Determinants of inequalities in antimicrobial-resistant bloodstream infections among socioeconomically disadvantaged populations in high-income countries: a systematic review protocol

This protocol addresses an important and clinically relevant question. Inequalities in antimicrobial-resistant bloodstream infections remain considerably less well characterised than inequalities in infection or antimicrobial use more generally. The involvement of an information specialist, the inclusion of patient and public contributors, the intention to include non-English evidence, and the attempt to evaluate an LLM-supported screening workflow are notable strengths.

Major concerns_
1. Title and objective: “determinants” may imply causality that the review cannot establish
Page 1, Title:
“Determinants of inequalities in antimicrobial-resistant bloodstream infections…”
Page 5, Review objectives:
“The aim of this review is to ascertain the drivers and risk factors contributing to disparities…”
Most eligible studies are expected to be observational, ecological or surveillance based. Such studies may identify associations between social factors and AMR BSI, but it will not necessarily establish that these factors are causal “determinants” or “drivers.”
So, I suggest the authors to rewrite the “Title” _
Social and socioeconomic factors associated with inequalities in antimicrobial-resistant bloodstream infections in high-income countries: a systematic review protocol.

In Abstract, authors stated_
“Studies will be eligible if they report a) quantitative data on AMR BSI and b) factors associated with reported health disparities…”
However, Page 7, Study design states:
“…surveillance reports…containing quantitative data on AMR BSI or antibiotic prescribing by socioeconomic status will be included.”
The use of “or” permits inclusion of studies that report antimicrobial prescribing but contain no BSI or resistance outcome. Such studies cannot directly answer the primary question concerning inequalities in AMR BSI.
The authors should either exclude prescribing-only studies or define a separate secondary review question.
Page 4, “Introduction section”
“Over five million deaths annually are associated with AMR bacterial infections…”
Suggested revision:
“In 2019, an estimated 4.95 million deaths were associated with bacterial antimicrobial resistance, including approximately 1.27 million deaths attributable to resistance.”

• The distinction between deaths “associated with” and “attributable to” AMR should be retained throughout the manuscript.

Page 5 “To identify risk factors considered to be associated with disparities in the incidence and burden of AMR BSI…”
The terms “incidence” and “burden” currently encompass several distinct outcomes. A socioeconomic factor may increase the overall risk of bloodstream infection without increasing the probability that the infecting organism is resistant. Conversely, a factor may be associated with resistance among patients who already have BSI without materially changing the population incidence of BSI.
The review should therefore separate at least three questions:

1. Population-level question: Among people living in eligible countries, is socioeconomic or social disadvantage associated with the incidence rate of laboratory-confirmed AMR BSI?
2. Resistance-conditional question: Among patients with laboratory-confirmed BSI, is socioeconomic or social disadvantage associated with the probability that the causative organism is antimicrobial resistant?
3. Prognostic question: Among patients with established AMR BSI, is socioeconomic or social disadvantage associated with mortality, morbidity, length of stay or other clinical outcomes?

These should not be combined in one meta-analysis because their denominators are respectively the general population, all BSI episodes, and patients with AMR BSI.
Antimicrobial prescribing should be treated separately as a possible exposure, mediator or contextual outcome rather than being incorporated into the definition of AMR BSI burden.
Page 5, Rationale, references 34–36:
The protocol cites three recent systematic or scoping reviews addressing social determinants, health inequalities and AMR in high-income settings. The authors state that none focuses specifically on AMR BSI, which may justify the present review, but the degree of overlap is not described.
A brief comparison table would strengthen the rationale, covering:

• population;
• infection type;
• AMR outcome;
• socioeconomic exposures;
• geographical scope;
• search dates;
• and number of relevant BSI studies identified.

