Risk prediction models for sarcopenia in maintenance hemodialysis patients: a systematic review and meta-analysis Authors: Luchen Chen, Huajuan Shen, Yongze Dong, Xiujun Xu, Qi Zhong, Danfeng Zhuang, Mengjiao Zhao Published: 2025 ABSTRACT AbstractBackgroundSarcopenia can severely affect patients undergoing maintenance hemodialysis. A high-quality prediction model could facilitate early identification and prevention. Despite the growing number of risk prediction models for sarcopenia in these patients, their quality and clinical utility remain uncertain.ObjectiveThis study aims to systematically review existing studies on risk prediction models for sarcopenia in maintenance hemodialysis patients.MethodsA comprehensive literature search was conducted across PubMed, Web of Science, Embase, The Cochrane Library, CINAHL, CNKI, VIP, CBM, Wanfang databases, and Clinical Trials.gov from their inception until May 12, 2024. Studies on sarcopenia risk prediction models for maintenance hemodialysis patients were included. Two independent reviewers screened studies using the Prediction Model Risk of Bias Assessment Tool (PROBAST) and the CHARMS checklist, applying predefined inclusion and exclusion criteria. Relevant data were extracted, and the risk of bias in the included studies was assessed.ResultsEighteen studies, encompassing 21 prediction models, were included. Sample sizes ranged from 60 to 805 participants, with outcome event incidence rates varying between 6.6% and 70.0%. The reported risk factors were age, gender, body mass index, grip strength and so on. The area under the receiver operating characteristic curve (AUC) for the models ranged from 0.73 to 0.955. Most studies had a high risk of bias, primarily due to issues related to study population selection and data analysis, including inappropriate data sources, insufficient outcome events, and poor management of missing data. Only two studies raised concerns regarding applicability.ConclusionCurrent models for predicting sarcopenia in maintenance hemodialysis patients exhibit a high risk of bias, as determined by PROBAST criteria. Future research should focus on improving existing models or developing new ones using rigorous methodologies.RegistrationThis study is registered with PROSPERO (registration number: CRD42024544944).Clinical trial numberNot applicable.Supplementary informationThe online version contains supplementary material available at 10.1186/s12882-025-04662-7.Keywords: Maintenance hemodialysis, Sarcopenia, Risk prediction model, Systematic review, Meta-analysis Background End-stage renal disease (ESRD) has become a major global public health concern, with a prevalence of approximately 10%-13% and a mortality rate as high as 30% [1]. Maintenance hemodialysis (MHD) is one of the primary alternative therapies for ESRD patients. According to recent estimates, approximately 850 million people worldwide suffer from kidney disease, and about 84% of those requiring renal replacement therapy ultimately undergo hemodialysis []. Maintenance hemodialysis (MHD) is one of the primary alternative therapies for ESRD patients. According to recent estimates, approximately 850 million people worldwide suffer from kidney disease, and about 84% of those requiring renal replacement therapy ultimately undergo hemodialysis [2, 3]. Patients undergoing long-term maintenance hemodialysis frequently experience complications such as pruritus and hypocalcemia, among which sarcopenia is a common syndrome [4]. Progressive renal dysfunction leads to the accumulation of uremic toxins, metabolic acidosis, and other pathological alterations, predisposing MHD patients to reduced muscle strength, selective changes in muscle architecture, and pronounced muscle atrophy. Moreover, the hemodialysis process itself inhibits muscle protein synthesis while accelerating protein degradation, further contributing to the loss of muscle mass [5, 6]. The reported prevalence of sarcopenia among patients undergoing maintenance hemodialysis varies widely in the literature, but remains consistently high. A meta-analysis including 10 observational studies from Asia, encompassing 1830 participants, reported a sarcopenia prevalence of 33% among Asian MHD patients [7]. A study by Umakanthan et al. [8] found that the prevalence of sarcopenia in MHD patients reached 18% in non-Asian populations. Furthermore, sarcopenia has been associated with a range of adverse outcomes. Patients with sarcopenia are more likely to experience declines in physical function, an increased risk of falls and fractures, prolonged hospitalization, and higher healthcare costs. In the long term, sarcopenia may also contribute to elevated risks of cardiovascular events and all-cause mortality [9, 10]. Therefore, early identification of MHD patients at high risk for sarcopenia and timely intervention are crucial for preventing the onset of sarcopenia and improving patient quality of life. Sarcopenia risk prediction models aim to estimate the likelihood of sarcopenia by integrating various factors such as laboratory test results (e.g., serum albumin and creatinine levels), medical history (e.g., age and gender), and clinical signs (e.g., grip strength). These models generally use statistical methods or machine learning algorithms, assigning specific weights to different variables to calculate an overall risk score [11]. Since sarcopenia often presents with subtle or no obvious clinical symptoms [12], predicting its occurrence can help healthcare professionals implement timely interventions and reduce the risk of further deterioration. Currently, assessment tools such as the SARC-Calf and Ishii scores are used to evaluate the risk of sarcopenia in MHD patients. The SARC-Calf scale and Ishii score, developed by Barbosa-Silva and colleagues, have been widely applied in clinical practice [13, 14]. However, studies have indicated that when using the SARC-Calf score, the cutoff point for calf circumference is critical; setting the threshold too low or ignoring sex differences may result in under-detection [15]. Moreover, these tools lack specificity for the MHD population, cover a limited range of factors, and are unable to quantitatively link specific risk factors to sarcopenia risk. Although research on this topic has increased in recent years, the quality of existing studies on sarcopenia risk prediction models for MHD patients varies widely. Some studies have small sample sizes, lack external validation, or fail to adequately consider dialysis-specific factors such as dialysis adequacy, nutritional status, or inflammatory markers. These limitations hinder the widespread application of current models in clinical practice. Therefore, this study aims to systematically analyze and evaluate existing sarcopenia risk prediction models for MHD patients, focusing on their construction methods, applicability, and robustness. This analysis