Frailty prediction models in older adults with stroke in China: A systematic review and meta-analysis Authors: Lifang Huang, Juan Zhang, Shanjun Mao, Rulei Ding, Lan Luo, Guilin Shen, Kun Sun Published: 2026 ABSTRACT BackgroundFrailty is a crucial factor influencing the prognosis of elderly stroke patients. Recent prediction models have focused heavily on predictive performance while often neglecting rigorous methodological quality assessment, thereby limiting their clinical applicability. This systematic review and meta-analysis aims to evaluate the predictive performance and methodological quality of existing frailty prediction models for older stroke patients in China, identify the most frequently included predictors, and provide evidence-based guidance for future research.MethodsA comprehensive literature search was conducted in PubMed, Embase, Web of Science Core Collection, Cochrane Library, EBSCOhost, and several Chinese databases for studies published up to May 2025. Two reviewers independently extracted data using the Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies and assessed the risk of bias with The Prediction model Risk Of Bias ASsessment Tool. A random-effects meta-analysis was conducted using R software (version 4.5.1) to pool the area under the curve (AUC) values of the prognostic models for frailty.ResultsA total of 12 studies were included in the review, of which 11 were eligible for the meta-analysis. The model AUCs ranged from 0.629 to 0.915. Frequently identified key geriatric predictors included age, neurological function, activities of daily living, physical exercise, and nutritional status. Methodological appraisal indicated a high risk of bias across all models, primarily due to non-prospective designs and flaws in the analysis phase. The pooled AUC from the meta-analysis was 0.82 (95% CI: 0.76 - 0.88, I² = 92.0%, P < 0.0001). Limitations of the included studies were small sample sizes and inappropriate handling of continuous variables.ConclusionThis systematic review finds that frailty prediction models for older Chinese stroke patients have modest discriminative ability (pooled AUC 0.82). However, all models suffer from methodological flaws, including non-prospective designs, small sample sizes, and a lack of external validation. These limitations undermine their reliability and clinical utility. Rather than endorsing any specific model, this review maps the current evidence and identifies key areas for improvement. Future research should prioritize rigorous development and external validation in prospective studies. What is already known •Frailty is a key prognostic factor that adversely affects recovery and healthcare outcomes in older stroke survivors.•A variety of prediction models have been proposed to quantify frailty risk in this population.•Existing research on these models has predominantly emphasized predictive performance while paying limited systematic attention to their methodological rigor and clinical applicability. Alt-text: Unlabelled box dummy alt text What this paper adds •This study provides the first systematic evaluation of predictive performance and methodological quality of frailty models in stroke patients, revealing a substantial gap between reported accuracy and methodological rigor.•Universal methodological shortcomings in study design and statistical analysis fundamentally undermine the reliability and clinical applicability of existing models.•This review delineates critical methodological pitfalls and provides specific guidance for improving future research particularly in prospective multi-center studies. Alt-text: Unlabelled box dummy alt text What is already known •Frailty is a key prognostic factor that adversely affects recovery and healthcare outcomes in older stroke survivors.•A variety of prediction models have been proposed to quantify frailty risk in this population.•Existing research on these models has predominantly emphasized predictive performance while paying limited systematic attention to their methodological rigor and clinical applicability. •Frailty is a key prognostic factor that adversely affects recovery and healthcare outcomes in older stroke survivors. Frailty is a key prognostic factor that adversely affects recovery and healthcare outcomes in older stroke survivors. •A variety of prediction models have been proposed to quantify frailty risk in this population. A variety of prediction models have been proposed to quantify frailty risk in this population. •Existing research on these models has predominantly emphasized predictive performance while paying limited systematic attention to their methodological rigor and clinical applicability. Existing research on these models has predominantly emphasized predictive performance while paying limited systematic attention to their methodological rigor and clinical applicability. Alt-text: Unlabelled box dummy alt text What this paper adds •This study provides the first systematic evaluation of predictive performance and methodological quality of frailty models in stroke patients, revealing a substantial gap between reported accuracy and methodological rigor.•Universal methodological shortcomings in study design and statistical analysis fundamentally undermine the reliability and clinical applicability of existing models.•This review delineates critical methodological pitfalls and provides specific guidance for improving future research particularly in prospective multi-center studies. •This study provides the first systematic evaluation of predictive performance and methodological quality of frailty models in stroke patients, revealing a substantial gap between reported accuracy and methodological rigor. This study provides the first systematic evaluation of predictive performance and methodological quality of frailty models in stroke patients, revealing a substantial gap between reported accuracy and methodological rigor. •Universal methodological shortcomings in study design and statistical analysis fundamentally undermine the reliability and clinical applicability of existing models. Universal methodological shortcomings in study design and statistical analysis fundamentally undermine the reliability and clinical applicability of existing models. •This review delineates critical methodological pitfalls and provides specific guidance for improving future research particularly in prospective multi-center studies. This review delineates critical methodological pitfalls and provides specific guidance for improving future research particularly in prospective multi-center studies. Alt-text: Unlabelled box dummy alt text Introduction The population is aging, and frailty among elderly stroke patients presents a significant The population is aging, and frailty among elderly stroke patients presents a significant clinical challenge. Frailty is a geriatric syndrome characterized by multisystem dysregulation and a diminished physiological reserve. This condition is closely associated with advanced age and multiple health issues (Kim et al., 2024). The prevalence of frailty increases with age. It affects approximately 11% of individuals aged 50–59 years, and this figure rises to over 50% among those aged 90 and older (O’Caoimh et al., 2021). Stroke is a common and disabling disease associated with aging. A harmful bidirectional relationship exists between frailty and stroke, as they often coexist. Frailty increases the risk of stroke and is associated with poor functional recovery, a lower quality of life (Wæhler et al., 2021), and higher mortality (Li et al., 2024). Stroke can lead to or exacerbate frailty through neurological damage and reduced mobility. Neurobiological studies indicate that frailty is linked to vascular brain injury, characterized by increased white matter hyperintensity volume and reduced gray matter volume (Sciancalepore et al., 2026). Therefore, older adults who have experienced a stroke represent a highly vulnerable group requiring targeted and effective care strategies. A major challenge in clinical management is the under-recognition of frailty. Early identification of at-risk individuals is crucial for implementing proactive and personalized interventions. Strategies such as nutritional support and structured exercise can help prevent or delay functional decline, ultimately leading to improved long-term outcomes (Angulo et al., 2020; Lengelé et al., 2025). Effective prevention minimizes the adverse consequences associated with frailty. One key preventive step is to screen individuals who are at high risk of developing frailty (Shen et al., 2024). This stratification facilitates the efficient allocation of healthcare resources and enables targeted interventions, ultimately reducing the risk of disease progression. Recently, there has been significant growth in research focused on frailty risk prediction models specifically for elderly stroke patients (Gao et al., 2025; Ma et al., 2025). These models incorporate demographic, clinical, psychosocial, laboratory, and anthropometric data to assist healthcare providers in identifying