Search strategy

Two independent reviewers (CC and ZLY) performed database searches in PubMed, Embase, the Cochrane library, Web of Science, and ClinicalTrials by using the following items: (heart failure OR heart insufficiency OR cardiac insufficiency OR myocardial insufficiency OR myocardi* failure OR cardia* failure OR heart decompensation OR cardiac decompensation OR myocardial decompensation) AND (osteoporosis OR osteoporos* OR osteopenia OR bone mineral density OR bone density OR bone densit* OR bone conten* OR bone loss*), from inception to April 16, 2019. There was no language restriction in our searches. We also performed a limited updated search from April 16, 2019 to June 21, 2020. Detailed search strategies in each database are reported in Supplementary material, tables S3-S7.

Eligibility criteria and study selection

Studies were eligible if they met the following criteria: 1) had a longitudinal design; 2) included patients with osteoporosis or osteopenia (defined by bone mineral density [BMD] assessment using any validated tool according to a clear standard diagnostic criteria or any documented medical or insurance records); 3) included a control group (with normal BMD, or without osteoporosis); 4) the endpoints were incident HF (defined as diagnosis information from any medical or insurance records or self‑reported HF based on a physician diagnosis confirmed by standard diagnostic criteria using medical records documenting series of symptoms, physical signs, and other supporting clinical findings); and 5) provided hazard ratio (HR) for incident HF. The primary outcome of our analysis was the risk of incident HF in patients with osteoporosis, assessed through HRs for incident HF.

Two authors (GZJ and YYY) developed the lists of studies independently during the selection process. A third author (CC) was prepared to adjudicate the lists, and discussed with the above 2 authors to reach a consensus if discrepancies were found. If 2 studies were based on the same cohort, the study with bigger sample size was included. The references of included articles were searched to identify additional potentially relevant publications. We also considered conference abstracts in our database searches. We contacted the corresponding authors to acquire the data to evaluate the eligibility of potentially relevant conference abstracts.

Data extraction and quality assessment

Two reviewers (CC and ZLY) independently extracted data from the included studies using a standard form. Disagreements between the 2 reviewers were solved by a formal discussion to reach a consensus. Data were extracted on: 1) basic characteristics of included studies (first author, publication year, country, race, number of patients, mean age, percentage of female patients, follow‑up period, adjusted covariates); 2) descriptions of exposure measurement (osteoporosis or osteopenia) and endpoint (incident HF); and 3) outcome data (hazard ratio). The quality of prognostic studies was assessed according to the recommendations by Hayden et al¹¹ using the following domains: study participation, study attrition, prognostic factor measurement, outcome measurement, confounding measurement and control, and analysis. Two reviewers (GZJ and YYY) performed the quality assessment of each of the included studies independently. Discrepancies were solved by discussion with a third reviewer (ZLY). The domains would be rated as yes, partly, unsure, or no for its appropriateness, respectively.

Statistical analysis

We performed a random‑effects meta‑analysis using Comprehensive Meta‑Analysis (CMA) 2.0 (Biostat, Englewood, New Jersey, United States).¹² In the primary analysis, we calculated the pooled HR and 95% CI from all included studies. We only included the HR adjusted for the highest number of covariates for each study. The same procedure was performed in the post hoc subgroup analysis. Study heterogeneity was evaluated by using the I² statistics and the χ² test. A value of over 50% for the former and a P value of less than 0.05 for the latter suggested significant moderate or higher heterogeneity.¹³ Sensitivity analyses were conducted as follows: First, a fixed‑effect model was additionally used to estimate the pooled HR of all included studies in order to compare with a random‑effects model. Secondly, we estimated the respective HRs by a random‑effects model by excluding one study at a time. Publication bias was not performed because of the limited number of included studies (less than 10) in the present review. A P value of less than 0.05 was considered significant.

Ethics statement

This study was approved by the Academic Administration Committee of Maoming People’s Hospital. Informed consent was not required.

Description of included studies

A total of 5451 articles were identified in the search. All articles were derived from the abovementioned databases, and no other sources of studies or unpublished studies were used. Then, 5428 articles were excluded after the screening of their titles and abstracts. Twenty‑three articles (5 in PubMed, 9 in Web of Science, 9 in Embase) were reviewed. Thirteen duplicated articles were excluded. Out of the 10 remaining full‑text studies, 7 were further excluded (mainly because of overlapping cohorts, case series reports, and missing HR data for incident HF) (Supplementary material, table S8), and 3 studies were ultimately included (figure 1).^14-16 Among included studies, the cohort study conducted in the United States had a prospective design,¹⁴ while the other 2 from Taiwan had a retrospective design.^15,16 The main features of the included studies are shown in table 1. All participants were recruited from patients with end‑stage renal disease (ESRD) in the study of Yu et al.¹⁶