely presented/reviewed independently from other study types also reviewed). Cohort is longitudinal, observational, prospective or retrospective (eg, data linkage). Not limited to high-income countries.	Individuals included in studies in the reviews/meta-analyses are primarily involved in clinical cohorts, cross-sectional, case–control studies, randomised controlled trials, or in population/community cohorts that have been selected on pre-existing criteria (eg, disease free at time of recruitment).
Intervention/ exposures	Examination of potential intervention targets for preventing early death, where intervention target has been measured prior to the outcome. Further detail in ‘Intervention targets/exposures’ section.	Intervention targets measured prior to birth or concurrently/at the same time as the outcome. Reviews that focus on predictors (ie, not suited to intervention) of mortality, rather than intervention targets (ie, causes). Or where the focus is unclear.
Comparator	Not limited by comparator populations.	NA
Outcome	Review synthesises primary papers examining any cause of mortality during or after age 15 years (even if some, but not all, mortality exceeds 60 years). Following which, regardless of publication date, each relevant primary paper within each included review will be sourced, and also included in primary paper data synthesis if participants are aged between ≥15 years and ≤60 years. Further detail in ‘Outcomes’ section.	Review synthesises primary papers examining mortality measured only before age 15, or only after age 60. Regardless of publication date, each relevant primary paper within each included review will be sourced and the primary paper (not the review) will be excluded from primary paper data synthesis if mortality is only measured before age 15 or after age 60.
Publication date	Because intervention targets for preventing death vary with changing social, political and economic conditions, to ensure reviewed targets are relevant today we will only include reviews published during/after 2012.	Published before 2012.

Included population

To reduce the chances of reverse causation and selection bias, we limited included reviews to those of primarily prospective cohort studies. Because our review focuses on early-life targets of mortality, we excluded reviews primarily based on randomised controlled trials as they rarely are prospective and rarely is the outcome death. We excluded clinical cohorts to ensure that our study findings are generalisable to our population of interest (ie, general population).

Intervention targets/exposures

By definition, we consider intervention targets as causes of mortality that can, at least hypothetically, be modified to reduce risk of death between mid-adolescence and mid-adulthood. Because it is essential to develop prevention strategies that are multisectoral (eg, education, housing, welfare) to achieve population-level impact on the complex issue of early mortality, intervention targets will not be limited to those within the health system.

Guided by the bioecological model of development and social determinants of health framework,^{33 42 43} intervention targets will be grouped into three levels. We aim to identify and discuss gaps in the literature and identify potential intervention targets that have not yet been the subject of research inquiry. Exposure needs to have occurred during infancy, childhood, adolescence, or adulthood and precede mortality. Examples include, but are not limited to:

Community-level intervention targets such as neighbourhood deprivation, access to health services, community connectedness, education systems, geographic location, area-level disadvantage, economic policy and systems (eg, systems of resource or service allocation and distribution).

Family-level intervention targets such as family socioeconomic status, income inequality, housing instability/homelessness, relationship stability.

Individual-level intervention targets such as self-harm, risk-taking, physical activity, accidents, injuries, illnesses, school/education engagement, high-school completion, employment, absenteeism, workplace/occupational engagement.

Outcome

It is likely some systematic reviews will include a wide age range of deaths and/or only summarise age-standardised death outcomes. As such, we will conduct ‘review paper data extraction and synthesis’ and also source each primary research paper within each included review. Following, if the paper meets the inclusion criteria we will conduct ‘primary paper data extraction and synthesis’. This approach is congruent with current recommendations for conducting overviews of reviews.⁴⁴

Selection of review studies

Publication details for the retrieved studies will be imported into EndNote reference management software and duplicates will be removed. Screening will be facilitated by artificial intelligence (AI) supported software developed through Deakin University—referred to as the Living Knowledge System (LKS).⁴⁵ Using AI to support the systematic review process is gaining in popularity because scientific literature continues to increase in volume, meaning that executing high quality and complex scientific reviews can be time intensive, resource heavy and prone to human error.⁴⁶ The LKS has been specifically designed to semi-automate the review process by expediting screening and extraction across multiple reviewers. The LKS is a web-based platform that uses machine learning to expedite the screening of papers for inclusion in systematic reviews. By reducing the repetitiveness and time taken to screen papers, human fatigue and error should reduce.⁴⁶ In turn, the accuracy of the screening and the corresponding review results should increase. As a machine learning algorithm, the LKS learns by example and makes predictions on novel data.⁴⁷ To begin the machine learning processing within LKS, from the total database search, the user begins screening with traditional methods to identify a small set of highly relevant papers (eg, 20–50 primary research papers examining early-life intervention targets for premature death). The user then uploads this small set to LKS and tags these as relevant. LKS then uses item text from title and abstract to create word embeddings, train, and ultimately reorder the larger pool of items so that the most relevant (vs irrelevant) items are prioritised, enabling the user/s to more efficiently complete stage 1 of the screening process. That is, once exposed to examples of highly relevant and irrelevant publications, the LKS will predict the relevance of subsequent unscreened items and prioritise these items for the user to screen ahead of less relevant items. This machine learning process and item re-ordering is continuous—as the user continues to screen retrieved studies, the LKS model continues to become more precise in its predictions and re-orders all remaining items after the user screens every 15 items. Therefore, for every 15 items screened by the user, the model re-ranks relevant publications and presents them for user/s screening.