will provide valuable insights for improving clinical practice and guide future research directions. Methods Search strategy We conducted a systematic search across multiple databases, including PubMed, Web of Science, Cochrane Library, CINAHL (Cumulative Index to Nursing and Allied Health Literature), Embase, China National Knowledge Infrastructure (CNKI), VIP Database (Technology Journal Database), China Biomedical Literature Database, Wanfang Database and Clinical Trials.gov, to comprehensively review literature on risk prediction models for sarcopenia in patients undergoing maintenance hemodialysis. The search strategy combined controlled vocabulary (subject terms) and free-text keywords. Electronic searches served as the primary method and were supplemented by manual reference screening. The search period spanned from database inception to May 12, 2024. To ensure inclusivity and identify grey literature, additional searches were conducted using Baidu Scholar and Google Scholar. Detailed search strategies are provided in Appendix S1. Key search terms and their synonyms included “renal dialysis,” “sarcopenia,” and “risk prediction model.” Our systematic review followed the PICOTS framework, as recommended by the CHARMS checklist [16], which guided the review’s objectives, search strategy, and study inclusion/exclusion criteria [17]. The framework components are summarized as follows: P (Population): Patients receiving maintenance hemodialysis. I (Intervention): Published risk prediction models for sarcopenia in maintenance hemodialysis patients (predictors≥2). C (Comparator): No comparator model. O (Outcome): Predicting outcomes as sarcopenia (including probable sarcopenia, sarcopenia, or severe sarcopenia. T (Timing): Outcome predicted based on baseline information at admission, clinical scoring scales, and laboratory indicators. S (Setting): Risk prediction models intended to individualize sarcopenia prediction in MHD patients, facilitating preventive measures to avoid adverse events. Inclusion and exclusion criteria The inclusion criteria for studies were: (1) studies involving patients receiving regular hemodialysis; (2) studies that developed or validated a multivariable model for predicting sarcopenia risk in maintenance hemodialysis patients, including at least two predictors, with detailed descriptions of the model’s development, performance, validation, and reporting. No restrictions were applied regarding prediction time frames; (3) the outcome of interest was sarcopenia, encompassing probable sarcopenia, sarcopenia, or severe sarcopenia; (4) eligible study designs included case-control, cross-sectional, and cohort studies; (5) publications available in either Chinese or English. The exclusion criteria were: (1) models with fewer than two predictors; (2) reviews, commentaries, or meeting summaries; (3) prediction models based on virtual data; (4) studies for which the full text was inaccessible; (5) duplicate publications. Study selection and screening The study screening process was independently performed by two authors, Luchen Chen and Mengjiao Zhao. Duplicate studies were first removed, and the eligibility of the remaining studies was assessed based on their titles and abstracts. Full texts of studies meeting the inclusion and exclusion criteria were then thoroughly reviewed. Additionally, the reference lists of eligible studies were examined to identify any potentially relevant research. Any discrepancies were resolved through discussion with a third researcher, Yongze Dong, until consensus was achieved. Data extraction Data extraction was independently conducted by two researchers, Luchen Chen and Mengjiao Zhao, with cross-checked for accuracy. In cases where disagreements could not be resolved through discussion, a third researcher, Yongze Dong, was consulted for adjudication. A standardized data extraction form, developed and pilot-tested according to the CHARMS checklist [16], was independently applied by two researchers, Luchen Chen and Mengjiao Zhao, to assess the risk of bias and applicability of the included studies. In cases of disagreement, Yongze Dong was consulted for resolution. PROBAST is a tool specifically designed to critically evaluate studies focused on the development, validation, or updating of individualized prediction models. It consists of 20 signaling questions across four domains: study objectives (two questions), predictors (three questions), outcomes (six questions), and analysis (nine questions). A study is considered to have a high risk of bias in any domain if at least one signaling question is answered with “no” or “probably no.” The overall risk of bias is classified as low only when all domains are rated as low risk. If any domain is rated as “unclear,” but all other domains are rated as low risk, the overall risk is deemed unclear. The applicability assessment covers three domains: population, predictors and outcomes. The criteria for evaluating applicability are similar to those used in the risk of bias assessment. Statistical analysis A meta-analysis was performed to evaluate the predictive value of the model’s predictors using Stata software (version 15.0; Stata Corporation, College Station, Texas, USA). Effect estimates were reported as odds ratios (OR) with 95% confidence intervals (CI). Heterogeneity across the studies was assessed using the Cochrane Q test and the I2 statistic. A fixed-effects model was applied when p > 0.1 and I2 < 50%, indicating low heterogeneity. If p ≤ 0.1 and I2 ≥ 50%, significant heterogeneity was identified, and a sensitivity analysis was conducted. If heterogeneity remained, a random-effects model was used. Statistical significance was considered when p < 0.05. Results Study selection A total of 1889 studies were initially retrieved from the target databases. After removing duplicate records, 1380 studies remained. Following a step-by-step screening process, 18 studies were ultimately included in the analysis [18–35], encompassing 21 sarcopenia risk prediction models. Three of the studies, by Sánchez et al. [33], Liao et al. [24], and Bellafronte NT et al. [23], each contained two prediction models. The study selection process and reasons for exclusion are presented in S1 Appendix. Study characteristics All included studies were published in the past five years, primarily conducted in China [18–22, 24, 26–32, 34, 35], additional studies originated from Spain [33], Brazil [23], and Japan [25]. In terms of study design, the included studies comprised retrospective cohort, prospective cohort, cross-sectional, and case-control studies. Regarding diagnostic criteria for the predicted outcome, the majority of studies [18–22, 24, 26–32, 34, 35] applied the AWGS guidelines, some used the EWGSOP criteria, and one study [25] adopted the lumbar major muscle mass index. The key characteristics of the included studies are summarized in