high-risk individuals early and delivering appropriate interventions (Qiao et al., 2025). However, a concerning trend in current research is the excessive focus on reporting model predictive performance, particularly the area under the curve (AUC), while often neglecting a thorough evaluation of methodological quality. The reliability and clinical applicability of a model are fundamentally dependent on methodological rigor, which encompasses study design, sample size, handling of variables and missing data, and validation strategies (Moons et al., 2019). A model with a seemingly high AUC value, such as 0.915 (Shen et al., 2024), but a high risk of bias has limited clinical value. More importantly, it may mislead clinical decisions by providing inaccurate risk predictions. For example, although some studies have developed statistically robust prediction models with AUCs ranging from 0.78 to 0.94, many frailty prediction models carry a high risk of bias. One reason is their use of univariate screening to select predictors, which can severely compromise model validity (Zhang et al., 2024). Another issue is small sample sizes. For instance, one study included only 231 participants (Jiang et al., 2024), which may affect model stability and generalizability. Insufficient model validation is a common issue. Many models undergo only internal validation and lack external validation in independent, more representative populations. Consequently, their real-world performance remains uncertain. A scoping review of prediction model network meta-analyses highlights that important metrics, such as calibration—which assesses the alignment between predicted probabilities and actual observed risks—are frequently neglected (Yusufujiang et al., 2025). However, that review focused on network meta-analysis methodology across clinical fields and did not examine prognostic models for frailty in stroke patients, particularly in Chinese populations. These methodological weaknesses collectively impede the translation of existing models into clinical practice. Therefore, there is a pressing need for a comprehensive synthesis that evaluates both the predictive performance and methodological quality of available models. Such an effort should highlight current limitations and challenges to inform future research and enhance potential clinical applications. Accordingly, we conducted a systematic review and meta-analysis with the following objectives: (1) to systematically summarize the reported predictive performance of existing models for frailty in older stroke patients in China; (2) to rigorously assess the methodological quality and risk of bias in these models using The Prediction model Risk Of Bias ASsessment Tool; (3) to quantitatively synthesize their discriminative performance using meta-analysis; and (4) to identify commonly used geriatric predictors. This review systematically evaluates current evidence to highlight key methodological shortcomings and offer evidence-based recommendations to guide the development of more reliable, clinically applicable prediction tools. Given expected heterogeneity in model development and predictor selection, this review does not seek a single optimal model or aim to quantify independent predictive effects of individual variables. Instead, it provides a comprehensive overview of the current landscape, critically appraises methodological quality, and identifies common shortcomings to inform future research. Methods Search strategy To identify relevant studies on the development or validation of frailty prediction models in stroke patients, two reviewers [LFH and JZ] conducted comprehensive literature searches across multiple databases, including PubMed, Embase, Web of Science Core Collection, Cochrane Library, EBSCOhost (including MEDLINE, CINAHL, PsycINFO, and Academic Search Complete) and several Chinese databases (China National Knowledge Infrastructure, Wanfang Data, China Science and Technology Journal Database, and Chinese Biomedical Literature Database). Searches covered the period from database inception up to May 2025. We utilized a range of search terms, including: Aged / Aged, 60 and over / elderly / geriatric / older adult*, Stroke / Brain Ischemia / Cerebrovascular Accident/stroke* / cerebral infarct*, Frailty/frail* / physical frailty, and Predictive Model / risk prediction / nomogram / machine learning / random forest/logistic regression. Each database's search strategy was tailored individually, incorporating both controlled vocabulary and free-text terms to ensure comprehensive coverage. The complete search strategies for all databases are available in the Supplemental materials (Table S1). We also manually reviewed the reference lists of the included articles to identify additional relevant publications. To manage duplicates, we used EndNote to perform both automated deduplication and manual checks. Although our search was global, all eligible studies were from China, resulting in a final analysis focused on the Chinese population. Inclusion and exclusion criteria Studies were included if they met the following criteria: (1) used a cross-sectional, cohort, or case-control design; (2) enrolled adult stroke patients aged ≥60 years; (3) developed or validated a multivariable prediction model for frailty, with performance metrics reported (e.g., AUC); and (4) defined frailty using a validated instrument (e.g., Tilburg Frailty Indicator). We excluded secondary sources (non-original studies, e.g., reviews, protocols), studies involving ineligible populations or outcomes that focused on specific frailty subtypes not assessed by the included instruments (e.g., cognitive frailty, oral frailty), reported only univariate associations, or had full texts that were unavailable. Study selection Two reviewers [LF H and RL D] independently managed the search results using EndNote 2025 and performed the study selection process; disagreements resolved by consensus or by consulting a third reviewer [KS]. Following the removal of duplicates, we screened titles and abstracts to identify potentially eligible studies. The full texts of these records were then retrieved and assessed for final inclusion. Furthermore, the reference lists of all included studies were checked to identify additional relevant studies. This systematic review was conducted in accordance with the Joanna Briggs Institute Manual for Evidence Synthesis (which includes guidance for prognostic and prediction model studies) and reported following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses 2020 statement; the corresponding checklist is provided in Table S2 (supplementary materials). Data extraction and quality assessment Two reviewers [LF H and GL S] independently extracted data using a standardized form based on the checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies (Moons et al., 2014). Key study characteristics were extracted, including data sources, participant details, frailty definitions and instruments, model development and validation processes, and performance metrics. Risk of bias and applicability of the included prediction models were assessed independently by two reviewers [LF H and LL] using The Prediction model Risk Of Bias ASsessment Tool (Moons et al., 2019). This tool evaluates four key domains: participants, predictors, outcome, and analysis, through a series of signaling questions. For each domain, we assessed the following: for participants, we assessed whether appropriate inclusion criteria and data sources were used; for predictors, we examined whether they were clearly defined, available in practice, and measured without knowledge of the outcome; for outcome, we determined if it was well-defined and objectively measured; and for analysis, we checked if sample size was accounted for, missing data was handled appropriately, univariate screening of predictors was avoided, and both discrimination and calibration were assessed. A study was classified as having a low overall risk of bias only when all domains were rated as low risk. Conversely, a high overall risk classification was assigned when at least one domain was rated as high risk. An unclear overall rating was given if any domain was rated unclear, provided no domains were rated high risk. Data synthesis and statistical analysis Given the variation in methods and populations among studies, we employed both a narrative summary and a meta-analysis. Model performance was evaluated using established criteria. Discrimination was assessed by calculating the area under the curve (AUC), interpreted as follows: 0.7–0.9 indicates moderate accuracy, while values greater than 0.9 indicate high accuracy (Çorbacıoğlu and Aksel, 2023). Calibration was evaluated using the Hosmer-Lemeshow test (p > 0.05 indicating good fit) and calibration plots, where closer agreement between the curve and the line of identity indicates better calibration (Liang et al., 2024; Zhong and Liu, 2023). We cataloged all independent predictors identified in the final models and calculated their frequencies across studies. This ranking