We will follow best practice in design, conduct, and reporting of our prognosis review as detailed in the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2019). The review will be supported by the Cochrane Prognostic Methods Group, partners within the Cochrane Mental Health and Neuroscience Network, and the UK National Institute for Health Research Complex Reviews Support Unit (NIHR CRSU).

We used the PICOT (Patient/Problem; Intervention; Comparison; Outcome; Timing) system to design our review question (Schardt 2007) (Table 1). As recommended by the Cochrane Prognosis Methods Group, we will follow guidelines suggested by Riley 2019, to ensure that our review is designed, conducted, and reported in keeping with best practice recommendations.

Criteria for considering studies for this review

Types of studies

We will include prospective and retrospective longitudinal cohort and case‐control observational studies. We will not include cross‐sectional studies, as it is not possible to determine prognosis from this design. We will not include prospective case studies, defined here as having fewer than 20 participants. We will exclude studies that are published only as abstracts or posters at conferences, as these have not undergone peer review.

Types of participants

We will include any studies that recruited older adults (defined as mean age 50 years or above) that, at time of recruitment and at time of application of the anticholinergic burden measure, have either a known diagnosis of MCI or dementia established by a medical practitioner, cognitive impairment established via a cut‐off on a formal cognitive assessment, or are taking cholinesterase inhibitor drugs.  Where a mixed population was recruited, we will only include the study if the prevalence of dementia or MCI was more than 70%.

We will make no restrictions based on comorbidity or polypharmacy, but will record these factors in our data extraction. We will assess whether Acetylcholinesterase (AChE) inhibitor use was measured and consider any potential impact of this in our risk of bias assessment. We will include studies conducted in specific patient subgroups, such as Parkinson's disease, Schizophrenia, or stroke, provided they meet our other inclusion criteria.

We will include studies conducted in all settings. People recruited in various settings (e.g. care‐home vs community care) may differ in important demographics (e.g. mean age, dementia severity, clinical or lifestyle factors) that could alter the strength of the association between anticholinergic burden and cognitive decline or neuropsychiatric disturbance. If a study was conducted in a care‐home setting but does not report numbers with previous MCI or dementia, we will include the study in the review but remove it via a sensitivity analysis, if required.

Index prognostic factor

The prognostic factor of interest is anticholinergic burden from medications. We will include any study describing use of a scale that purports to measure cumulative exposure to medications with anticholinergic properties. Scales do not need to be described as validated for prediction of cognitive outcomes. Previously identified scales are listed in Appendix 1.

We will not choose a particular measure of primary interest as there is no consensus on the preferred measure, and there is substantial heterogeneity in clinical practice. However, if the Drug Burden Index (DBI) scale is utilised, we will only include data if anticholinergic burden data were reported separately.

Due to our expectation of a relatively sparse literature, we will not exclude studies that simply use a dichotomised present/absent method to investigate the association between anticholinergic medication use and risk of cognitive decline or neuropsychiatric disturbance. However, as severity of anticholinergic burden may play an important role in identifying any association, we will consider the potential impact of this approach in our risk of bias assessment.

We will not include studies that only measure use of anticholinergic burden via serum radioreceptor assay (SAA) measurement as this has limited clinical applicability.

Comparator prognostic factors

We are interested in the value of anticholinergic burden as a prognostic factor over and above other prognostic factors that may be common in this population. Hence, while we will include studies that only assess the unadjusted anticholinergic burden prognosis, we will also evaluate the prognostic effect of anticholinergic burden adjustment for core variables identified as fundamental to the putative link between anticholinergic burden and further cognitive decline or neuropsychiatric disturbance in people with MCI or dementia. We selected these variables on the basis of a Delphi discussion between the review authors and a wider multicentre collaborative, working in the field of anticholinergic burden research (Appendix 2). The chosen core variables were age, sex, comorbidities, and use of AChE inhibitors.

We will assess use of additional adjustments in our risk of bias assessment.

Outcome measures

Primary outcomes: we will include any study that assesses cognitive decline (i.e. change on a measure of cognitive function) or neuropsychiatric disturbance (defined as stressed and distressed behaviours, such as those measured via the Neuropsychiatric Inventory) as an outcome. In the case of people with MCI, we will also include studies that assess incident dementia as an outcome. For the outcome of cognitive decline, we will accept any multidomain cognitive assessment tool that is validated for the direct assessment of cognition. We will not include papers that only measure a single cognitive domain.  Only primary outcomes will be included in our summary of findings table. 

Secondary outcomes: we will also include studies that assess for risk of mortality, decline in physical functioning, and institutionalisation‐‐defined as  admission to a care home ‐‐ in people with pre‐existing cognitive impairment.

