Commencing on September 29, 2023, a computerized search of various bibliographic databases, including EMBASE, Cochrane Library, and PubMed, was initiated. This comprehensive search utilized specific terms such as “Carcinoma, Pancreatic Ductal” [Mesh], “Pancreatic Neoplasms” [Mesh], “Cholangiocarcinoma” [Mesh], “Gallbladder Neoplasms” [Mesh], “Pancreatic Intraductal Neoplasms” [Mesh], “Bile Duct Neoplasms” [Mesh], “Liver Neoplasms” [Mesh], “Carcinoma, Hepatocellular” [Mesh], “Biliary Tract Neoplasms” [Mesh], “Bone Density” [Mesh], “Osteosarcopenia” [Title/Abstract], “Osteopenia” [Title/Abstract], among others. This search was limited to studies conducted in the English language involving human subjects. For a more detailed insight into our search strategy, we refer readers to Supplementary Material 1 . Furthermore, we conducted supplementary searches in grey literature using Google Scholar and performed a manual examination of the reference lists from qualifying studies. Following the protocols established by the Cochrane collaboration, the results from both manual and electronic sources were consolidated using the Covidence software to facilitate effective data management.

We have defined a predefined set of inclusion and exclusion criteria for the selection of articles. The inclusion criteria encompassed the following aspects: (i) Inclusion of articles centered on patients with a diagnosis of HBPC. (ii) Investigation of the prognostic value of the baseline of preoperative osteopenia and osteosarcopenia. (iii) The presence of at least one among the designated endpoints, including overall survival (OS), recurrence-free survival (RFS), or postoperative major complications (Clavien-Dindo grade ≥ III). Conversely, exclusion criteria consisted of: (i) Research of alternative typologies, such as animal studies, reviews, case reports, or conference abstracts. (ii) Research projects wherein the essential data required for hazard ratio (HR) or odds ratio (OR) computation regarding the designated endpoints was notably absent from both the textual content and the publicly accessible records. In cases where multiple studies contained overlapping cohorts of patients, priority was given to articles that presented comprehensive data and adhered to rigorous methodological standards (16).

During the process of data extraction, we systematically and comprehensively gathered essential information. This encompassed a wide range of details, including the author, publication year, study characteristics (region, period, and design), treatment, cancer type, demographic characteristics (such as sample size and gender distribution), outcome, the assessment of method and site, and the cut-off points. HR, OR, and their respective 95% confidence intervals (CIs) are predominantly sourced from multivariate analysis. In situations where such statistical values were not available, we either turned to univariate analysis or employed the Engauge Digitizer software to extract data from survival analysis graphs. In evaluating the caliber of the observational studies, we employed the Newcastle-Ottawa Scale (NOS) score. Studies that achieved a score of six points or higher were considered to be of high quality. It is worth highlighting that each stage of this procedure, encompassing literature retrieval, filtration, data acquisition, and evaluation of quality, underwent comprehensive and autonomous execution by a trio of investigators. In cases of disparities or contentions, such issues were referred to the senior author for adjudication.

In this study, Stata 15.0 software was utilized for the purpose of statistical analysis. The data presentation was executed using forest plots. To gauge the presence of heterogeneity within the dataset, we employed both Cochran’s Q test and I² statistics. Either “P-value < 0.1” or “I² value > 50%” was considered statistically significant heterogeneity. In cases of significant heterogeneity, we adopted a random-effects model, employing the DerSimonian-Laird method. Conversely, in the absence of significant heterogeneity, we adopted a fixed-effect model with the Inverse Variance method. We also conducted an assessment of potential publication bias, utilizing both Begg’s test (17) and Egger’s test (18). In order to evaluate the robustness of our observations, a sensitivity analysis was systematically executed, entailing the methodical exclusion of each individual study. Furthermore, we conducted subgroup analyses, taking into account factors such as Cox regression analysis, cancer type, and cut-off.

Figure 1 visually presents the results, as shown in the PRISMA flow diagram. In the preliminary stage, a cumulative sum of 623 research papers was discerned by means of database queries in conjunction with manual exploration. After removing duplicate entries, 515 distinct articles remained. A thorough review of the titles and abstracts led to the exclusion of 486 articles that did not meet the eligibility criteria. From the remaining pool, 29 articles underwent a comprehensive full-text review, resulting in the inclusion of 16 studies that met the established criteria for analysis (8, 19–35).

Table 1 provides an overview of the key characteristics of the studies included in this analysis. A total of 2,599 patients were enrolled in the research, with 65.81% being male. Geographically, thirteen studies were conducted in Japan, two in USA, and one in the Germany. These studies exhibited diversity in terms of the specific types of HBPC they enrolled. Five studies focused on patients with hepatocellular carcinoma (HCC), six on BTC, and five on PC. Fifteen studies used CT scans to measure BMD at the 11th thoracic vertebra to diagnose osteoporosis. Six studies defined osteopenia as an actual BMD lower than the calculated standardized BMD (308.82-2.49×age in men and 311.84-2.41×age in women). In five studies, patients with BMD < 160 HU were considered to have osteopenia. Furthermore, the 16 cohorts received NOS scores ranging from 6 to 8 ( Table 1 ).

In total, 16 studies investigated the impact of preoperative osteopenia on the OS of patients with HBPC, encompassing 2,599 individuals. We applied a random-effects model due to significant heterogeneity, as indicated by Cochran’s Q test and I² statistics (I² = 72.8%, p < 0.001). Our study revealed that individuals diagnosed with osteopenia exhibited significantly worse OS (HR: 2.27, 95% CI: 1.70-3.03, p < 0.001, Figure 2 ) in contrast to those devoid of osteopenia.