Table 1. Table 1Overview of basic data of the included studiesStudyCountryStudy designNumber of modelParticipantsdiagnostic criteriaMain outcomeSample SizeDevelopmentValidationShi,Qingbo [20](2024)ChinaCase-control study1Non-diabetic ≥18 years, HD ≥ 3 monthsAWGS 20191、2、339/126(30.95%)—Sánchez-Tocino ML [33](2023)SpainA retrospective study2Age 75–95, HD > 3 months, 210 min/session blood flow ≤300 ml/min.EWGSOP 20194、5、6Sarcopenia:24(40%);Severe sarcopenia17(28.3%)—Senzaki [25](2023)JapanA multi-center prospective cohort study1HD > 6 months, with non-enhanced abdominal CT.CT7150/441(34.01%)79/178(42.25%)Liao,Hualong [24](2023)ChinaCase-control study2 > 18 years, HD > 3 months.AWGS 2019—Male: 58/140(41.43%)Female: 65/102(63.73%)—Qin,Hongju [27](2023)ChinaCase-control study1MHDAWGS 20141、6、858/246(23.58%)—Xie,Danshu [32](2023)ChinaCross-sectional study1≥18 years, HD ≥3 monthsAWGS 20191、2、977/510(15.07%)47/246(19.11%)TianR [35] (2023)ChinaCase-control study1≥18 years, HD ≥ 8 weeksAWGS 20191、863/167(37.28%)—Cai,Qian [18](2023)ChinaCase-control study1MHDAWGS 20191、2、372/197(36.55%)—Yang,Wenjuan [28](2023)ChinaCross-sectional study1MHDAWGS 20191、2、378/350(22.3%)23/110(20.9%)Guo L [22] (2023)ChinaCross-sectional study1ESRDAWGS 20191、263 (37.28%)—Du,Xiaoju [26](2022)ChinaCross-sectional study1 > 18 years, CKD stage Ⅶ (eGFR < 15 ml/min), HD > 3 monthsAWGS 20191、2、3101/589(17.1%)35(18.1%)Bao,Qun [30](2022)ChinaA retrospective study1≥18 years, HD > 3 monthsAWGS 20191、2、1056(27.2%)—Ding,Yan [29](2022)ChinaCross-sectional study1≥18 years, HD > 3 monthsAWGS 20191、2、3——Zhou,Qifan [31](2022)ChinaCase-control study118–60 years, HD > 3 months.AWGS 20191、2、341(17.30%)—Cai,Genlian [34](2022)ChinaA retrospectively study1≥18 years, HD ≥ 3 months.AWGS 20191、2、359(15.99%)43 (17.48%)BellafronteNT [23](2021)BrazilA prospective cross-sectional study218–60 years, HD ≥ 6 months,with AVFEWGSOP 201911、1255 (70%)—Lin,Tingyun [21](2021)Taiwan of ChinaA retrospectively study1 > 20 years, HD > 3 monthsAWGS 201911、217 (6.60%)—Du,Xiaoju [19](2021)ChinaCross-sectional study1Non-diabetic ≥18 years, HD > 3 monthsAWGS 2019174/304(24.3%)68/309(22.01%)1 = skeletal muscle index was < 7.0 kg/m2 in men and 5.7 kg/m2 in women; 2 = Male and female patients with HGS < 28 and < 18 kg, respectively; 3 = 6-m walk speed < 1.0 m/s; 4 = Skeletal muscle mass was < 20 kg in men and 15 kg in women; 5 = Male and female patients with HGS < 27 and < 16 kg, respectively; 6 = 4-m walk speed < 0.8 m/s; 7 = CT (PMI < 7.17 cm2/m2 for males and < 5.13 cm2/m2 for females); 8 = Male and female patients with HGS < 28 and < 18 kg; 9 = 4-m walk speed < 1.0 m/; 10 = 8-m walk speed < 1.0 m/s; 11 = SMI: DXA and BIA; 12 = Grip strength (male < 20 kg, female < 15 kg) Overview of basic data of the included studies Sánchez-Tocino ML [33] (2023) Sarcopenia:24(40%); Severe sarcopenia17(28.3%) Male: 58/140(41.43%) Female: 65/102(63.73%) Cai,Qian [18] (2023) Bao,Qun [30] (2022) BellafronteNT [23] (2021) Du,Xiaoju [19] (2021) 1 = skeletal muscle index was < 7.0 kg/m2 in men and 5.7 kg/m2 in women; 2 = Male and female patients with HGS < 28 and < 18 kg, respectively; 3 = 6-m walk speed < 1.0 m/s; 4 = Skeletal muscle mass was < 20 kg in men and 15 kg in women; 5 = Male and female patients with HGS < 27 and < 16 kg, respectively; 6 = 4-m walk speed < 0.8 m/s; 7 = CT (PMI < 7.17 cm2/m2 for males and < 5.13 cm2/m2 for females); 8 = Male and female patients with HGS < 28 and < 18 kg; 9 = 4-m walk speed < 1.0 m/; 10 = 8-m walk speed < 1.0 m/s; 11 = SMI: DXA and BIA; 12 = Grip strength (male < 20 kg, female < 15 kg) Characteristics of included prediction models A total of 21 predictive models were developed, primarily using logistic regression, while one study applied machine learning algorithms for model comparison [18]. In terms of handling of missing data, five studies [21, 24–26, 34] reported missing data and treatment methods. Of these, two studies [21, 26] excluded missing data directly, while the other three studies [24, 25, 34] used mean substitution, complete case analysis and multiple imputation methods to process the missing data. The specific details regarding model establishment are presented in Table 2. Table 2Characteristics of included prediction modelsStudyThe Number of Candidate VariablesVariable SelectionMissing Data treatmentModeling ApproachesPredictorsShi,Qingbo [20](2024)19The association was evaluated using the chi-square test. Then binomial logistic—Multivariate logistic regression1、2Sánchez-Tocino ML [33] (2023)36The association was evaluated using the chi-square test. Then binomial logistic__Binomial logistic regressionModel 1: 3、4、5、6、7;Model 2: 4、5、7、8、9、10Senzaki [25](2023)22Backward stepwise selection (logistic regression model)Eleven patients were excluded, and missing variables were imputed using multivariate chained equations across 20 datasets.Logistic regression5、11–17Liao,Hualong [24](2023)75Random Forest (RF) model、lasso regressionComplete case analysisLogistic regression model、RF、KNN、GaussianNB、SVM、MLP、DT、AdaBoost、GBDT、LGBMModel 1: 4、18、19Model 2: 4、20、21、22Qin,Hongju [27](2023)28Univariate analysis, then multivariate logistic regression__Logistic regression4、11、23、24、25Xie,Danshu [32](2023)19Univariate analysis, then multivariate logistic regression—Multiple linear regression4、5、11、13TianR [35](2023)26Univariate analysis, then multivariate logistic regression—Logistic regression4、5、11、26、27Cai,Qian [18](2023)44Univariate analysis, then multivariate logistic regression—Logistic regression4、5、11、28Yang,Wenjuan [28](2023)20Univariate analysis, then multivariate logistic regression—Logistic regression4、5、11、25、26、49Guo L [22] (2023)31regressions were used Univariate analysis, Pearson’s test, then the multivariate logistic regression—Multivariate logistic regression4、26 、29、30、31、32Du,Xiaoju [26](2022)36Univariate and multivariable logistic regressionExcludedBinary logistic regression4、26、33、34Bao,Qun [30](2022)15regressions were used Univariate analysis, Pearson’s test, then the multivariate logistic regression—Multivariate logistic regression8、23、35、36、37Ding,Yan [29](2022)24Univariate analysis, then multivariate logistic regression—Binomial logistic regression4、26、38Zhou,Qifan [31](2022)17Univariate analysis, then multivariate logistic regression—Logistic regression26、39、40Cai,Genlian [34](2022)27Univariate and multivariable logistic regressionReplaced by the series mean of the continuous variablesMultivariable logistic regression4、26、30、41、42BellafronteNT [23] (2021)14Binary and multivariate logistic regression——Model 1:5、43、44、45Model 2 : 4、5、44、45、46Lin,Tingyun [21](2021)12A stepwise multiple linear regression4 patients were excludedA stepwise multiple linear regression model4、5、45、48Du,Xiaoju [19] (2021)13Univariate and multivariable logistic regression—Logistic regression5,、26、33、471 = TyG; 2 = BDNF; 3 = albumin; 4 = age; 5 = gender; 6 = phosphate; 7 = cholesterol; 8 = MIS; 9 = lymphocytes; 10 = sub scapular fold; 11 = HGS; 12 = height; 13 = dry weight; 14 = primary cause of ESRD; 15 = diastolic blood pressure at start of session, 16 = pre-dialysis potassium; 17 = dialysis water removal in a session; 18 = parathyroid hormone; 19 = fasting blood glucose; 20 = creatinine after dialysis; 21 = arm without vascular access; 22 = total bilirubin; 23 = dialysis age; 24 = limb skeletal muscle mass; 25 = MQSGA score; 26 = BMI; 27 = GS; 28 = FFM; 29 = muscle thickness; 30 = hs-CRP; 31 = Young’s modulus value in the stretched state; 32 = SWV in the stretched state; 33 = calf circumference; 34 = Scr; 35 = combined with cognitive dysfunction; 36 = Hcy; 37 = irisin; 38 = phase angle; 39 = hemoglobin; 40 = upper arm muscle circumference; 41 = mid-upper arm muscle circumference; 42 = serum phosphorus; 43 = resistance index; 44 = reactance; 45 = weight; 46 = height2/resistance; 47 = predialysis creatinine; 48 = total body water; 49 = body fat percentage Characteristics of included prediction models Sánchez-Tocino ML [33] (2023) Model 1: 3、4、5、6、7; Model 2: 4、5、7、8、9、10 Model 1: 4、18、19 Model 2: 4、20、21、22 Model 1:5、43、44、45 Model 2 : 4、5、44、45、46 1 = TyG; 2 = BDNF; 3 = albumin; 4 = age; 5 = gender; 6 = phosphate; 7 = cholesterol; 8 = MIS; 9 = lymphocytes; 10 = sub scapular fold; 11 = HGS; 12 = height; 13 = dry weight; 14 = primary cause of ESRD; 15 = diastolic blood pressure at start of session, 16 = pre-dialysis potassium; 17 = dialysis water removal in a session; 18 = parathyroid hormone; 19 = fasting blood glucose; 20 = creatinine after dialysis; 21 = arm without vascular access; 22 = total bilirubin; 23 = dialysis age; 24 = limb skeletal muscle mass; 25 = MQSGA score; 26 = BMI; 27 = GS; 28 = FFM; 29 = muscle thickness; 30 = hs-CRP; 31 = Young’s modulus value in the stretched state; 32 = SWV in the stretched state; 33 = calf circumference; 34 = Scr; 35 = combined with cognitive dysfunction; 36 = Hcy; 37 = irisin; 38 = phase angle; 39 = hemoglobin; 40 = upper arm muscle circumference; 41 = mid-upper arm muscle circumference; 42 = serum phosphorus; 43 = resistance index; 44 = reactance; 45 = weight; 46 = height2/resistance; 47 = predialysis creatinine; 48 = total body water; 49 = body fat percentage Performance of risk prediction models Model AUCs ranged from 0.73 to 0.955, indicating moderate to excellent discrimination. Regarding model validation, five studies [19, 23, 25, 26, 28] conducted external validation, while nine studies [22, 25, 27, 29–32, 34, 35] performed internal validation. Most models were presented as nomograms (n = 7), while others used risk-scoring formulas (n = 5) or tables (n = 1). The Discrimination and Calibration Measures of Prediction Models are illustrated in Table 3. Table 3Discrimination and calibration measures of prediction modelsStudyValidation methodAUCCalibration methodModel PresentationInternal VerificationExternal ValidationDevelopmentValidationShi,Qingbo [20](2024)——0.939———Sánchez-Tocino ML [33](2023)——Model 1: 0.8 model 2: 0.866——FormulaSenzaki [25](2023)Bootstrapspatial0.810.73Calibration curve—Liao,Hualong [24](2023)——Male: 0.874(0.829–0.918)Female: 0.776(0.697–0.853)———Qin,Hongju [27](2023)Bootstrap——0.924—NomogramXie,Danshu [32](2023)Bootstrap0.929(0.904–0.953)0.955(0.931–0.979)Calibration curveNomogramTianR [35] (2023)Randomly divided—0.906(0.862–0.940)0.917(0.846–0.962)calibration curveNomogramCai,Qian [18](2023)——0.817(0.759–0.875)———Yang,Wenjuan [28](2023)—Geographic0.902(0.869–0.935)0.894(0.832–0.955)Hosmer-Lemeshow testNomogramGuo L [22] (2023)Bootstrap—0.8520.845Calibration curveNomogramDu,Xiaoju [26](2022)—Temporal0.922(0.899–0.946)0.913(0.870–0.956)calibration curve、Hosmer-Lemeshow testNomogramBao,Qun [30](2022)Bootstrap——0.829(0.742–0.916)calibration curve、Hosmer-Lemeshow testNomogramDing,Yan [29](2022)Randomly divided—0.925(0.892–0.958)0.917(0.872–0.961)Hosmer-Lemeshow testFormulaZhou,Qifan [31](2022)Randomly divided—0.862(0.792–0.932)0.866(0.788–0.0.943)Hosmer-Lemeshow testFormulaCai,Genlian [34](2022)Bootstrap—0.869(0.822–0.915)0.832(0.765–0.900)calibration curve、Hosmer-Lemeshow testRisk scoringBellafronteNT [23](2021)—Domain—Model 1: 0.759Model 2: 0.738—FormulaLin,Tingyun [21](2021)____Male: 0.945female: 0.940______Du,Xiaoju [19](2021)__Temporal0.9110.819Hosmer-Lemeshow testFormula Discrimination and calibration measures of prediction models Sánchez-Tocino ML [33] (2023) Male: 0.874(0.829–0.918) Female: 0.776(0.697–0.853) Cai,Qian [18] (2023) BellafronteNT [23] (2021) Model 1: 0.759 Model 2: 0.738 Male: 0.945 female: 0.940 Results of quality assessment Table 4 summarizes the risk of bias and applicability of the included studies. All included studies were judged to have a high overall risk of bias. In the domain of study participants, nine studies [19, 22, 26, 28–30, 32–34] exhibited high risk due to retrospective data collection, inconsistent predictor definitions across centers, or complex biomarker measurements that limited data reliability. In the domain of outcomes, six studies [18, 25, 27, 28, 32, 35] used predictors overlapping with the diagnostic criteria of sarcopenia, potentially inflating model performance. Table 4PROBAST results of the included studiesStudyParticipantPredictorsOutcomeAnalysisOverall risk of biasApplicability①②risk of bias③④⑤risk of bias⑥⑦⑧⑨⑩⑪risk of bias⑫⑬⑭⑮⑯⑰⑱⑲⑳risk of biasShi,Qing [14] (2024)11+112—111111+231311221——+Sánchez-Tocino,M.L [27](2023)23—221—111111+231311221——+Qin,Hongju [21] (2023)31+111+112111—221321211——+Liao,Hualong [18] (2023)11+111+111131+111111221——+Senzaki [19](2023)11+211—112111—211111111———Cai,Qian [12](2023)11+111—112111—211321221——+Xie,Danshu [26](2023)21—221+112111—111321111——+Yang,Wenjuan [22](2023)21—121—112111—211321121——+Guo L[16](2023)21+111+111111+211321111———Tian R [29](2023)11+111+112111—211321121——+Du,Xiaoju [20](2022)21—121—111111+211221121——+Bao,Qun [24](2022)21—122—111111+211321111——+Ding,Yan [23](2022)21—121—111111+211321121——+Zhou,Qifan [25](2022)11+111+111111+211321121——+Cai,Genlian [28](2022)21—121—111111+211121111——+Lin,Tingyun [15](2021)11+111+111111+111211221——+Bellafronte NT[17] (2021)11+111+111111+231311221——+Du,Xiaoju [13] (2021)21—121—111111+111321121——+1: Yes / maybe; 2: No / probably not; 3: No information; "+": low risk of bias / high applicability; "−": high risk of bias / low applicability. ① Is the data source appropriate? ② Are the inclusion and exclusion criteria of the participant reasonable? ③ Are the definitions and evaluations of predictors the same for all participants? ④ Is the evaluation of the predictor performed without knowing the outcome data? ⑤ Are predictors included in the predictive model valid? ⑥ Is the classification of the results reasonable? ⑦ Is the definition of the result reasonable? ⑧ Does the definition of the outcome exclude predictors? ⑨ Is the definition of results the same for all participants? ⑩ Is the