Timing: on the basis that anticholinergic effects on cognition or neuropsychiatric disturbance may be more rapid in a dementia population than in a cognitively‐unimpaired population, we will accept assessment for cognitive decline or neuropsychiatric disturbance at one month or longer following baseline anticholinergic burden assessment. We will evaluate the risk of reverse causality based on duration of follow‐up in our risk of bias assessment.

Search methods for identification of studies

Electronic searches

As reporting of prognostic factor studies is variable, it can be challenging to identify all relevant studies. We will adopt the procedure proposed by Geersing 2012 to maximise our ability to identify relevant prognostic studies. Specifically, as we will search for one prognostic factor, we will not adopt any specific search filter, but instead adopt a search that combines our prognostic factor (anticholinergic burden) with the population of interest (people with MCI or dementia).

We will search the following databases: MEDLINE (1946 to present; OvidSP), Embase (1974 to present; OvidSP), PsycINFO (1806 to present; OvidSP), CINAHL (Cumulative Index to Nursing and Allied Health Literature) (1950 to present; EBSCOhost), and ISI Web of Science Core Collection (1928 to present; ISI Web of Science) (Appendix 3). We will apply no language restrictions.

Searching other resources

We will supplement this with hand‐searches of references of all included studies and identified systematic reviews.

Data collection and analysis

Selection of studies

We will use Covidence systematic review software to identify relevant studies (Covidence). The review group Information Scientist will perform a 'first pass' screen to remove clearly irrelevant titles.

Two review authors (OK and CK) will independently screen studies identified via our search method. Titles and abstracts will be screened in the first instance, with the full text of potentially relevant studies then accessed to determine if the study meets our inclusion criteria. In cases of disagreement, a third review author (MT) will act as arbiter and make the final decision on study inclusion/exclusion.

Data extraction and management

Two review authors (OK and CK) will independently extract the data to a piloted proforma based on the CHARMS‐PF (CHecklist for critical Appraisal and data extraction for systematic Reviews of prediction Modelling Studies, adapted for prognostic factors) template (Riley 2019). We will contact authors for missing data where required. We will select two studies to trial our data extraction proforma (Fox 2011; Bishara 2020). We will extract all data onto a standard form (Appendix 4).

Assessment of methodological quality

Two review authors (OK and CK) will independently use the QUIPS (Quality in Prognosis Studies) checklist (Hayden 2012), assessing the included studies across the domains of: study participation; study attrition; prognostic factor measurement; outcome measurement; adjustment for covariates; reverse causation; statistical analyses; and reporting. We will use the QUIPS anchoring statements, but modify the content to suit our review topic based on consensus within the review author team.

We will judge each domain as low risk of bias, moderate/unclear risk of bias, or high risk of bias (Appendix 5). In cases of uncertainty we will contact original study authors for clarification, where possible.

Discussing reporting deficiencies: prognosis research is frequently confounded by poor reporting and possible publication bias. We will supplement our risk of bias assessment with a narrative discussion of reporting issues, highlighting when missing information may have affected results. Prognostic factor studies often do not register protocols, increasing the risk that not all studies (published and unpublished) will be identified, and there is a risk of small‐study effects (in which smaller studies with higher odds ratios (ORs) are more likely to be published than smaller studies with non‐significant ORs), which can bias meta‐analyses (Peat 2014; Riley 2019). We will use sensitive search filters for the population (people with MCI or dementia) and the prognostic factor (anticholinergic burden) without any specific filter for prognostic research to increase retrieval, and attempt to examine the likelihood of small‐study effects in our review by generating a funnel plot.

Data synthesis

We will evaluate risk of future cognitive decline or neuropsychiatric disturbance for anticholinergic drug users against non‐users. Where possible, we will pool summary estimates for each anticholinergic burden tool individually; then, as an exploratory analysis, pool summary estimates across all scales. We will conduct each meta‐analysis in two ways. First, as an all‐encompassing 'any anticholinergic drug use' variable; second, as an ordinal, hierarchical variable in which low, moderate, and high users are pooled separately, to investigate possible differential relationships based on anticholinergic burden severity. Low users are defined as those with a cumulative score of 1 on an anticholinergic scale; moderate users are defined as those with a cumulative score of 2 on an anticholinergic scale; high users are defined as those with a cumulative anticholinergic scale score of 3 or above.

We will pool data in two separate ways. In the first instance, we will pool data obtained from unadjusted analyses. In the second instance, we will pool data from fully adjusted analyses, provided age, sex, and comorbidities are controlled for as a minimum. We will pool ORs and hazard ratios (HRs) separately. We will calculate standardised mean difference (SMD) for linear data and pool this separately from dichotomous outcome data. Where data are not available, we will attempt to estimate data based on methods suggested by Tierney 2007. Where data are sufficiently similar to permit pooling, we will use a random‐effects approach given our expectation of high heterogeneity between studies. We will use Comprehensive Meta‐Analysis software to conduct all meta‐analyses (Comprehensive Meta‐Analysis Version 3).