This is particularly important because the authors of the current protocol overlap with authors of some of the cited reviews. The manuscript should demonstrate clearly that the present review addresses a distinct question rather than subdividing an existing evidence base without sufficient clinical justification.
The protocol should also explain how primary studies and surveillance datasets already identified in those reviews will be used—for example, for citation chasing—without treating the reviews themselves as primary evidence.
Page 5, Secondary objectives:
“reported interventions to mitigate against AMR BSIs including preventative measures such as vaccination and healthcare infection, prevention and control measures”
An intervention that reduces AMR BSI overall does not necessarily reduce inequality. To address the stated equity question, an intervention study would need to report:

• differential effectiveness across socioeconomic groups;
• an interaction between intervention effect and a social stratifier;
• or a change in an absolute or relative inequality measure.

Otherwise, the intervention component becomes a separate review of AMR BSI prevention.
The authors should either remove this secondary objective or restrict it to studies that evaluate whether an intervention changes inequalities in AMR BSI.
Suggested revision:
To identify interventions evaluated for their effect on absolute or relative socioeconomic inequalities in AMR BSI incidence or outcomes.
Pages 6–7 “Socioeconomically disadvantaged populations in HICs with AMR BSI.”
“The primary outcome of interest…is incidence/burden of AMR BSI.”
The protocol does not define an eligible AMR BSI episode. Given the absence of publication-date restrictions and the expected diversity of surveillance systems, definitions are likely to vary considerably.
The protocol should specify:

• Whether BSI must be confirmed by blood culture
• Whether molecular detection without culture is eligible
• How common blood-culture contaminants will be handled
• Whether resistance must be phenotypically confirmed or whether resistance genes alone are sufficient
• Whether “intermediate” susceptibility will be classified as resistant
• Whether study-defined resistance is acceptable when CLSI or EUCAST criteria are not reported
• How recurrent isolates from the same patient will be handled
• Minimum interval defining a new BSI episode
• How polymicrobial BSI will be classified, and
• Whether community-onset and healthcare-onset BSI will be analysed separately.

“Sepsis” should not independently satisfy the BSI eligibility criterion because sepsis may occur without bloodstream infection, while bacteraemia may occur without sepsis.
A possible eligibility statement would be:
Studies will be eligible when they report microbiologically confirmed bloodstream infection, or a clearly defined national surveillance BSI episode, with phenotypic or genotypic antimicrobial resistance and provide data stratified by, or analytically associated with, at least one prespecified social or socioeconomic factor.
Page 7 “Exposures will include factors…such as healthcare access, antibiotic misuse, disability and comorbidity…”
This list combines socioeconomic exposures, structural determinants, clinical risk factors, potential mediators and possible confounders. For example:

• Socioeconomic deprivation may be the principal exposure
• Healthcare access and antimicrobial use may lie on the causal pathway
• Age and comorbidity may confound or modify the association;
• Ethnicity may represent a social stratifier rather than a biological exposure;
• Disability may be an exposure, a mediator, or a consequence, depending on the analysis.

The protocol would benefit from a conceptual framework or directed acyclic graph identifying the principal exposure, outcome, confounders, mediators and effect modifiers. Without this, the review may pool adjusted estimates that answer different causal questions.
This is particularly relevant to prior antimicrobial exposure. If the intended question is the total association between socioeconomic disadvantage and AMR BSI, adjusting for antimicrobial use may constitute overadjustment when antimicrobial use is part of the causal pathway. Conversely, if the authors seek a direct effect independent of antimicrobial exposure, that should be stated explicitly.
The authors should prespecify which adjusted estimate will be extracted when a study reports several models and should record the covariates included in each model.
Page 6, Table 1: “bloodstream infections OR bacteraemia OR sepsis OR Escherichia coli OR methicillin-resistant Staphylococcus aureus…”
The protocol does not state whether all bacterial pathogens and resistance phenotypes will be considered together. The epidemiology and social distribution of community-onset ESBL-producing E. coli, MRSA bacteraemia, carbapenem-resistant Enterobacterales, VRE, resistant Pseudomonas aeruginosa and resistant Acinetobacter baumannii are unlikely to be directly comparable.
The authors should prespecify clinically meaningful organism–resistance categories, for example:

• third-generation cephalosporin-resistant or ESBL-producing Enterobacterales;
• carbapenem-resistant Enterobacterales;
• methicillin-resistant Staphylococcus aureus;
• vancomycin-resistant enterococci;
• carbapenem-resistant P. aeruginosa;
• and carbapenem-resistant A. baumannii.