allowed us to identify the predictors most frequently included in existing models. For the meta-analysis, we used R software (version 4.5.1). Of the 12 included studies, 11 reported AUC values eligible for pooling; one was excluded for not doing so. These 11 studies included 13 prediction models, as two studies developed both a conventional regression model and a machine learning model from the same cohort. To avoid statistical dependence from multiple models derived from the same study population, we did not perform a single overall meta-analysis including all 13 models. Instead, we conducted separate meta-analyses stratified by model type: traditional regression models (e.g., logistic regression, nomogram; n = 9 models from 9 studies) and machine learning models (n = 4 models from 4 studies), with each study contributing only one model to each subgroup analysis. We quantified heterogeneity using the I² statistic; a random-effects model was applied when I² > 50%, while a fixed-effect model was used for I² values of 50% or lower. To investigate the sources of heterogeneity, we conducted subgroup analyses. Additionally, a leave-one-out sensitivity analysis was performed to evaluate the robustness of the pooled results. We examined publication bias through funnel plots and statistical tests (Egger's and Begg's tests), considering a p-value > 0.05 as indicative of a low risk of bias. As an exploratory analysis, we also performed an overall meta-analysis that included all 13 models simultaneously, acknowledging the potential statistical dependence. (Supplementary Material Figs. S1–S4). Results Search results From an initial pool of 1,729 records, the study selection process was conducted as follows: 685 duplicates were eliminated, and 1,044 titles and abstracts were screened, resulting in the exclusion of 1,004 records. A full-text review of the remaining 40 articles subsequently led to the exclusion of 28, culminating in a final selection of 12 eligible studies (Chen et al., 2023; Gao et al., 2025; Jiang et al., 2024; Liang et al., 2024; Ma et al., 2025; Qiao et al., 2025; Shen et al., 2024; Wang et al., 2022; Zhang et al., 2024; Zhang et al., 2025; Zhang et al., 2023; Zhong and Liu, 2023). The final analysis included 14 prediction models derived from 12 studies. Among them, two studies each developed two distinct models, namely logistic regression with random forest (Zhang et al., 2025) and logistic regression with an artificial neural network (Zhang et al., 2023). A flowchart of the study selection process is provided in Fig. 1.Fig. 1Flowchart of literature search for prediction models for frailty in Elderly Stroke Patients.Fig 1 dummy alt text Flowchart of literature search for prediction models for frailty in Elderly Stroke Patients. Characteristics of included studies This review analyzed 12 studies, all published between 2022 and 2025, which together developed 14 prediction models for frailty in older adults in China. The study designs consisted of cross-sectional (6 models, 50.0%), retrospective (3 models, 25.0%), and prospective cohort (3 models, 25.0%) designs. All models were developed in single-center hospital settings. The sample sizes for model development varied from 98 to 904 participants. Within these cohorts, the proportion of females ranged from 28.85% to 54.94%, while the reported prevalence of frailty varied from 15.3% to 64.1%. The most commonly used instrument for assessing frailty was the Fried Frailty Phenotype (5 models), followed by the FRAIL Scale (3 models), the Tilburg Frailty Indicator (2 models), the Comprehensive Geriatric Assessment–Frailty Index (1 model), and the Elderly Frailty Assessment Scale (1 model). A summary of these study characteristics is provided in Table 1.Table 1Characteristics of included studies.Table 1 dummy alt textAuthor (year)CountryStudy designData source (collection period)SettingsParticipantsOutcome measuresFrailty cases/sample size (%)Candidate predictorsNo. of model(Shen et al., 2024)ChinaCross-sectional studyOne class A tertiary hospital (Mar–Dec 2023)HospitalOlder ischemic stroke patients, Aged ≥60 years,mean age 70.6 years, 43.9% femaleTFI328/567 (57.88%)42(sociodemographic,behavioral, lifestyle,laboratory, clinical, functional,anthropometric)1(Chen et al., 2023)ChinaRetrospective studyOne class A tertiary hospital (Jan 2020–Dec 2022)HospitalElderly AIS patients , Aged ≥60 years,mean age 74.44years,28.85%FFP57/156 (36.5%)13 (demographics, lifestyle,clinical, functional,anthropometric)1Female(Gao et al., 2025)ChinaCross-sectional studyOne class A tertiary hospital (Dec 2021–Jul 2022)HospitalElderly ischemic stroke patients, aged ≥60 years,mean age 68.01years, 32.63% femaleFRAIL Scale196/331 (59.21%)21 (demographics, clinical,functional,psychological social,anthropometric)1(Jiang et al., 2024)ChinaCross-sectional studyOne class A tertiary hospital (Apr 2021–Mar 2022)HospitalElderly ischemic stroke patients, aged ≥60 years,mean age 73.74 (estimated) years,43.29%femaleFRAIL Scale109/231 (47.19%)22 (demographic, clinical,functional, laboratory, lifestyle,psychological social,anthropometric)1(Liang et al., 2024)ChinaCross-sectional studyOne class A tertiary hospital(Mar–Nov 2022)HospitalOlder cerebral infarction patients, aged ≥60 years, mean age 73.68 years, 40.9%FRAIL Scale28/110 (25.45%)17 (demographic, clinical,psychological social,anthropometric.)1Female(Ma et al., 2025)ChinaCross-sectional studyOne class A tertiary hospital(Jan–Jun 2022)HospitalOlder patients with mild ischemic stroke, aged ≥60 years,mean age 72.4 (estimated) years,54.97%FFP61/322 (18.94%)17 (demographic, lifestyle,behavioral, clinical,psychological social,anthropometric)1female(Qiao et al., 2025)Chinaprospective cohortOne class A tertiary hospital(Aug 2022–Jul 2023)HospitalOlder AIS patients, aged ≥60 years, mean age 70.5(estimated) years, 50.35%%FFP375/904 (41.5%)16 (demographic,medical, lifestyle,behavioral, functional, psychological social,anthropometric)1female(Wang et al., 2022)ChinaCross-sectional studyOne tertiary hospital, (Jan 2019–May 2020)Hospitalcerebral infarction patients, aged ≥65 years,mean age 70.65 (estimated) years, 46.94%femaleCGA-FI15/98 (15.3%)16(demographic, medical, lifestyle,behavioral, clinical)1(Zhang et al., 2024)ChinaProspective studyone tertiary hospital (July 2022 – March 2023)HospitalOlder ischemic stroke patients, aged ≥ 60 years,mean age 66 years,37.4% femaleTFI134/441 (30.4%)19(sociodemographic,laboratory, psychological social,clinical,anthropometric)1(Zhang et al., 2025)Chinaretrospective studyone secondary grade A hospital(Jan 2021 – Jan 2023)HospitalOlder stroke patients, aged ≥60 years, median age 70.42 years,44.15%FFP98/167 (58.68%)20(demographic, laboratory, psychological social,clinical,functional,anthropometric)2Female(Zhang et al., 2023)ChinaProspective studyone tertiary hospital (Mar 2021 – May 2022)HospitalOlder stroke patients, aged ≥60 years, mean age 74(estimated) years,42.29% femaleEFAS249/532 (46.8%)18 (demographic, lifestyle,behavioral, clinical,functional)2(Zhong and Liu, 2023)Chinaretrospective studyone tertiary hospital( Feb 2020 - Mar 2022)hospitalOlder ACI and dysphagia patients , aged ≥60 years; mean age 73.6 years,43.08%FFP125/195 (64.1%)18(demographic, psychological social,clinical,functional,lifestyle)1female(estimated of the modeling group)AIS, Acute Ischemic Stroke;TFI, Tilburg Frailty Indicator ;FFP, Fried Frailty Phenot ype;PSM, Propensity Score Matching;CGA-FI, Comprehensive Geriatric Assessment–Frailty Index;EFAS, Elderly Frailty Assessment Scale;ACI,Acute Cerebral Infarction.Note: Candidate predictors are summarized by category in this table. For the complete list of final predictors for each included study, please refer to Table 3. Table 1Characteristics of included studies.Table 1 dummy alt textAuthor (year)CountryStudy designData source (collection period)SettingsParticipantsOutcome measuresFrailty cases/sample size (%)Candidate predictorsNo. of model(Shen et al., 2024)ChinaCross-sectional studyOne class A tertiary hospital (Mar–Dec 2023)HospitalOlder ischemic stroke patients, Aged ≥60 years,mean age 70.6 years, 43.9% femaleTFI328/567 (57.88%)42(sociodemographic,behavioral, lifestyle,laboratory, clinical, functional,anthropometric)1(Chen et al., 2023)ChinaRetrospective studyOne class A tertiary hospital (Jan 2020–Dec 2022)HospitalElderly AIS patients , Aged ≥60 years,mean age 74.44years,28.85%FFP57/156 (36.5%)13 (demographics, lifestyle,clinical, functional,anthropometric)1Female(Gao et al., 2025)ChinaCross-sectional studyOne class A tertiary hospital (Dec 2021–Jul 2022)HospitalElderly ischemic stroke patients, aged ≥60 years,mean age 68.01years, 32.63% femaleFRAIL Scale196/331 (59.21%)21 (demographics, clinical,functional,psychological social,anthropometric)1(Jiang et al., 2024)ChinaCross-sectional studyOne class A tertiary hospital (Apr 2021–Mar 2022)HospitalElderly ischemic stroke patients, aged ≥60 years,mean age 73.74 (estimated) years,43.29%femaleFRAIL Scale109/231 (47.19%)22 (demographic, clinical,functional, laboratory, lifestyle,psychological social,anthropometric)1(Liang