Protocol registration and standard reporting
We used the Preferred Reporting Items for Systematic Reviews and Meta-Analyses Protocols checklist when writing this protocol (Additional file 1) (31). We have submitted and registered our study protocol with PROSPERO (CRD42020171366) and will follow PRISMA-IPD guidance for reporting the results of this review (32).
Information sources
We will perform systematic searches of MEDLINE, EMBASE, and CINAHL using OVID platform from their inception date to May 2020. A review of the gray literature will also be performed using Google Scholar. We will screen reference lists of included studies and relevant reviews to find additional studies that meet our inclusion criteria. An experienced librarian will refine the searches for individual databases. Additional file 2 provide a draft of our search strategy.
Study selection
We will include randomized trials, non-randomized studies, case-control studies, retrospective and prospective cohorts, single-arm longitudinal studies that: (1) enroll older adults (³ 50 years) undergoing elective or urgent surgery, (2) assess preoperative risk factors for post-operative delirium, and (3) measure the incidence of delirium (up to 1-week post-surgery) using the Diagnostic and Statistical Manual of Mental Disorders (DSM) criteria by a trained individual, Confusion Assessment Method (CAM) or CAM-ICU, or other validated delirium assessment tools. We will exclude studies that evaluated delirium retrospectively, studies conducted in the ICU setting, and studies that address delirium tremens. In the case of studies investigating heterogeneous patient populations, we will only use data from eligible patients for the analysis. We will not restrict our study selection based on language of publication.
Eligible preoperative risk factors for this review are: age, sex, education, body mass index, type of surgery, frailty, cognitive impairment, comorbidities (such as diabetes, renal insufficiency, cerebrovascular disease, coronary artery disease, myocardial infarction, neurological disorders, dyslipidemia, hypertension, and obstructive lung disease), hypoalbuminemia, psychotropic medication use, history of psychiatric disorders and depression, impairment of activities of daily living, smoking status, alcohol consumption, American Society of Anaesthesiologists physical status classification, Charlson Comorbidity Index, caregiver support, type of anesthesia, and pre-hospitalization.
Pairs of reviewers will independently screen titles, abstracts, and full-text articles of records retrieved through the searches using standardized, pilot-tested forms in Covidence, an online systematic review software (https://www.covidence.org). Any disagreements between reviewers will be resolved through discussion, and if necessary, through participation of a third reviewer.
Collection of study-level data and data storage
We will contact principal investigators or corresponding authors of eligible studies to inform them about the study and ask if they are willing to share IPD. Those who express an interest in collaboration will be invited to read the protocol and to discuss the requirements for data sharing. The data custodian/representative of the institution will then be invited to sign a data sharing agreement, which specifies the data requested, obligations, ownership of data, terms, and authorship of potential publications. We have developed a specific web link for collaborators with limited access to acquire more information and access the relevant information (https://www.pipra.ch/collaboration.html). We will request authors of eligible studies to provide ethics-approved de-identified data as an encrypted file. All cleaned datasets will be stored in password-protected files on a secured server at McMaster University only accessible to members of the IPD meta-analysis group. All datasets will be in their original formats and will be converted to a master dataset for analysis. Additional file 2 provide a draft copy of our data sharing agreement.
Study authors will be asked to share relevant data regarding candidate predictors and delirium outcomes regardless of whether such data has previously reported in relevant study publication(s). The IPD received will not be used for any other research apart from that described in the data sharing agreement. All collaborators (one representative per study) will be invited to be co-authors on manuscripts describing the principal IPDMA analyses, for which their data contributed, as outlined in this protocol, subject to them meeting the International Committee of Medical Journal Editors (ICMJE) criteria for authorship.
Risk of bias assessment
We will evaluate validity and bias in studies of prognostic factors using Quality In Prognosis Studies (QUIPS) tool (33). We will assess the study risk of bias using five domains (study participation, study attrition, prognostic factors, outcome measurement, and study confounding). We will use a modified QUIPS tool to rate each domain as low or high risk of bias as opposed to the original low, medium, and high risk of bias. We will use the individual domains, rated as low, or high risk of bias, to inform the overall risk of bias in each study. Studies with four or five low-risk domains will be considered at overall low risk of bias, and studies with two or more high-risk domains will be considered at overall high risk of bias. Risk of bias assessments will be performed in duplicate and independently.
Data synthesis and analysis
Data will be inspected for missing data and unusual outliers via range check for all included variables. We will resolve any issues with the data provided through communication and in collaboration with the original authors. Datasets will be accepted in any form, provided that all data are anonymized, and variables and categories are adequately labelled in English. Following satisfactory data control, each dataset will be converted to a common format and variables will be renamed in a consistent manner. The original scales of covariate measurements that are reported will be utilized, where possible. To ensure compatibility across studies, when required, attempts will be made to convert variables to the same scale for all studies. Once each study data is cleaned and standardized, individual study datasets will then be combined to form a new master dataset with a variable added to indicate the original study.
IPD meta-analysis can be conducted using either a one-stage or two-stage approach (34, 35). Random-effects linear regression fitted with mixed or logistic two-level regression model (with patients nested within studies) will be used in the one-stage approach (using ipdforst package in Stata). In this approach, the IPD from all studies will be pooled in a single step, while accounting for clustering of patients within studies. This approach is preferred when few studies or few patients per study are available for synthesis (34, 36).
Two-stage IPD meta-analysis methods can also incorporate covariates and interactions; however, the one-stage approach is likely to be more flexible and practical for modelling of non-linear associations of candidate predictors with outcome. Therefore, our primary analysis will be conducted using the one-stage approach (hierarchical model which will include both study-level and patient-level covariates in the same model). We will use two-stage approach as a sensitivity analysis on the final dataset. All models will allow for random effects and will be estimated using restricted maximum likelihood. Our outcome of interest is post-operative delirium, and we will estimate odds ratios with 95% confidence intervals to quantify the magnitude of outcome-factor associations. There will be no subgroup analyses required. We will use Stata (StataCorp., Release 15.1. College Station, TX) for all data analyses.
Assessment of certainty in evidence
For assessing our certainty in the prognostic value of the predictors, we will use the Grading of Recommendations, Assessment, Development, and Evaluation (GRADE) approach for prognostic factors (37). Based on GRADE guidance for prognostic studies because we are addressing the issue of prognosis and not causation, observational studies will start as high certainty but may be rated down based on the limitations in risk of bias, imprecision, inconsistency, indirectness, and publication bias (37, 38). Certainty can be rated as high, moderate, low, or very low. The GRADE approach is typically applied at the outcome level. Since the focus of our review is prognostic factors for the same outcome, assessment of our certainty will be applied at the level of each individual prognostic factor.