An overall pooled “AMR BSI” estimate should be avoided unless the authors can demonstrate that the pathogens, resistance definitions and epidemiological contexts are sufficiently comparable.
The title uses “antimicrobial-resistant,” but the search strategy and rationale appear to concern bacterial antibiotic resistance. The authors should clarify whether fungal bloodstream infections with antifungal resistance are eligible. If they are not, “antibiotic-resistant bacterial bloodstream infections” may be more precise.
The table provides only selected examples. It does not include a complete line-by-line strategy for any database. The extended-data section indicates that the OSF repository contains the PRISMA-P checklist, but it does not state that the full database strategies are available.
PRISMA-P requires a draft search strategy for at least one electronic database to be reported in sufficient detail for replication.

Page 6, “There will be no publication date…restrictions.”
Resistance classification has changed over time because susceptibility breakpoints, laboratory methods and definitions such as MDR, XDR, ESBL and carbapenem resistance have evolved. A microorganism categorised as susceptible under an older breakpoint may be classified differently under a later standard.
The data-extraction form should therefore include:

• susceptibility-testing method;
• CLSI, EUCAST or other interpretive standard;
• version or year of the breakpoint;
• whether “resistant” or “non-susceptible” was reported;
• and the study’s definition of multidrug resistance.

Studies that do not provide a reproducible definition of resistance should not be combined quantitatively with those that use recognised standards. Sensitivity analysis by breakpoint era or resistance-definition quality may be required.
This is particularly important because unrestricted publication

Page 6, Information sources:
“Dissertations and theses will be searched using Core, Researchgate and Academia databases.”
ResearchGate and Academia.edu are not structured bibliographic databases with stable and fully reproducible search functions. The authors should avoid referring to them simply as databases and provide the exact procedure used, including: search terms, search date, account or access restrictions, sorting method, number of results screened and any stopping rule.
Page 7, Comparator(s):
“Populations not experiencing socioeconomic disadvantage within the same HICs.”
Many relevant studies will report deprivation deciles or quintiles, a continuous deprivation score, household income categories or educational gradients rather than a binary disadvantaged/non-disadvantaged comparison.
The comparator criterion should therefore permit:

• comparisons across ordered socioeconomic categories
• continuous associations
• lowest-versus-highest category comparisons, and
• within-study social gradients.

The authors should prespecify how differently scaled deprivation measures will be harmonised. Area-level indices such as the English Index of Multiple Deprivation, the US Area Deprivation Index and the Australian SEIFA indices are country-specific relative measures and should not be treated as directly equivalent absolute exposures.
Where data permit, the authors could distinguish:

• extreme-group comparisons
• effect per unit or standard-deviation increase
• relative index of inequality, and
• slope index of inequality.

In Abstract, it stated that
“Thirteen reviewers will complete screening…”
And here at Page 8 Initial title and abstract screening:
“Ten reviewers will be involved in the study selection process…”
Figure 2 also refers more generally to “all reviewers.”
The exact number of human reviewers participating in title and abstract screening, full-text screening, data extraction and risk-of-bias assessment should be reported consistently. If different numbers are involved at different stages, this should be clearly stated rather than using one overall number.