et al., 2024)ChinaCross-sectional studyOne class A tertiary hospital(Mar–Nov 2022)HospitalOlder cerebral infarction patients, aged ≥60 years, mean age 73.68 years, 40.9%FRAIL Scale28/110 (25.45%)17 (demographic, clinical,psychological social,anthropometric.)1Female(Ma et al., 2025)ChinaCross-sectional studyOne class A tertiary hospital(Jan–Jun 2022)HospitalOlder patients with mild ischemic stroke, aged ≥60 years,mean age 72.4 (estimated) years,54.97%FFP61/322 (18.94%)17 (demographic, lifestyle,behavioral, clinical,psychological social,anthropometric)1female(Qiao et al., 2025)Chinaprospective cohortOne class A tertiary hospital(Aug 2022–Jul 2023)HospitalOlder AIS patients, aged ≥60 years, mean age 70.5(estimated) years, 50.35%%FFP375/904 (41.5%)16 (demographic,medical, lifestyle,behavioral, functional, psychological social,anthropometric)1female(Wang et al., 2022)ChinaCross-sectional studyOne tertiary hospital, (Jan 2019–May 2020)Hospitalcerebral infarction patients, aged ≥65 years,mean age 70.65 (estimated) years, 46.94%femaleCGA-FI15/98 (15.3%)16(demographic, medical, lifestyle,behavioral, clinical)1(Zhang et al., 2024)ChinaProspective studyone tertiary hospital (July 2022 – March 2023)HospitalOlder ischemic stroke patients, aged ≥ 60 years,mean age 66 years,37.4% femaleTFI134/441 (30.4%)19(sociodemographic,laboratory, psychological social,clinical,anthropometric)1(Zhang et al., 2025)Chinaretrospective studyone secondary grade A hospital(Jan 2021 – Jan 2023)HospitalOlder stroke patients, aged ≥60 years, median age 70.42 years,44.15%FFP98/167 (58.68%)20(demographic, laboratory, psychological social,clinical,functional,anthropometric)2Female(Zhang et al., 2023)ChinaProspective studyone tertiary hospital (Mar 2021 – May 2022)HospitalOlder stroke patients, aged ≥60 years, mean age 74(estimated) years,42.29% femaleEFAS249/532 (46.8%)18 (demographic, lifestyle,behavioral, clinical,functional)2(Zhong and Liu, 2023)Chinaretrospective studyone tertiary hospital( Feb 2020 - Mar 2022)hospitalOlder ACI and dysphagia patients , aged ≥60 years; mean age 73.6 years,43.08%FFP125/195 (64.1%)18(demographic, psychological social,clinical,functional,lifestyle)1female(estimated of the modeling group)AIS, Acute Ischemic Stroke;TFI, Tilburg Frailty Indicator ;FFP, Fried Frailty Phenot ype;PSM, Propensity Score Matching;CGA-FI, Comprehensive Geriatric Assessment–Frailty Index;EFAS, Elderly Frailty Assessment Scale;ACI,Acute Cerebral Infarction.Note: Candidate predictors are summarized by category in this table. For the complete list of final predictors for each included study, please refer to Table 3. Characteristics of included studies. AIS, Acute Ischemic Stroke;TFI, Tilburg Frailty Indicator ;FFP, Fried Frailty Phenot ype;PSM, Propensity Score Matching;CGA-FI, Comprehensive Geriatric Assessment–Frailty Index;EFAS, Elderly Frailty Assessment Scale;ACI,Acute Cerebral Infarction. Note: Candidate predictors are summarized by category in this table. For the complete list of final predictors for each included study, please refer to Table 3. Model development The detailed characteristics of the included risk prediction models are summarized in Table 2. The candidate predictors were multi-dimensional, encompassing sociodemographic, clinical, lifestyle, anthropometric, laboratory, and social psychological domains. Across the studies, the number of candidate variables considered ranged from 13 to 42. In terms of data preprocessing, only one study used median imputation to address missing data (Shen et al., 2024). Final predictor selection methods varied by modeling technique: Machine learning was used for feature selection; traditional regression models typically applied univariable screening followed by multivariable stepwise selection; nomograms derived final predictors from their underlying regression models.Table 2Development and Validation Characteristics of the Frailty Prediction Models.Table 2 dummy alt textAuthor (year)Missing dataContinuous variablesVariables selectionModeling methodModel validationDiscrimination(AUC/C-index)CalibrationOther indexesModel presentation(Shen et al., 2024)Median imputation (<30%) ,deletion (≥30%)continuous categoricalLASSO ,RF-RFENomogram for PredictingBootstrap internal validation and temporally external validationAUC(train)= 0.904, AUC(internal)=0.915, AUC(external)= 0.920H-L test: P = 0.210 (train), P = 0.213 (internal); Brier scores(train)=0.108, Brier scores(internal)= 0.119; Calibration curve:good agreementSe(internal) = 84.1%, Sp(internal) = 84.8%;DCA(train): net benefit( 0.02–0.96),DCA(internal): (0.03–0.93),nomogram(Chen et al., 2023)Not reportedcontinuous categoricalunivariable LRmultivariate LRinternal validation (simple cross validation)AUC(train )= 0.84, AUC(internal) = 0.676H-L test: P(train) = 0.072;--(Gao et al., 2025)Not reportedcategoricalunivariable LRNomogram for PredictingBootstrap internal validationAUC(internal) = 0.937Calibration curve : good agreementSe = 80.1%, Sp = 96.3%;DCA:higher net benefitnomogram(Jiang et al., 2024)Not reportedcategorical continuousunivariable LRmultivariate LR-AUC = 0.780H-L test:P > 0.05Se = 80.7%, Sp = 68.0%;Youden's index=0.487equation(Liang et al., 2024)Not reportedcategorical continuousunivariable LRmultivariate LR-C-index= 0.8253;AUC = 0.825H-L test: P = 0.754 > 0.05; calibration curve:good agreementSe = 67.9%, Sp = 86.6%;Youden's index=0.545nomogram(Ma et al., 2025)Not reportedcategorical continuousunivariable analysisRandom ForestBootstrap internal validationAUC(train )= 0.795, AUC(internal) = 0.816-Se = 88.24༅, Sp = 88.89༅;F1 = 0.933-(Qiao et al., 2025)Not reportedcategoricalunivariable LRBayesian Network----Bayesian graph + conditional probability table(Wang et al., 2022)excludedcategorical continuousunivariable LRmultivariate LR-AUC = 0.796---(Zhang et al., 2024)excludedcategorical continuousunivariable LRCARTinternal cross-validation(5-fold cross-validation)AUC(train) = 0.94-Se(train)= 74%Tree diagrams, classification rules and prediction tablesAUC(internal) =0.92Sp(train)= 94%Se(internal)= 76%Sp(internal)= 93%(Zhang et al., 2025)Not reportedcategorical continuousunivariable LRmultivariate LR, RFinternal cross-validation(5-fold cross-validation)RF:AUC = 0.728, LR:AUC =0.629--model formula(Zhang et al., 2023)Not reportedcategoricalunivariable LRmultivariate LR, ANNInternal validationLR:AUC(train) = 0.908 ,AUC(internal) =0.923 ; ANN:AUC(train) = 0.904 ,AUC(internal) =0.916-LR:Se = 82.02%, Sp = 83.04%, ANN:Se = 84.54%, Sp = 83.08%LR:nomogram , model formula(Zhong and Liu, 2023)Not reportedcategorical continuousunivariable LRmultivariate LRInternal validationAUC(train)=0.853, AUC(internal)=0.844calibration curve:good agreementDCA:higher net benefit; Se(train)= 78. 31%nomogramSp(train)= 85. 11%Se(internal)= 76. 19%Sp(internal)= 82. 61%CART,classification and regression tree;LR, logistic regression;LASSO, least absolute shrinkage and selection operator; RF-RFE, Random Forest, Recursive Feature Elimination;VIF,Variance Inflation Factor;RF, random forest;ANN, Neural network;BN,Bayesian Network; Table 2Development and Validation Characteristics of the Frailty Prediction Models.Table 2 dummy alt textAuthor (year)Missing dataContinuous variablesVariables selectionModeling methodModel validationDiscrimination(AUC/C-index)CalibrationOther indexesModel presentation(Shen et al., 2024)Median imputation (<30%) ,deletion (≥30%)continuous categoricalLASSO ,RF-RFENomogram for PredictingBootstrap internal validation and temporally external validationAUC(train)= 0.904, AUC(internal)=0.915, AUC(external)= 0.920H-L test: P = 0.210 (train), P = 0.213 (internal); Brier scores(train)=0.108, Brier scores(internal)= 0.119; Calibration curve:good agreementSe(internal) = 84.1%, Sp(internal) = 84.8%;DCA(train): net benefit( 0.02–0.96),DCA(internal): (0.03–0.93),nomogram(Chen et al., 2023)Not reportedcontinuous categoricalunivariable LRmultivariate LRinternal validation (simple cross validation)AUC(train )= 0.84, AUC(internal) = 0.676H-L test: P(train) = 0.072;--(Gao et al., 2025)Not reportedcategoricalunivariable LRNomogram for PredictingBootstrap internal validationAUC(internal) = 0.937Calibration curve : good agreementSe = 80.1%, Sp = 96.3%;DCA:higher net benefitnomogram(Jiang et al., 2024)Not reportedcategorical continuousunivariable LRmultivariate LR-AUC = 0.780H-L test:P > 0.05Se = 80.7%, Sp = 68.0%;Youden's index=0.487equation(Liang et al., 2024)Not reportedcategorical continuousunivariable LRmultivariate LR-C-index= 0.8253;AUC = 0.825H-L test: P = 0.754 > 0.05; calibration curve:good agreementSe = 67.9%, Sp = 86.6%;Youden's index=0.545nomogram(Ma et al., 2025)Not reportedcategorical continuousunivariable analysisRandom ForestBootstrap internal validationAUC(train )= 0.795, AUC(internal) = 0.816-Se = 88.24༅, Sp = 88.89༅;F1 = 0.933-(Qiao et al., 2025)Not reportedcategoricalunivariable LRBayesian Network----Bayesian graph + conditional probability table(Wang et al., 2022)excludedcategorical continuousunivariable LRmultivariate LR-AUC = 0.796---(Zhang et al., 2024)excludedcategorical continuousunivariable LRCARTinternal cross-validation(5-fold cross-validation)AUC(train) = 0.94-Se(train)= 74%Tree diagrams, classification rules and prediction tablesAUC(internal) =0.92Sp(train)= 94%Se(internal)= 76%Sp(internal)= 93%(Zhang et al., 2025)Not reportedcategorical continuousunivariable LRmultivariate LR, RFinternal cross-validation(5-fold cross-validation)RF:AUC = 0.728, LR:AUC =0.629--model formula(Zhang et al., 2023)Not reportedcategoricalunivariable LRmultivariate LR, ANNInternal validationLR:AUC(train) = 0.908 ,AUC(internal) =0.923 ; ANN:AUC(train) = 0.904 ,AUC(internal) =0.916-LR:Se = 82.02%, Sp = 83.04%, ANN:Se = 84.54%, Sp = 83.08%LR:nomogram , model formula(Zhong and Liu, 2023)Not reportedcategorical continuousunivariable LRmultivariate LRInternal validationAUC(train)=0.853, AUC(internal)=0.844calibration curve:good agreementDCA:higher net benefit; Se(train)= 78. 