We performed subgroup analysis based on Cox regression analysis, cancer type, and cut-off values ( Table 2 ). Both univariate and multivariate analyses consistently supported that osteoporosis predicted shorter OS (multivariate analyses, HR: 2.46, 95% CI: 1.79-3.37, p < 0.001; univariate analyses, HR: 2.06, 95% CI: 1.40-2.80, p < 0.001), reaffirming the robustness of our research findings. Subgroup examination, stratified by cancer classifications, demonstrated that for individuals afflicted with HCC (HR: 2.31, 95% CI: 1.63-3.25, p < 0.001), BTC (HR: 2.75, 95% CI: 2.03-3.72, p < 0.001), and PC (HR: 1.76, 95% CI: 1.05-2.95, p = 0.031), the presence of osteopenia was associated with markedly diminished OS ( Table 2 ). Furthermore, it is noteworthy that the adverse impact on overall survival attributed to osteopenia persisted, irrespective of whether the diagnostic criterion applied was a 160 HU cut-off or age-adjusted BMD.

Eleven studies, involving 1556 patients, examined the relationship between osteopenia and RFS in HBPC patients. Due to significant heterogeneity among the studies (I² = 75.6%, p < 0.001), a random-effects model was employed. The pooled analysis revealed a markedly shortened RFS in HPBC patients suffering from osteopenia (HR: 1.96, 95% CI: 1.42-2.71, p < 0.001, Figure 3A ).

The results of both univariate and multivariate analyses substantiate the reliability of our research findings, which consistently supported the prediction of shorter RFS by osteopenia (multivariate analyses, HR: 2.64, 95% CI: 1.87-3.72, p < 0.001; univariate analyses, HR: 1.68, 95% CI: 1.14-2.48, p = 0.009; Table 2 ). Subgroup analysis by cancer type indicated that osteopenia significantly predicted worse RFS in patients with HCC (HR: 1.72, 95% CI: 1.21-2.45, p = 0.003) and BTC (HR: 2.61, 95% CI: 1.93-3.51, p < 0.001; Table 2 ). Nevertheless, in the case of PC, a tendency towards reduced RFS was observed among osteopenia patients; however, this trend did not attain statistical significance (HR: 1.56, 95% CI: 0.93-2.62, p = 0.090; Table 2 ). Considering that only four PC studies were included, a larger number of studies are needed to further confirm the relationship between osteopenia and the prognosis of patients treated with surgery for PC. Moreover, irrespective of the criteria employed to diagnose osteopenia, be it through the application of a threshold of 160 HU or the utilization of age-adjusted BMD, a consistent and robust correlation was observed with a poorer RFS outcome.

This analysis, depicted in Figure 3B , assesses the influence of osteopenia on the incidence of major postoperative complications. A low level of heterogeneity was observed, as indicated by the I² value (I² = 25.5, p = 0.243), prompting the utilization of a fixed-effects model. The results demonstrated a pooled OR of 1.66 (95% CI: 1.19-2.33) based on six studies encompassing 1135 patients. This signifies a statistically significant elevation in the risk of major postoperative complications among individuals with osteopenia in comparison to those without this condition.

Sensitivity analysis and investigations of publication bias were conducted to assess the robustness of the association between osteopenia and survival outcomes. The exclusion of individual studies did not exert a significant impact on pooled HR for OS, which ranged from 2.07 (95% CI: 1.59-2.71, post-exclusion of Abe K et al., 2021) to 2.43 (95% CI: 1.78-3.32, post-exclusion of Abe T et al., 2021, Figure 4 ). Similarly, for RFS, the HR exhibited minimal variations, ranging from 1.83 (95% CI: 1.33-2.51, post-exclusion of Yao et al., 2019) to 2.12 (95% CI: 1.46-3.06, after excluding Abe T et al., 2021, Figure 5A ).

The possibility of publication bias was assessed using Begg’s and Egger’s tests. Notably, the results of these tests did not indicate any significant publication bias for RFS (Egger’s test: p = 0.533, Begg’s test: p = 0.108). However, the funnel plot, Begg’s and Egger’s tests confirmed a significant publication bias in the case of osteopenia and OS (Egger’s test: p = 0.001, Begg’s test: p = 0.017, Supplementary Figure S1A ). To address this issue, we employed the trim and fill method, a valuable tool for estimating the number of potentially missing studies in OS. The results of this analysis indicated that even after incorporating these potentially missing studies, there was no discernible alteration in the pooled HR for OS ( Supplementary Figure S1B ).

In a similar vein, sensitivity analysis and assessment of publication bias have been employed to investigate the stability and reliability of the connection between osteopenia and postoperative major complications. The OR of the primary analysis remained unaltered upon the exclusion of any individual study ( Figure 5B ). Rigorous scrutiny through Egger’s and Begg’s tests affirmed the absence of notable publication bias (Egger’s test: p = 0.325, Begg’s test: p = 0.452).

A total of four studies, involving 671 patients, were included in this analysis, examining the impact of preoperative osteosarcopenia on OS or RFS in HBPC patients. Notably, there was significant heterogeneity among the included studies (OS: I² = 64.2%, p = 0.039; RFS: I² = 70.7%, p = 0.017), so a random-effects model was employed. Our findings demonstrated that patients with osteosarcopenia had significantly inferior OS (HR: 3.31, 95% CI: 2.00-5.48, p < 0.001, Figure 6A ) and RFS (HR: 2.50, 95% CI: 1.62-3.84, p < 0.001, Figure 6B ) compared to those without osteosarcopenia.

Sensitivity analyses and publication bias tests for the correlation between osteosarcopenia and OS as well as RFS revealed that the exclusion of any single study did not exert a substantial impact on the pooled results, and no indications of publication bias were detected ( Supplementary Figure S2 ). As elucidated above, our findings exhibit robustness and reliability.