information about the predictor unclear when determining the result? ⑪ Is the time interval between the predictor evaluation and the result determination reasonable? ⑫ Is the sample size reasonable? ⑬ Is the treatment of continuous and categorical independent variables appropriate? ⑭ Are all included participants included in the statistical analysis? ⑮ Are missing data included in the appropriate treatment? ⑯ Does the use of single factor analysis to screen predictors avoided? ⑰ Is the complexity of the data considered? ⑱ Is the performance of the predictive model evaluated? ⑲ Whether to consider the over-fitting, under-fitting and best-fitting of the predictive model? ⑳ Are the predictors and their weights consistent with the reported results? PROBAST results of the included studies Guo L[16] (2023) Tian R [29] (2023) 1: Yes / maybe; 2: No / probably not; 3: No information; "+": low risk of bias / high applicability; "−": high risk of bias / low applicability. ① Is the data source appropriate? ② Are the inclusion and exclusion criteria of the participant reasonable? ③ Are the definitions and evaluations of predictors the same for all participants? ④ Is the evaluation of the predictor performed without knowing the outcome data? ⑤ Are predictors included in the predictive model valid? ⑥ Is the classification of the results reasonable? ⑦ Is the definition of the result reasonable? ⑧ Does the definition of the outcome exclude predictors? ⑨ Is the definition of results the same for all participants? ⑩ Is the information about the predictor unclear when determining the result? ⑪ Is the time interval between the predictor evaluation and the result determination reasonable? ⑫ Is the sample size reasonable? ⑬ Is the treatment of continuous and categorical independent variables appropriate? ⑭ Are all included participants included in the statistical analysis? ⑮ Are missing data included in the appropriate treatment? ⑯ Does the use of single factor analysis to screen predictors avoided? ⑰ Is the complexity of the data considered? ⑱ Is the performance of the predictive model evaluated? ⑲ Whether to consider the over-fitting, under-fitting and best-fitting of the predictive model? ⑳ Are the predictors and their weights consistent with the reported results? All studies were also judged at high risk in the analysis domain. Fourteen studies [18, 20, 22, 23, 25–31, 33–35] did not meet the recommended events-per-variable (EPV) criterion, increasing the risk of overfitting. Most studies relied on univariate analysis for variable selection, while only one [24] applied advanced methods such as LASSO and random forest. Furthermore, nine studies [18–21, 23, 24, 26, 28, 33] lacked model validation, and three studies [29, 31, 35] used simple data splitting, likely overestimating model accuracy. Two studies [22, 25] presented a high risk of concerns regarding applicability; specifically, Guo L et al. [22] raised issues in the participant domain, while Senzaki et al. [25] had concerns in the outcome domain. A detailed quality assessment is presented in Table 4. Meta-analysis results of predictors A meta-analysis was performed for predictors reported in at least three studies. Age, body mass index (BMI), and handgrip strength were identified as independent risk factors for sarcopenia in MHD patients. Substantial heterogeneity was observed for both age and BMI. Sensitivity analysis indicated that one study [34] contributed most to the heterogeneity in age due to design differences. After its exclusion, heterogeneity decreased (I2 = 47.8%, p = 0.074), and the fixed-effects model showed a pooled OR of 1.066 (95% CI: 1.052–1.080), confirming a robust association between age and sarcopenia. For BMI, a leave-one-out sensitivity analysis was performed to explore the source of heterogeneity. The results showed that after sequentially excluding each study, heterogeneity decreased markedly upon removal of the study by Du et al. [19], with a pooled OR of 0.660 (95% CI: 0.550–0.790). This heterogeneity may be attributable to population selection bias, as that study specifically included non-diabetic MHD patients. Egger’s test showed no evidence of publication bias (P>0.05). Corresponding funnel plots are presented in Appendix S1. Meta-analysis results of AUC Due to incomplete reporting of model development and validation procedures, only ten studies [18, 19, 24, 26, 28, 29, 31, 32, 34, 35] were eligible for inclusion in the meta-analysis of training set AUCs. Six studies [22, 29–31, 34, 35] were included for the internal validation set, and four studies [19, 23, 28, 32] for the external validation set. For the training set, substantial heterogeneity was observed (I2 = 65.5%, p < 0.001). Using a random-effects model, the pooled AUC was 0.894 (95% CI: 0.875–0.915), indicating good overall predictive performance. Sensitivity analysis showed that excluding Liao et al. [24] reduced heterogeneity and increased the pooled AUC to 0.904 (95% CI: 0.886–0.922), possibly due to the study’s small sample size and the use of multiple machine learning algorithms. Funnel plot asymmetry and Egger’s test (−7.66, p < 0.001) suggested potential publication bias, which might have led to an overestimation of predictive performance (Fig. 1). Forest plot of the random effects meta-analysis of pooled AUC estimates for training models In the internal validation set, considerable heterogeneity was also observed. Subgroup analysis by predictor type revealed lower heterogeneity within each subgroup. Egger’s test for this set was 0.62 (p = 0.569), indicating no significant publication bias (Appendix S1). Heterogeneity was particularly pronounced in the external validation set (I2 = 95.0%, p < 0.001). The pooled AUC from the random-effects model was 0.871 (95% CI: 0.809–0.939) demonstrating good predictive accuracy. A leave-one-out sensitivity analysis identified Bellafronte et al. [23] as a major source of heterogeneity. When this study was excluded, heterogeneity decreased to 60.2%, and the pooled AUC increased to 0.928 (95% CI: 0.891–0.967). Egger’s test (−5.69, p = 0.111) suggested no significant publication bias. Discussion Overall risk of bias in sarcopenia prediction models for MHD patients This study included 21 risk prediction models, with logistic regression as the predominant modeling approach. All included studies reported moderate to good predictive performance during model development or validation, with AUC values ranging from 0.73 to 0.955. However, according to the PROBAST checklist, all studies were assessed as having a high risk of bias, which limits the practical utility of these prediction models. Further optimization is required in sample selection, data handling, and model development and validation. The majority of studies were single-center with relatively small sample sizes. Eight studies [19, 26, 28–30, 32–34] used retrospective or cross-sectional