31%nomogramSp(train)= 85. 11%Se(internal)= 76. 19%Sp(internal)= 82. 61%CART,classification and regression tree;LR, logistic regression;LASSO, least absolute shrinkage and selection operator; RF-RFE, Random Forest, Recursive Feature Elimination;VIF,Variance Inflation Factor;RF, random forest;ANN, Neural network;BN,Bayesian Network; Development and Validation Characteristics of the Frailty Prediction Models. CART,classification and regression tree;LR, logistic regression;LASSO, least absolute shrinkage and selection operator; RF-RFE, Random Forest, Recursive Feature Elimination;VIF,Variance Inflation Factor;RF, random forest;ANN, Neural network;BN,Bayesian Network; The methods for final predictor selection varied across the models. Seven models employed multivariate logistic regression (Chen et al., 2023; Jiang et al., 2024; Liang et al., 2024; Wang et al., 2022; Zhang et al., 2025; Zhang et al., 2023; Zhong and Liu, 2023). Nomograms were applied in two studies (Gao et al., 2025; Shen et al., 2024), while four studies (Ma et al., 2025; Qiao et al., 2025; Zhang et al., 2024; Zhang et al., 2023) utilized machine learning algorithms to generate and compare models. The final model sizes ranged from 2 to 12 predictors, encompassing a total of 68 distinct variables. Age was the most prevalent predictor, appearing in 7 models (50.0%), followed by neurological function in 6 models (42.9%). Physical exercise and nutritional status each featured in 4 models (28.6%). Additionally, several common predictors were present in three models, including history of falls, dysphagia, multimorbidity, depression, and economic level. Seven predictors, such as polypharmacy, albumin level, and cognitive function, were included in two models, while 12 predictors were unique to a single model. The distribution of these predictors is illustrated in Fig. 2, and a complete listing can be found in Table 3.Fig. 2Frequency of Predictors Across the Included Models.Fig 2 dummy alt textTable 3List of Final Predictors in the Frailty Prediction Models.Table 3 dummy alt textAuthor (year)No. of predictorsFinal predictors(Shen et al., 2024)9marital status, smoking, history of falls(in the preceding year),physical exercise, polypharmacy, albumin level, ADL(BI), dysphagia(WST), cognition function(MMSE)(Chen et al., 2023)4age, diabetes mellitus, massive cerebral infarction, dysphagia(Gao et al., 2025)6recurrent stroke, motor dysfunction, multimorbidity, nutritional status(MNA-SF), neurological function(mRS), and ADL(Jiang et al., 2024)6age, physical exercise, nutritional status(NRS-2002), ADL(BI), neurological function(NIHSS), aging perceptions(B-APQ )(Liang et al., 2024)2age, anxiety(SAS)(Ma et al., 2025)5multimorbidity, Sleep status, living quality(SF-36), ADL, monthly personal income(Qiao et al., 2025)4ADL, physical exercise, nutritional status(PNI), hypertension(Wang et al., 2022)2age, history of falls(Zhang et al., 2024)5neurological function(NIHSS),depression(GDS), grip strength, general self-efficacy, social support(SSRS)(Zhang et al., 2025)7LR and RF: neurological function(NIHSS), cognition function(MoCA),age,Neuroelectrophysiological assessment(TMS cortical silent period,qEEG delta wave power),albumin level,Depression(GDS-15)(Zhang et al., 2023)6LR and ANN: age, sleep disturbance, IADL, history of falls, living alone, physical exercise.(Zhong and Liu, 2023)12age, course of dysphagia, living alone, self-financing, economic level, multimorbidity, Number of medications, motor disorder, depression(HAMD-17), malnutrition(MNA-SF), social support(MSPSS),dietary guidanceADL, activities of daily living;BI, Barthel Index;WST,The water-swallowing test;MMSE,The Chinese version of the Mini- Mental State Examination;mRS,modified rankin scale ;MNA-SF,mini-nutritional assessment short-form;NIHSS,National Institutes of Health Stroke Scale;NRS-2002,Nutritional Risk Screening 2002;B-APQ ,Brief Aging Perceptions Questionnaire;SAS,Self-Rating Anxiety Scale;SF-36,Short Form 36-Item;PNI,Prognostic Nutritional Index;GDS, Geriatric Depression Scale;SSRS,Social Support Rating Scale;MoCA,Montreal Cognitive Assessment;TMS,Transcranial Magnetic Stimulation;qEEG,Quantitative Electroencephalography;GDS-15, Geriatric Depression Scale-15;IADL,impaired instrumental activities of daily living ;LR, logistic regression;RF, random forest;ANN, Neural network;CART,classification and regression tree;BN,Bayesian Network;HAMD-17,Hamilton Depression Rating Scale-17 items;MSPSS,Multidimensional Scale of Perceived Social Support;Hcy,Homocysteine;BADL,Basic Activities of Daily Living Frequency of Predictors Across the Included Models. Table 3List of Final Predictors in the Frailty Prediction Models.Table 3 dummy alt textAuthor (year)No. of predictorsFinal predictors(Shen et al., 2024)9marital status, smoking, history of falls(in the preceding year),physical exercise, polypharmacy, albumin level, ADL(BI), dysphagia(WST), cognition function(MMSE)(Chen et al., 2023)4age, diabetes mellitus, massive cerebral infarction, dysphagia(Gao et al., 2025)6recurrent stroke, motor dysfunction, multimorbidity, nutritional status(MNA-SF), neurological function(mRS), and ADL(Jiang et al., 2024)6age, physical exercise, nutritional status(NRS-2002), ADL(BI), neurological function(NIHSS), aging perceptions(B-APQ )(Liang et al., 2024)2age, anxiety(SAS)(Ma et al., 2025)5multimorbidity, Sleep status, living quality(SF-36), ADL, monthly personal income(Qiao et al., 2025)4ADL, physical exercise, nutritional status(PNI), hypertension(Wang et al., 2022)2age, history of falls(Zhang et al., 2024)5neurological function(NIHSS),depression(GDS), grip strength, general self-efficacy, social support(SSRS)(Zhang et al., 2025)7LR and RF: neurological function(NIHSS), cognition function(MoCA),age,Neuroelectrophysiological assessment(TMS cortical silent period,qEEG delta wave power),albumin level,Depression(GDS-15)(Zhang et al., 2023)6LR and ANN: age, sleep disturbance, IADL, history of falls, living alone, physical exercise.(Zhong and Liu, 2023)12age, course of dysphagia, living alone, self-financing, economic level, multimorbidity, Number of medications, motor disorder, depression(HAMD-17), malnutrition(MNA-SF), social support(MSPSS),dietary guidanceADL, activities of daily living;BI, Barthel Index;WST,The water-swallowing test;MMSE,The Chinese version of the Mini- Mental State Examination;mRS,modified rankin scale ;MNA-SF,mini-nutritional assessment short-form;NIHSS,National Institutes of Health Stroke Scale;NRS-2002,Nutritional Risk Screening 2002;B-APQ ,Brief Aging Perceptions Questionnaire;SAS,Self-Rating Anxiety Scale;SF-36,Short Form 36-Item;PNI,Prognostic Nutritional Index;GDS, Geriatric Depression Scale;SSRS,Social Support Rating Scale;MoCA,Montreal Cognitive Assessment;TMS,Transcranial Magnetic Stimulation;qEEG,Quantitative Electroencephalography;GDS-15, Geriatric Depression Scale-15;IADL,impaired instrumental activities of daily living ;LR, logistic regression;RF, random forest;ANN, Neural network;CART,classification and regression tree;BN,Bayesian Network;HAMD-17,Hamilton Depression Rating Scale-17 items;MSPSS,Multidimensional Scale of Perceived Social Support;Hcy,Homocysteine;BADL,Basic Activities of Daily Living List of Final Predictors in the Frailty Prediction Models. ADL, activities of daily living;BI, Barthel Index;WST,The water-swallowing test;MMSE,The Chinese version of the Mini- Mental State Examination;mRS,modified rankin scale ;MNA-SF,mini-nutritional assessment short-form;NIHSS,National Institutes of Health Stroke Scale;NRS-2002,Nutritional Risk Screening 2002;B-APQ ,Brief Aging Perceptions Questionnaire;SAS,Self-Rating Anxiety Scale;SF-36,Short Form 36-Item;PNI,Prognostic Nutritional Index;GDS, Geriatric Depression Scale;SSRS,Social Support Rating Scale;MoCA,Montreal Cognitive Assessment;TMS,Transcranial Magnetic Stimulation;qEEG,Quantitative Electroencephalography;GDS-15, Geriatric Depression Scale-15;IADL,impaired instrumental activities of daily living ;LR, logistic regression;RF, random forest;ANN, Neural network;CART,classification and regression tree;BN,Bayesian Network;HAMD-17,Hamilton Depression Rating Scale-17 items;MSPSS,Multidimensional Scale of Perceived Social Support;Hcy,Homocysteine;BADL,Basic Activities of