Page 8, Use of a large language model for screening:
“An initial set of 100 abstracts will be independently reviewed…”.  “a recall threshold of 95% is set as the primary target.”
Recall is calculated among genuinely eligible records. Therefore, the precision of the recall estimate depends on the number of eligible abstracts in the validation set, not merely on a total sample of 100 abstracts.
If only five or ten of the 100 abstracts are eligible, an observed recall of 100% would still have considerable statistical uncertainty. If the authors intend to demonstrate, rather than merely observe, recall of at least 95%, approximately 59 genuinely eligible abstracts with zero false-negative classifications would be needed for a one-sided 95% lower confidence bound close to 95%.
The validation sample should therefore be enriched with known eligible and near-eligible records while retaining a representative random component. The authors should report:

• the number of eligible records in each dataset;
• recall with a confidence interval;
• the number and characteristics of false negatives;
• precision and specificity;
• and performance for difficult exclusion categories.

The clinically important issue is not only the overall recall but whether the model disproportionately misses studies with unusual terminology, non-English abstracts, ecological designs or non-standard resistance definitions.
“Across successive iterations, the LLM’s outputs will be compared to the majority expert labels…”
The text suggests that the same 100 abstracts may be used repeatedly for prompt refinement and performance assessment. If the prompt is optimised against this dataset, performance measured on the same records will be optimistic.
Figure 2 appears to show an “additional set” for final evaluation, but the corresponding text does not describe this independent set adequately.
The authors should clearly separate:

• a development dataset used for prompt refinement
• an independent, locked validation dataset used only once after the prompt is finalised, and
• the complete dataset to which the validated procedure will be applied.

Page 8, “The Llama 3.3 70b model will be used…”
“A free-text prompt will be drafted…”
The protocol should also clarify the screening logic. It appears that all human-included records will progress to full text and that LLM-included records will be added even when the human reviewer excluded them. This important safety rule should be stated explicitly.
My suggestion_
Protocol should report:

• Exact model identifier and version
• Complete final prompt
• System instructions (if any)
• Temperature and other sampling parameters
• How non-binary, missing or malformed responses will be handled
• Date on which the model was run
• Whether the internal UKHSA tool applies any preprocessing or post-processing.

Page 9, “ROBINS-I will be used for cohort studies…”
ROBINS-I is intended for non-randomised studies evaluating interventions, whereas most studies in this review will evaluate the effects of socioeconomic or social exposures. The current ROBINS-E framework is specifically designed for non-randomised studies of exposure effects. 
A single tool should not be assigned solely based on a “cohort” label. The appropriate tool depends on the study question and result being evaluated.
The authors should provide a design-to-tool matrix for:

• cohort studies of exposure effects;
• case-control studies;
• analytical cross-sectional studies;
• prevalence studies;
• ecological or spatial studies;
• surveillance reports;
• randomised intervention studies;
• and mixed-method studies.

A further difficulty is that national surveillance reports may not be adequately assessed by NOS, ROBINS-E or the JBI prevalence checklist alone. Important surveillance-specific domains include:

• completeness of laboratory participation;
• consistency of blood-culture sampling;
• episode deduplication;
• changes in resistance definitions;
• completeness of population denominators;
• linkage success for socioeconomic variables;
• and differential missingness.

These domains should be incorporated explicitly.
Page 10, “PROGRESS domains include Place of residence…Race/ethnicity…Gender/sex, Religion, Education, Socio-economic status…”
The title and population criteria imply that an eligible study must include a measure of socioeconomic disadvantage. However, the Introduction and proposed subgroup analyses also include ethnicity, age, sex, disability, migration status, religion and geographical region.
The eligibility implications are unclear. For example, would a study comparing AMR BSI rates between ethnic groups, but reporting no income, education, occupation or deprivation measure, be eligible?
The authors should adopt one of two approaches:

1. retain the socioeconomic focus and require every included study to report an individual-level or area-level socioeconomic indicator; or
2. broaden the review to social inequalities and revise the title, objectives and eligibility criteria accordingly.

If the broader approach is selected, a more accurate title may be:
Social and socioeconomic inequalities in antimicrobial-resistant bloodstream infections in high-income countries: a systematic review protocol.
Age, sex and comorbidity should not be described automatically as indicators of socioeconomic disadvantage. They may be examined as stratifiers or effect modifiers, but they should be conceptually separated from socioeconomic position.