The LKS provides a recommended point for the user/s to stop screening. By default, the LKS prompt to stop screening appears once the user has selected 40 consecutive items to be irrelevant/ineligible and when the remaining unscreened items are predicted by the LKS algorithm to have less than 50% chance of being relevant (ie, <50% level of confidence). Previous reviews supplemented by machine learning have used a similar approach to ending screening.^{48 49} This LKS stopping criterion can be altered by users. For example, a more conservative approach would involve using a larger number of consecutive irrelevant items and/or a lower confidence level: a change we will make.

For the current review, reviewer ANG will use our inclusion and exclusion criteria to identify highly relevant and irrelevant primary papers to train the machine learning algorithm ready for stage 1 screening (this is stage 0, figure 1). As previously described, the LKS algorithm will continuously assess all items based on title/abstract text, and then rank these items for users to screen; this process is continuous, and items will be re-ranked for the users after every 15 items screened. Two users/reviewers (JAK and ANG) will then apply the inclusion criteria independently to screen a proportion of publication titles and abstracts (stage 1) to determine eligibility for full-text review (stage 2). At the time of screening, reviewers will be blinded to each other’s screening. During stage 1 and stage 2, the LKS will track reviewers’ conflicting votes related to study inclusion and notify reviewers of disagreements that need resolving. These will be discussed by the two reviewers to reach consensus, with a third team member (MO) consulted if an initial agreement cannot be reached. During stage 1, publications will be reviewed until the LKS identifies that the two reviewers have specified 40 consecutive items as irrelevant and that the remaining unscreened items are less than 20% likely to be relevant. We have set this conservative stopping criterion (<20%) within the LKS to override the default setting of <50% relevant.

Figure 1.

Selection of studies. AI, artificial intelligence; LKS, Living Knowledge System.

After stage 1 screening is complete, we will review the eligibility criteria to use during stage 2 with all team members. These discussions will ensure that the inclusion and exclusion criteria remain relevant to the reviews that have been identified in stage 1. Once eligibility criteria are re-finalised, stage 2 screening will begin by obtaining all identified full-text reviews and uploading to LKS. JAK and a second reviewer will then independently screen 100% of these full-text items for review inclusion. Conflicts about review inclusion during stage 2 will be discussed between the two reviewers, and by a third team member (MO) when necessary. To measure the degree of accuracy and reliability between the screening conducted by JAK and the second reviewer, we will calculate and report Cohen’s kappa coefficient.⁵⁰

Selection of primary studies from included review studies

As noted in table 2, each primary paper within each included review will be sourced, and if the participants are aged between ≥15 years and ≤60 years, we will complete a subset of data extraction from this primary source.

Data extraction and management: review studies

Once review articles are finalised for inclusion, two reviewers (JAK and a second reviewer) will examine each paper to extract relevant data in the LKS. Discrepancies between reviewers will be discussed until consensus is met with a third reviewer (MO) if necessary. JAK will contact the corresponding review author twice via email if important data are incomplete or missing, if there is no reply from the author then the study will be excluded from review.

Following guidance for overview umbrella reviews and the AMSTAR 2 (A MeaSurement Tool to Assess systematic Reviews) checklist,^{44 51} the following details will be extracted from the included reviews: publication date, country, population, participant characteristics, sample size, review type, research question (including components of PICO), aims and objectives, setting, searched databases (years), inclusion/exclusion criteria (including study design), quality assessment method and result, risk of bias assessment and result (including funding sources), analysis/synthesis method, subgroup analysis, duplicate study selection and data extraction, protocol registration, protocol deviations and justification, list of excluded studies with reasons, declaration of funding, number of papers included in review, intervention targets reviewed (including their timing), outcomes included (all-cause or cause-specific mortality; age of death), time period between intervention target and outcome, summary of results including effect size summary statistics, publication bias statistics if meta-analysis performed, and reviewer interpretations, conclusions and identified gaps.