Daily Living Model validation The validation characteristics of the included models are presented in Table 2. Among the 12 studies, only one (Shen et al., 2024) performed both internal and external validation. Internal validation was reported for 7 models (58.3%), whereas 4 models (33.3%) lacked validation after development. Among the 7 models that reported internal validation, bootstrap resampling was the most common method (3 models) (Gao et al., 2025; Ma et al., 2025; Shen et al., 2024), followed by five-fold cross-validation (2 models) (Zhang et al., 2024; Zhang et al., 2025) and simple cross-validation (1 model) (Chen et al., 2023). The single external validation instance (Shen et al., 2024) employed temporal validation. Model performance As summarized in Table 2, all models except one (Qiao et al., 2025) reported discrimination of model, assessed primarily by the AUC (13 models, 92.9%) or c-index (1 model, 7.14%) (Liang et al., 2024). The reported AUC or c-index values for model development ranged from 0.629 (Zhang et al., 2025) to 0.915 (Shen et al., 2024). All internally validated models maintained good discriminative ability, with AUC values ranging between 0.676 (Chen et al., 2023) and 0.937 (Gao et al., 2025). Calibration was reported in 6 models (42.9%). During model development, Two models (Chen et al., 2023; Jiang et al., 2024) demonstrated acceptable calibration per the Hosmer-Lemeshow test (P > 0.05), though no further interpretation was provided. For validation, three models employed calibration curves, which suggested favorable calibration (Gao et al., 2025; Shen et al., 2024; Zhong and Liu, 2023). The Hosmer-Lemeshow test and Brier score were additionally used (Shen et al., 2024). Eight models reported other indexes, including sensitivity, specificity, Youden's index, and the F1 score. Three models (Gao et al., 2025; Shen et al., 2024; Zhong and Liu, 2023) also performed decision curve analysis, demonstrating good potential for clinical application. Model presentation Among the 14 models, 9 (64.3%) provided details on their presentation format. The formats varied across the studies, with nomograms being the most common; they were used in five of the models (Gao et al., 2025; Liang et al., 2024; Shen et al., 2024; Zhang et al., 2023; Zhong and Liu, 2023). One model each was presented as a Bayesian graph with conditional probability table (Qiao et al., 2025), tree diagrams with classification rules and prediction tables (Zhang et al., 2024), and a nomogram with model formula (Zhang et al., 2023). The remaining models were presented as an equation (Jiang et al., 2024) and a model formula (Zhang et al., 2025), respectively. Summary of risk of bias and applicability Table 4 summarizes the risk of bias and applicability assessments for the included models. All models were evaluated as having an overall high risk of bias. In the participants domain, nine models were rated as high risk because they relied on data from cross-sectional or retrospective studies (Chen et al., 2023; Gao et al., 2025; Jiang et al., 2024; Liang et al., 2024; Ma et al., 2025; Shen et al., 2024; Wang et al., 2022; Zhang et al., 2025; Zhong and Liu, 2023). Additionally, one model was rated as high risk in this domain solely because it excluded patients with "pre-stroke frailty" (Zhang et al., 2024). In the Predictors domain, eight models had an unclear risk of bias (due to unreported blinding of predictor assessment) (Gao et al., 2025; Jiang et al., 2024; Liang et al., 2024; Ma et al., 2025; Qiao et al., 2025; Shen et al., 2024; Wang et al., 2022; Zhang et al., 2023), two had a high risk (due to clinically unavailable predictors, e.g., quantitative electroencephalography delta power and nutritional guidance requiring extra assessment) (Zhang et al., 2025; Zhong and Liu, 2023), and two had a low risk. In the outcome domain, one model was assessed as high risk because it defined the outcome as a combination of frailty and pre-frailty (Liang et al., 2024). Six models were deemed high risk due to their inability to exclude predictors from the outcome definition (Gao et al., 2025; Liang et al., 2024; Ma et al., 2025; Shen et al., 2024; Wang et al., 2022; Zhang et al., 2024). Three models demonstrated an unclear risk of bias attributable to inadequate reporting regarding assessor training and inter-rater reliability (Jiang et al., 2024; Ma et al., 2025; Shen et al., 2024). No models provided information on blinded outcome assessment, leading to an unclear risk for this item. Additionally, eleven models showed an unclear risk in the analysis domain because of insufficient descriptions of the data collection processes and timelines for both predictors and outcomes (Gao et al., 2025; Jiang et al., 2024; Liang et al., 2024; Ma et al., 2025; Shen et al., 2024; Wang et al., 2022; Zhang et al., 2025; Zhang et al., 2023; Zhong and Liu, 2023).Table 4PROBAST Assessment of Risk of Bias and Applicability.Table 4 dummy alt textAuthor (year)No. of modelRisk of biasApplicabilityOverallParticipantsPredictorsOutcomeAnalysisParticipantsPredictorsOutcomeRisk of biasApplicability(Shen et al., 2024)1-?--+++-+(Chen et al., 2023)1-+?--++--(Gao et al., 2025)1-?--+++-+(Jiang et al., 2024)1-?--+++-+(Liang et al., 2024)1-?---+---(Ma et al., 2025)1-?--+++-+(Qiao et al., 2025)1+??-+++-+(Wang et al., 2022)1-?--+++-+(Zhang et al., 2024)1-+---++--(Zhang et al., 2025)2--?---+--(Zhang et al., 2023)2+??-+++-+(Zhong and Liu, 2023)1--?--++-- Table 4PROBAST Assessment of Risk of Bias and Applicability.Table 4 dummy alt textAuthor (year)No. of modelRisk of biasApplicabilityOverallParticipantsPredictorsOutcomeAnalysisParticipantsPredictorsOutcomeRisk of biasApplicability(Shen et al., 2024)1-?--+++-+(Chen et al., 2023)1-+?--++--(Gao et al., 2025)1-?--+++-+(Jiang et al., 2024)1-?--+++-+(Liang et al., 2024)1-?---+---(Ma et al., 2025)1-?--+++-+(Qiao et al., 2025)1+??-+++-+(Wang et al., 2022)1-?--+++-+(Zhang et al., 2024)1-+---++--(Zhang et al., 2025)2--?---+--(Zhang et al., 2023)2+??-+++-+(Zhong and Liu, 2023)1--?--++-- PROBAST Assessment of Risk of Bias and Applicability. Within the analysis domain, all models exhibited a high risk of bias. Additionally, ten studies were deemed to have insufficient sample sizes, as they had an events-per-variable ratio of less than 10 (Chen et al., 2023; Gao et al., 2025; Jiang et al., 2024; Liang et al., 2024; Ma et al., 2025; Shen et al., 2024; Wang et al., 2022; Zhang et al., 2024; Zhang et al., 2025; Zhong and Liu, 2023).Only two models (Gao et al., 2025; Qiao et al., 2025) employed exclusively categorical predictors. Among the remaining 12 models that included continuous predictors, six categorized these predictors without providing clear justification or established cut-off standards (Chen et al., 2023; Ma et al., 2025; Qiao et al., 2025; Zhang et al., 2023; Zhong and Liu, 2023).Four models did not include all enrolled participants in the final analysis (Wang et al., 2022; Zhang et al., 2024; Zhang et al., 2023), and one explicitly excluded cases with missing data (Shen et al., 2024). Eleven models did not provide any information concerning the management of missing data (Gao et al., 2025; Jiang et al., 2024; Liang et al., 2024; Ma et al., 2025; Qiao et al., 2025; Wang et al., 2022; Zhang et al., 2024; Zhang et al., 2025; Zhang et al., 2023). One model (Qiao et al., 2025) did not assess discrimination, and six models lacked any form of calibration assessment (Chen et al., 2023; Gao et al., 2025; Jiang et al., 2024; Liang et al., 2024; Shen et al., 2024; Zhong and Liu, 2023). No models reported on the management of data complexity. Seven models exhibited a low risk concerning applicability, whereas five models presented a high risk. Within the participants' domain, five models were identified as having a high applicability risk. Reasons included the exclusion of individuals who already had frailty at baseline (Chen et al., 2023; Zhang et al., 2024; Zhong and Liu, 2023), active exclusion of all non-frail patients (Liang et al., 2024), exclusion of patients with a history of prior stroke (Zhang et al., 2025), and a cohort with a notably low proportion of females (28.85%) (Chen et al., 2023). In the predictors domain, one model (Zhang et al., 2025) was deemed to have high applicability risk because its predictive variables lacked generalizability. In the outcome domain, one model (Liang et al., 2024) was rated as high risk due to its inclusion of pre-frailty within the frailty group. Meta-analysis results To investigate potential sources of heterogeneity, a subgroup analysis was conducted based on the core methodology of model construction. The results indicated that the modeling approach significantly contributed to performance heterogeneity. Traditional Statistical Models Subgroup (Group 1): This subgroup consisted of models developed using conventional regression techniques (e.g., logistic regression or nomogram) . The pooled AUC was 0.82 (95% CI: 0.74–0.90, P < 0.0001), demonstrating that these traditional approaches can achieve good discriminatory performance. Nevertheless, substantial heterogeneity was observed within this subgroup (I² = 93.1%), suggesting significant performance variation even among models based on classical statistical methods. This variation may be attributed to differences in variable selection, population characteristics, or specific modeling details (Fig. 3, Fig. 4).Fig. 3Risk of bias included studies.Fig 3 dummy alt textFig. 