designs. Developing models solely from existing data or medical records is not recommended, as it may introduce recall, information, and selection biases, thereby limiting clinical applicability. Additionally, fourteen studies [18, 20, 22, 23, 25–31, 33–35] reported events per variable (EPV) < 20. According to PROBAST [36], an EPV < 20 increases the risk of overfitting, potentially resulting in overestimation of model performance. Sample size calculation should follow either EPV≥20 or the clinical prediction model method proposed by Riley et al. [37]. Most studies, however, did not report their sample size determination method. Adequate sample size is essential for robust predictive modeling; future studies should adopt established guidelines to minimize overfitting and ensure reliable estimation of key parameters. Regarding data handling, only five studies [21, 24–26, 34] reported methods for addressing missing data, with two studies [21, 26] using simple deletion. Direct deletion may introduce bias in predictor-outcome relationships, compromising model reliability. PROBAST [36] recommends multiple imputation with transparent reporting. Proper handling of continuous variables and missing data, guided by expert knowledge, is crucial to minimize negative effects on model performance. Most studies selected variables based on univariate analysis, which may overlook interactions or collinearity, potentially omitting key risk factors. Techniques such as Ridge regression, stepwise regression, or LASSO regression, combined with evidence-based findings and clinical considerations, are recommended [36]. Only one study [24] employed multiple machine learning algorithms; the remaining studies used logistic regression. Logistic regression provides strong interpretability and robustness in small samples but may be less accurate when strong correlations or interactions exist between variables, potentially leading to bias or overfitting. Additionally, logistic regression has limited capacity to capture complex nonlinear relationships. Machine learning approaches can overcome these limitations, enhancing predictive accuracy and enabling automated feature extraction. However, this systematic review found that the AUC values of machine learning models were lower than those of logistic regression models, which is likely attributable to insufficient sample sizes leading to an underestimation of predictive performance. Furthermore, the lack of external validation for these machine learning models further limits the generalizability of the results. Therefore, researchers should select appropriate model development methods based on the specific research context. Before clinical implementation, models should undergo internal and external validation. The lack of validation is a common limitation in current clinical risk prediction models, often due to difficulties in obtaining clinical data and limited research resources. Only nine studies [22, 25, 27, 29–32, 34, 35] conducted internal validation, of which three studies [29, 31, 35] used random split validation; however, this method may reduce effective sample size and increase overfitting risk. More robust methods, such as bootstrap resampling, are recommended. Five studies [19, 23, 25, 26, 28] onducted external validation, and one study [25] combined internal and external validation, highlighting the limited external applicability of current models and the challenges in verifying their performance. Compared with redevelopment, direct external validation is more time- and cost-efficient. Future research should prioritize external validation of high-quality models or develop models with improved predictive performance by integrating traditional methods with machine learning approaches. Furthermore, during model evaluation, we observed that several studies demonstrated insufficient adherence to the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) statement [38]. This lack of transparency introduces uncertainty and potential bias into the models. Simultaneously, a high risk of bias further limits their applicability in real-world clinical settings. It is important to emphasize that without rigorous external validation, standardized calibration, and assessment based on clinical decision-making benefits, even models with high discriminatory performance may have their true predictive ability overestimated. This overestimation limits their effectiveness in guiding clinical practice or improving patient outcomes. Based on these limitations, the current models are not yet suitable for use as standalone clinical decision-making tools. Clinicians and renal dietitians should avoid over-reliance on model-generated predictions; however, these models may still serve as useful adjuncts for early screening and risk stratification. Future research should place greater emphasis on model generalizability and clinical feasibility, promoting a shift from “statistical performance” to “clinical utility”. Studies with larger sample sizes and multicenter designs are recommended, employing more rigorous methodological approaches and strictly adhering to frameworks such as TRIPOD [38] and PROBAST [36] criteria to enhance reporting transparency. Such efforts are essential for developing predictive models that are reliable, interpretable, and implementable in clinical practice. Differences and commonalities in predictive factor selection This study systematically reviewed existing risk prediction models for sarcopenia in patients receiving maintenance hemodialysis. Variations in study design and included variables resulted in differences in the risk factors incorporated and their relative weights. However, certain commonalities were observed. Independent predictors frequently reported in multivariable models included age, gender, body mass index, handgrip strength, albumin levels, malnutrition scores, and serum creatinine. These factors are characterized by their ow cost and easy accessibility. Age is a key predictor of sarcopenia. The results of our meta-analysis indicate that age is a significant predictor of sarcopenia in MHD patients, providing robust supporting evidence. Aging exacerbates muscle loss primarily through mechanisms such as the transition of muscle fiber types from type II (fast-twitch) to type I (slow-twitch) and increased fat infiltration [39]. These changes reduce the number of fast-twitch fibers, decreasing muscle speed and strength, ultimately impairing muscle mass and function. Additionally, the vascular network surrounding muscle fibers plays a critical role in muscle contraction, yet aging reduces the number of capillaries in contact with muscle fibers, further weakening muscle strength. Age-related changes also occur in motor neuron function and the stiffness of the extracellular matrix. Although the precise sequence of these changes, their contributions to muscle mass and function, and