We will conduct a sensitivity analysis, restricting to studies that have no high risk of bias domains.

For secondary (subgroup) analysis, we will assess risk by type of dementia, severity of dementia, APOe4 status, and by setting. We will also conduct analysis based on duration of follow‐up. We will create categories of <1 year and > 1 year, considering pooled rates for each timeframe individually.

These additional outcomes were decided upon through discussion among the review authors. We also plan to assess exposure, including exposure before enrolment into the study and exposure during the study.

Finally, if sufficient data are available, we plan to conduct a comparative analysis of the prognostic performance of the differing anticholinergic burden measures using a network meta‐analysis.

Investigation/description of heterogeneity

We will describe heterogeneity narratively based on consistency of association and effect size between anticholinergic burden and cognitive decline or neuropsychiatric disturbance, measurement of prognostic factor, outcome measurement and definition, and study design. We will not employ the I² statistic in our evaluation of heterogeneity. In prognosis research, individual studies often have large sample sizes resulting in narrow confidence intervals (CIs); this can cause high I² values even if inconsistency between studies is moderate (Iorio 2015).

It is possible that observed associations between anticholinergic drugs and cognitive decline or neuropsychiatric disturbance in people with MCI or dementia are driven by uncontrolled variables. Number of medical conditions or overall polypharmacy are possible moderators of observed effects and could differ by setting. To investigate this, if data allow, we will conduct a meta‐regression based on study recruitment setting (nursing home versus non‐nursing home), severity of dementia (mild/moderate/severe), comorbidity ('number of comorbidities' controlled for as a covariate, yes/no), and polypharmacy (controlled for as a covariate, yes/no).

We will also investigate potential differences in effect size based on control for concomitant use of AChE medications (control for AChE medication (yes/no).

Grading the evidence

We will use the GRADE approach to evaluate our overall confidence in the results. We will adapt the GRADE approach to suit prognosis research using methods consistent with Huguet 2013. Specifically, we will evaluate reported evidence in the following eight areas.

• Phase of investigation: phase 3 explanatory studies derived from bespoke cohort study designs that seek to explain the mechanisms behind an underlying association between anticholinergic burden and cognitive decline/ neuropsychiatric disturbances in people with MCI or dementia are considered to be a high level of evidence. Phase 2 explanatory studies that seek to confirm an independent association between anticholinergic burden and cognitive decline/neuropsychiatric disturbances are treated as moderate evidence, and hypothesis‐generating phase 1 explanatory studies are treated as weak evidence for any association between anticholinergic burden and cognitive decline/ neuropsychiatric disturbances.

• Study limitations: we will use the previously described QUIPS tool to evaluate the overall risk of bias of included studies. Our GRADE judgement will be based upon the overall certainty of the evidence. That is, if most (more than 50%) included studies are considered at high risk of bias in their reported association between anticholinergic burden and cognitive decline or neuropsychiatric disturbance, we will downgrade the evidence accordingly.

• Inconsistency: we will downgrade the evidence if associations between anticholinergic burden and cognitive decline/neuropsychiatric disturbances are heterogeneous (i.e. estimates of effect are variable across studies with regards to showing beneficial or detrimental effects and their confidence intervals show minimal or no overlap; the measure of the prognostic factor is highly variable; outcome measurement is highly variable; and there is methodological heterogeneity due to study design); and if the P value is low for the test of the null hypothesis that all studies in a meta‐analysis have the same underlying magnitude of effect.

• Indirectness: we will downgrade studies where their investigation does not fully match with our broader review question. Specifically, if the population in the included studies only represent a subset of the population of interest (e.g. a specific subtype of dementia only) then the evidence for the association between anticholinergic burden and cognitive decline/ neuropsychiatric disturbances will be downgraded for indirectness.

• Imprecision: we will downgrade if the evidence is generated by a few studies involving a small number of participants and most of the studies provide imprecise results; if there are insufficient numbers to meet the optimal information size in the meta‐analysis (i.e. if the total number of patients included is less than the number of patients generated by a conventional sample size calculation for a single adequately powered study), or if the CIs fail to exclude important benefit or important harm.

• Publication bias: due to inherent issues regarding publication bias in prognostic research, we will adopt the default position that publication bias is likely and downgrade the evidence unless our assessment of publication bias provided significant evidence to the contrary (i.e. a symmetrically distributed funnel plot, and evidence that the prognostic factor has been investigated in numerous cohort studies).

• Effect size: we will upgrade our confidence in the effect estimate when the effect size was moderate to large (e.g. a hazard ratio of 2.5 or above).

• Exposure‐response gradient: we will upgrade the evidence if there is an incremental increase in effect size with increasing anticholinergic burden.