4Forest Plot for the Subgroup Meta-Analysis of Traditional Statistical Models.Fig 4 dummy alt text Risk of bias included studies. Forest Plot for the Subgroup Meta-Analysis of Traditional Statistical Models. Machine Learning Models Subgroup (Group 2): This subgroup included models constructed using machine learning algorithms. The pooled AUC was 0.83 (95% CI: 0.73–0.94, P < 0.05), comparable to that of the traditional statistical models. Although the heterogeneity in this subgroup remained substantial (I² = 64.2%), it was significantly lower than that of the traditional statistical models subgroup. This suggests that machine learning models, as a category, may demonstrate more consistent performance (Fig. 5).Fig. 5Forest Plot for the Subgroup Meta-Analysis of Machine-Learning Models.Fig 5 dummy alt text Forest Plot for the Subgroup Meta-Analysis of Machine-Learning Models. A total of 13 prediction models were included in this meta-analysis. The comprehensive evaluation demonstrated that these models exhibited good overall discriminatory ability, with a pooled AUC of 0.82 (95% CI: 0.76 - 0.88). However, the heterogeneity analysis revealed substantial inconsistency among the models (I² = 92.0%, P < 0.0001), highlighting considerable variation in predictive performance across the included studies. As a result, a random-effects model was employed for the meta-analysis (Fig. S1). Leave-one-out analysis showed pooled AUC 0.81–0.85 (95% CI >0.80), with no dominant study. One study had the largest heterogeneity impact: exclusion reduced I² from 92% to 86.7%. The stability of the AUC estimate supports the robustness of the main findings. (Zhang et al., 2025). (Fig. S2). Publication bias was assessed through visual inspection of a funnel plot and statistical tests. The funnel plot, presented in Fig. S3, displayed some asymmetry, indicating a possible presence of publication bias. This suggests that studies with poorer predictive performance may be less likely to be published. This visual indication was corroborated by the results of Egger's linear regression test (t = -3.440, P = 0.006). In contrast, Begg's rank correlation test did not reveal significant bias (z = 0.000, P = 1.000). Given that Egger's test is generally regarded as having higher statistical power for detecting bias in such contexts, the possibility of publication bias cannot be dismissed. Discussion This systematic review and meta-analysis identified 12 studies introduced 14 frailty prediction models for older adults with stroke. We evaluated their performance, methodological quality, and applicability to map the current research landscape and identify shortcomings to guide future research. Core predictors and their geriatric significance A precise frailty risk prediction model is crucial for early intervention in elderly stroke patients (Kong et al., 2025). This review identified five common geriatric predictors: age, neurological function, activities of daily living, physical exercise, and nutritional status. Importantly, the frequency of inclusion does not equate to predictive strength but rather reflects factors commonly considered in this field. Age was the most frequently included predictor in the models (50% of models). The relationship between advanced age and frailty is explained by inflammaging (Franck et al., 2025), vascular aging, and immunosenescence, which exacerbate brain damage and poor outcomes (Chen et al., 2023). Accelerated biological age (PhenoAge) is associated with poorer short- and long-term outcomes after ischemic stroke (Wang et al., 2025). Among elderly patients aged ≥80 years, outcomes include poorer functional status, higher mortality, and lower quality of life (Bala et al., 2023). Therefore, age should be a fundamental component of frailty risk assessment, integrated into clinical warning systems to facilitate early identification of high‑risk individuals. Neurological deficit, typically assessed by National Institutes of Health Stroke Scale or the modified Rankin Scale, is a key predictor. Frail patients have significantly higher admission scores on the National Institutes of Health Stroke Scale (Pilotto et al., 2022). Neuroelectrophysiological indicators (e.g., transcranial magnetic stimulation cortical silent period) provide novel insights into the neurophysiological mechanisms linking stroke to frailty, such as enhanced cortical inhibition and impaired neural network function (Zhang et al., 2025). Mechanistically, neurological deficits exacerbate frailty primarily by limiting physical activity, leading to sarcopenia and impaired cardiopulmonary function (Hassan et al., 2021). Decline in activities of daily living reflects functional status and is linked to poor post-stroke recovery (Heshmatollah et al., 2021). Neurological deficits and decline in activities of daily living create a vicious cycle of inactivity, muscle loss, and accelerated frailty. Physical exercise is a modifiable protective factor via muscle protein synthesis and neuromuscular coordination (Akbar et al., 2022). Four studies confirmed that regular exercise reduces frailty risk (Talar et al., 2021), helping to delay sarcopenia and frailty in both early and late stages. Nutritional status is also crucial. Low albumin and trace element deficiencies impede muscle synthesis (Wunderle et al., 2024). Dysphagia increases malnutrition risk (Karisik et al., 2025). One study found albumin an independent predictor of post‑stroke frailty (OR=0.89) (Zhang et al., 2025). Nutritional risk (Nutritional Risk Screening 2002) is also an independent factor (Jiang et al., 2024). Importantly, these five predictors do not operate independently but interact in complex ways. For example, nutritional interventions may be most effective when combined with structured exercise (Sirikul et al., 2024), and optimal timing may depend on baseline neurological function and activities of daily living status (De Keyser et al., 2025). Therefore, future research should investigate these interactions through well-designed prospective studies to develop more targeted intervention strategies, moving beyond simply listing variables toward examining their independent and interactive effects. Pervasive methodological limitations undermining validity Research into frailty risk prediction models for elderly stroke patients has intensified in recent years, with all identified studies having been published in China within the past four years. The Prediction model Risk Of Bias ASsessment Tool evaluation indicated a high risk of bias across all 14 models, primarily due to methodological limitations in study design. Despite the proliferation of models, none meet the methodological standards required for reliable clinical application, limiting the utility of any specific model’s predictions regardless of reported AUC values. In terms of participant selection, only three studies employed a prospective design, while the remaining studies utilized cross-sectional or retrospective approaches (Qiao et al., 2025; Zhang et al., 2024; Zhang et al., 2023). Such designs cannot establish temporal relationships between predictors and frailty outcomes, which is essential for true prediction. Therefore, future studies must prioritise prospective cohorts, even if more resource-intensive. Two studies were rated as high risk in the predictor domain. One incorporated neuroelectrophysiological indicators (e.g., Quantitative Electroencephalography Delta Power, transcranial magnetic stimulation cortical silent period) (Zhang et al., 2025), which require expensive equipment and expert operators, limiting clinical availability (Varone et al., 2025). The other used “nutritional guidance” as a predictor (Zhang et al., 2023), which was given after baseline data collection – effectively an intervention rather than a baseline characteristic. Such a predictor may not be available during external validation, potentially limiting model applicability. Additionally, dietary data reliant on recall are vulnerable to memory and social desirability bias (Bajunaid et al., 2025). As a result, the reliability of such data as predictive factors is questionable, fundamentally undermining the model's predictive performance. Conceptual overlap within the outcome domain further compromises validity. Examples include combining frailty and pre-frailty into a composite endpoint (Liang et al., 2024), or using predictors that are components of the frailty assessment tool (e.g., multimorbidity in the FRAIL scale), which leads to circular reasoning that inflates performance (Gao et al., 2025; Liang et al., 2024). Existing studies confirm the close relationship between these factors and frailty (Chen and Liang, 2025). Therefore, it is crucial to maintain a strict