their roles at different life stages remain incompletely understood [40], some alterations may be irreversible. Nevertheless, supervised, personalized resistance training programs, combined with progressively increased intensity and protein supplementation, can partially mitigate adverse outcomes. Regarding gender, as a frequently reported predictor, remains controversial in its association with sarcopenia among MHD patients. Our meta-analysis aligns with some previous studies indicating that male patients are more prone to sarcopenia, potentially due to reduced sex hormone levels. In men, decreased testosterone levels induce overexpression of myostatin, which impairs muscle protein synthesis [41, 42]. In women, early estradiol deficiency during disease progression can lead to an earlier decline in muscle strength, increasing susceptibility to sarcopenia [10]. However, other studies [43] have reported no significant sex differences in sarcopenia incidence. These findings suggest that developing sex-specific prediction models may improve risk assessment accuracy. BMI is recommended for nutritional assessment in dialysis patients according to the 2020 Clinical Practice Guideline for Nutrition in CKD [44]. Low BMI indicates malnutrition and reduced muscle mass, whereas high BMI is generally considered protective against sarcopenia [29]. However, BMI alone does not distinguish between muscle and fat mass. High BMI due to greater muscle mass may be beneficial, whereas high BMI driven by fat accumulation may offer no protective effect or even pose additional risks. Among obese MHD patients, a subset presents with sarcopenic obesity, which is associated with higher body fat, elevated risk of cardiovascular disease, metabolic disorders, and mortality compared to patients with sarcopenia or simple obesity [45, 46]. Therefore, when using BMI to predict sarcopenia risk in MHD patients, further assessment of muscle-to-fat ratio is essential. Muscle strength, particularly handgrip strength, has also been identified as a predictive factor for sarcopenia in MHD patients. Handgrip strength is a key indicator for diagnosing sarcopenia and is commonly used to assess upper limb strength. Evidence suggests that muscle strength is a more effective predictor of adverse outcomes than muscle mass [47]. However, using handgrip strength as a predictive factor may introduce bias, and future studies should comprehensively consider additional relevant factors. Biochemical indicators such as albumin, malnutrition scores, and serum creatinine are also predictive of sarcopenia risk in MHD patients, reflecting protein-energy metabolism. Albumin, a common marker of nutritional status, is influenced by inflammation and fluid balance. Low albumin levels are strongly correlated with poor nutritional status and mortality [48], though its predictive power decreases after adjusting for inflammation. In MHD patients with anorexia and sarcopenia, hypoalbuminemia and inflammation often coexist, complicating differentiation. Serum creatinine, a metabolic byproduct of skeletal muscle, serves as a surrogate marker for muscle and lean body mass in patients without residual renal function [49]. Regular monitoring every 1–3 months is recommended to facilitate timely assessment and intervention. In summary, healthcare professionals should remain vigilant in identifying common risk factors in clinical practice, provide patient education to raise awareness of modifiable factors, and implement individualized management and adjustments based on each patient’s specific condition. Sources of heterogeneity A meta-analysis was performed on six internal models, resulting in a pooled AUC of 0.873 (95% confidence interval: 0.84–0.906), indicating a relatively high discriminative performance. However, an I2 value of 59.2% (p = 0.031) suggested moderate heterogeneity among the included studies. Subgroup analysis indicated that the type of predictors was an important source of heterogeneity: models incorporating only clinical predictors (pooled AUC = 0.92) demonstrated better and more stable discriminative performance than those combining clinical and laboratory indicators (pooled AUC = 0.85). This finding may be attributed to the fact that key clinical variables more accurately capture the underlying determinants of the outcome, whereas certain laboratory indicators are subject to short-term variability and may not reliably reflect long-term risk, thereby diminishing their predictive value. Therefore, in the development of sarcopenia prediction models, priority should be given to core clinical indicators with good stability and practical clinical applicability. Despite the overall satisfactory performance of the included models, considerable heterogeneity remained in several indicators, such as age, BMI, and AUC values. This heterogeneity may arise from multiple clinical and methodological factors, including differences in sample characteristics, outcome definitions, study design, and modeling approaches. First, sample characteristics varied, the included studies originated from various countries and regions, and the study populations differed in age distribution, body composition, and nutritional status. Such variability in population profiles may lead to inconsistent effect estimates. Second, inconsistency in sarcopenia diagnostic criteria likely contributed. The EWGSOP criteria are primarily based on European populations, using reference thresholds derived from young healthy adults [50], whereas the AWGS (2019) criteria are adapted to East and Southeast Asian populations [10]. Given the generally smaller body size, higher body fat proportion, and different lifestyle or physical activity patterns in Asian populations, these two standards may classify patients differently, thereby introducing applicability bias and effect variation. In addition, heterogeneity may also stem from differences in modeling strategies. Most studies employed logistic regression, while a minority adopted machine learning approaches. A notable example is the study by Bellafronte et al. [23], which was identified as the main source of heterogeneity in the external validation set, likely due to its small sample size and insufficient parameter tuning. Moreover, a considerable number of studies did not report internal or external validation results, which restricted the ability to further explore sources of heterogeneity through subgroup analysis or meta-regression. Due to the limited number of eligible studies for stratified analysis, the primary source of heterogeneity could not be conclusively determined. Taken together, future studies should aim to standardize methodological practices, including the harmonization of diagnostic criteria, detailed reporting of sample characteristics, transparent model development processes, and comprehensive validation strategies. Such efforts are essential