Data extraction: primary papers

From each included primary paper within each included systematic review, two reviewers (JAK and a second reviewer) will extract information on country, population, participant characteristics, exposure/intervention target detail, period between intervention target and outcome, subgroup analysis, specific age and cause of death. Discrepancies between reviewers will be discussed with a third reviewer (MO) if necessary.

Data synthesis

We anticipate variability in the presentation and availability of age-specific data in the obtained reviews (which may include deaths beyond age 60) and associated primary papers. Therefore, guided by life course theory,³¹ we intend to synthesise the data from the reviews and included primary papers on intervention targets for each age/stage of the life course (eg, childhood (0–10 years), adolescence (10–24 years),⁵² and young-to-middle adulthood (25–40 years, 41–60 years), and later adulthood (past age 60)), or across the life course. The aim of doing so is to provide clarity and to guide investment in preventive intervention targets to reduce the incidence of mortality between years 15 and 60, compared with reducing mortality beyond 60 years when deaths are more likely due to diseases of ageing.

Once all data have been extracted, they will be critically examined by JAK. The data will be synthesised from the eligible reviews and their primary papers that examine intervention targets preceding mortality. If data allow, we plan to narratively compare males versus females because young males have two times the risk of females of dying young.¹ Therefore, we expect that causes and drivers may vary between sexes. Additionally, while this is a global population-level review, in recognition that risk-reduction priorities vary across settings, if data is sufficient, we will compare the different targets relevant for HICs, compared with LMICs. Moreover, and in accordance with PROGRESS-Plus,⁵³ other indicators of social disadvantage may also be summarised separately if the reviews provide adequate subgroup information (place of residence, race/ethnicity, occupation, religion, education, social capital, socioeconomic status, plus age, disability, sexual orientation and gender identity).⁵³

All findings will be reported in accordance with Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines.⁵⁴ We will synthesise the results narratively because we anticipate clinical, methodological and statistical heterogeneity between the reviews and papers (eg, differences in comparison groups, exposure variable definitions and measurement, and review methodologies),⁴⁴ such that in-depth quantitative synthesis will not be appropriate. Narratively, we will summarise, compare and contrast the findings within eligible reviews and eligible primary papers within these reviews.

Where appropriate, we will describe and summarise any quantitative results available in any reviews with included meta-analysis. To avoid bias, we will avoid direct comparisons of the effect sizes for intervention targets between different reviews.⁵⁵ If multiple reviews on the same/similar topic are included, we will avoid introducing bias (ie, double counting outcome data arising from duplicate primary papers within overlapping reviews) by following the evidence-based decision tool developed by Pollock and colleagues to determine whether and how to include the overlapping reviews in our umbrella review.^{56 57} Although this decision tool was designed for reviews of healthcare interventions, it is applicable to multiple forms of umbrella reviews.

Assessment of reviews methodological quality

The conclusions drawn within reviewed studies are dependent on the quality and strength of evidence synthesised in that study and will likely vary across studies. Ideally, reviewed studies should meet a minimum standard of quality and have a low risk of bias. We will critically appraise the included reviews using the 16-item AMSTAR 2.⁵¹ This tool is suitable for reviews of randomised trials, non-randomised studies and observational evidence. AMSTAR 2 measures the review’s design, search strategy, criteria for inclusion and exclusion, quality assessment undertaken for individual studies, methods for data synthesis, publication bias and any conflicts of interest. Two reviewers will independently complete this quality assessment for each eligible review. Any disagreements will be discussed and resolved with a third author (JAK) if necessary. Based on AMSTAR 2 guidelines, each review will be categorised as having high, moderate, low or critically low confidence in review results. This quality rating will be presented as part of our data synthesis, and critically low reviews will be excluded from conclusions. In addition, depending on the number of included meta-analyses, we will assess the credibility of evidence and degree of meta-bias within such reviews.⁵⁸

Patient and public involvement

There is no patient and public involvement in the design, conduct, reporting or dissemination plans of this study.

Ethics and dissemination

To our knowledge, this will be the first systematic review to synthesise the evidence and quality of published reviews and meta-analyses, and their primary papers, that have evaluated intervention targets to reduce the risk of dying between the period of middle adolescence and middle adulthood, in general population samples. The findings of our review will inform the design of an international cross-cohort research consortium examining intervention targets for preventing mortality during this early period of the life course.

As the current study will review published review studies in which the primary studies reviewed have already obtained ethics committee approval, the current study does not require additional ethics committee approval. We will disseminate our findings in a peer-reviewed journal article and via presentations at national and international conferences. Links to publications will be circulated via social media.

Ethics statements

Patient consent for publication

Not applicable.