to enable more robust meta-analyses and cross-population comparisons, thereby enhancing the generalizability and clinical credibility of prediction models. Future prospects and challenges With the rise of routine medical data and big data applications, machine learning and artificial intelligence models have become increasingly popular. Compared to traditional modeling methods, machine learning models can integrate large-scale, multidimensional, and complex data from electronic health records and manage a large number of input variables. However, variations in case information collection systems and nursing documentation formats across hospitals may prevent access to comprehensive data needed for modeling. This limitation restricts the clinical application and dissemination of such models. To facilitate the clinical adoption of these models, standardized electronic health record systems need to be established to enhance machine learning models. During model development, both simplicity and clinical applicability should be carefully considered. Simplicity refers to the accessibility of predictors and the ease of the assessment process. When selecting predictors, the “parsimony principle” should be followed, minimizing hard-to-obtain variables to reduce the risk of overfitting and enhance model generalizability. Based on the results of this meta-analysis, age, BMI, and handgrip strength are independent risk factors for sarcopenia in MHD patients; however, as handgrip strength serves as an outcome measure of sarcopenia, its inclusion as a predictor may introduce bias and should be carefully considered in future studies. A“core model”including age and BMI could be prioritized initially, with additional variable domains reasonably incorporated thereafter. For example, research by Liang et al. [51] indicates that active social participation and family interactions are important protective factors against potential sarcopenia. However, existing models have seldom incorporated psychosocial factors. Compared with previous models, social and familial factors are underrepresented in current prediction models. Social and familial factors, often overlooked, contribute to maintaining cognitive function and enhancing muscle and bone health, which may help prevent sarcopenia and osteoporosis [51]. Future research should comprehensively explore potential predictive domains to more accurately identify MHD patients at high risk of sarcopenia. Currently, there is a lack of empirical research evaluating the effectiveness of sarcopenia risk prediction models, which limits their practical use in guiding precise interventions for high-risk patients. Nevertheless, studies in other fields have preliminarily validated that predictive models can reduce the risk of adverse outcomes through risk alerts. For instance, Han Jing et al. [52] developed a hypoglycemia risk prediction model for newborns and implemented a graded management plan, which significantly reduced the incidence of hypoglycemia in newborns. Burdick et al. [53] applied a sepsis prediction model to 75,147 patients, finding that its use shortened hospital stays and reduced in-hospital mortality by 39.5%. Golijar et al. [54] applied the ASFPSI fall prediction model to 232 stroke patients, classifying them into low, medium, and high-risk groups and implementing targeted interventions. Fall rates decreased from 0.71% in 2006 to 0.28% in 2011. These findings suggest that future efforts should focus on developing high-quality risk prediction models based on methodological guidelines to identify individual sarcopenia risk. These models can then guide targeted, stratified prevention and management strategies to improve muscle mass, strength, and function in MHD patients. Limitation There are several potential limitations to this review. First, it was restricted to studies published in Chinese and English, which may have introduced publication bias. Second, most included studies were conducted in mainland China, potentially limiting the generalizability of the findings to Western populations. Adjustments may be required when applying these models in different regions. Consequently, future research should prioritize developing sarcopenia risk prediction models for maintenance hemodialysis patients across diverse populations. Lastly, only four predictors were included in our meta-analysis due to variations in study design and methodology among the included studies. This limitation hindered further exploration of heterogeneity sources and reduced the statistical power of the analysis. Although these findings reflect the methodological and reporting issues identified elsewhere, they did not affect the evaluation of model performance. Future studies should standardize methodological and reporting procedures to enhance comparability and reliability. Conclusion This analysis included 18 studies comprising 21 prediction models, all of which demonstrated good overall predictive performance. The pooled AUC of six internally validated models was 0.873 (95% CI: 0.84–0.906), indicating reasonable discriminative ability. However, according to the PROBAST assessment, all included studies were judged to have a high risk of bias; two raised concerns regarding applicability, and only a few models underwent external validation. Approximately 90% of the studies used traditional logistic regression for model development. Future research should focus on simplifying machine learning approaches to develop clinically practical and efficient intelligent prediction systems. In addition, investigators are encouraged to follow the PROBAST framework to optimize study design and to adhere strictly to the TRIPOD reporting guidelines, thereby enhancing the reliability and reproducibility of predictive models. Electronic supplementary material Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements None. Author contributions LuChen Chen: Writing – original draft, Visualization, Methodology, Formal analysis, Conceptualization. Huajuan Shen: Visualization, Resources, Investigation, Funding acquisition. Yongze Dong: Writing – review & editing, Validation, Data curation. Mengjiao Zhao: Writing – review & editing, Software, Formal analysis, Data curation. Xiujun Xu: Validation, Data curation. Qi Zhong: Resources, Funding acquisition. Danfeng Zhuang: Resources, Funding acquisition. Funding This research project is supported by the Health Science and Technology Plan of Zhejiang Province (2021KY468), and the Medical and Health Science and Technology Plan of Zhejiang Province (2022KY614). Data availability The data that support the findings of this study are available from the corresponding author upon reasonable request. Declaration Ethical approval The study had no ethical implications or concerns. Consent to participate Not applicable. Consent to publish Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request.