separation between predictor and outcome definitions to ensure authentic clinical predictive value. Researchers must carefully examine frailty assessment tools to ensure that candidate predictors are not already embedded within the outcome definition, and sensitivity analyses should explore whether removing such predictors affects model performance. Methodological flaws were also identified in the analytical domain of the included prediction models. A significant issue was the inadequate handling of continuous variables. Among the 12 models that included continuous predictors, six dichotomized them—for example, categorizing age as under or over 60 years—without providing a clear justification for the cut-off points. This practice results in a loss of information, decreased statistical power, and potential bias. According to the British Medical Journal methodology guidelines, dichotomizing continuous variables significantly undermines model performance (Lopez-Ayala et al., 2025). Future studies should adopt flexible approaches, such as restricted cubic splines, to model potential nonlinear relationships between continuous predictors and outcomes. Another common limitation was incomplete data analysis. Four models did not include all participants, one eliminated records with missing data, and eleven models failed to report any methods for handling missing data. These approaches can introduce selection bias and compromise the generalizability of the models. Multiple imputation is the recommended method for addressing missing data in prediction model studies (Nijman et al., 2022). Future work should proactively address missing data by clearly documenting its patterns, extent, and handling procedures. Variable selection methods exhibited significant limitations. Many models relied solely on univariate screening, which does not account for correlations between predictors and may retain variables that are spuriously associated (Jin and Kattan, 2023). Furthermore, even advanced methods like Least Absolute Shrinkage and Selection Operator regression can select irrelevant variables, particularly with limited samples or correlated predictors (Lo et al., 2015). Therefore, future research should utilize multivariable methodologies—such as regularization techniques or pre-specified models grounded in clinical rationale—to enhance robustness. Model evaluation was notably insufficient. Among the models analyzed, one failed to report discrimination measures, six lacked calibration assessments, and all neglected to account for data clustering structures. More concerningly, only one study referenced external validation without providing results, while all others entirely omitted this crucial validation. Discrimination and calibration are essential components of predictive performance. Discrimination differentiates between high-risk and low-risk individuals, while calibration ensures that predicted probabilities align with observed outcomes. Inadequate calibration results in erroneous risk estimates that compromise clinical utility, even when discrimination appears adequate (de Oliveira et al., 2025). Furthermore, without external validation, model generalizability to new populations remains unverified (Mansmann and Ön, 2025). Models developed in single-center settings may reflect site‑specific factors that do not generalize. External validation in diverse, multicenter cohorts is essential to assess transportability and identify sources of performance decline. Moreover, the lack of calibration assessment is particularly problematic because poorly calibrated models can systematically overestimate or underestimate risk, potentially leading to inappropriate clinical decisions. Future research should prioritize external validation of existing models over developing new ones, and should report calibration alongside discrimination. Meta-analysis results: performance and heterogeneity This meta-analysis revealed substantial heterogeneity among frailty prediction models (I² = 92%), which is indicative of methodological diversity across the included studies. It is important to clarify that this pooled AUC merely describes the range of reported discriminative performance across different models and does not represent a meta-analytic validation of any specific model, given varying predictors and populations. The observed heterogeneity arises from variations in study design, population, frailty definitions, predictor selection, and modeling techniques. These factors are interrelated and difficult to disentangle in an aggregate-data meta-analysis. Subgroup analysis revealed differences in performance between modeling approaches. While both traditional statistical models and machine learning models demonstrated similar discriminative performance (AUC 0.82 vs 0.83), the machine learning models displayed significantly lower heterogeneity (I² = 64.2% vs 93.1%). Heterogeneity decreased substantially from 93.1% in traditional models to 64.2% in machine learning models. This provides direct evidence that the modeling approach contributes to the observed performance variation. Machine learning models demonstrated more consistent performance across studies, suggesting that algorithmic choices influence performance variability across populations and settings. However, substantial heterogeneity persists within both subgroups, indicating that modeling approach alone does not fully explain the variation—other factors, such as population characteristics, frailty definitions, and predictor selection, continue to play important roles. Machine learning algorithms may achieve more consistent performance across different datasets, possibly due to their built-in regularization and ability to capture complex variable relationships (Zhao et al., 2025). In contrast, traditional models demonstrate greater susceptibility to subjective decisions regarding variable selection and model specification, which subsequently contributes to increased variability. Nevertheless, this interpretation must be tempered by the small number of machine learning models in our analysis, and future studies should further explore whether machine learning approaches consistently outperform traditional methods across diverse settings. Moreover, substantial heterogeneity persists within both subgroups, indicating that modeling approach alone does not fully explain the observed variation. This finding highlights limitations of subgroup analysis based on a single study-level characteristic, underscoring the need for more standardized approaches to model development and reporting to enable meaningful comparisons. Sensitivity analysis confirmed the stability of our pooled estimates, as removing any single study did not substantially alter the overall conclusions (Sappani et al., 2025). Egger's test suggested potential publication bias in the available literature. These findings provide complementary rather than contradictory evidence. The sensitivity analysis shows that the published studies we analyzed are robust; however, the presence of publication bias suggests that our results may overestimate true performance due to the absence of studies with poorer outcomes. Substantial heterogeneity was observed, due to models developed in different populations with different predictor sets. Furthermore, methodological limitations common in the primary studies — including non-prospective designs, inadequate sample sizes, and suboptimal handling of predictors (e.g., continuous variables categorised or missing data poorly addressed) — also contribute to the uncertainty of the pooled estimates. Therefore, the pooled AUC must be interpreted with caution. While it offers a broad benchmark of current model performance, it does not support recommending any specific model for clinical practice. Future research should focus on two priorities: externally validating existing models in independent populations and, when possible, comparing them directly on the same datasets to determine relative effectiveness. Strengths and limitations This study’s key strengths include a comprehensive methodological appraisal using The Prediction model Risk Of Bias ASsessment Tool, identification of consistent shortcomings, synthesis of frequently included predictors, and a quantitative summary of model performance despite heterogeneity. Limitations stem from the quality and scope of primary evidence: all models were from single-center Chinese studies, requiring broader validation; predictor frequency does not indicate predictive strength; and substantial unexplained heterogeneity persists, likely due to unreported differences in population or frailty definitions. Implications for practice, policy, and future research No current model is ready for routine clinical use due to high risk of bias and lack of external validation. The five core predictors (age, neurological function, activities of daily living, exercise, and nutrition) should nonetheless be routinely assessed as part of comprehensive geriatric evaluation to guide early supportive interventions. Assessment of these factors may help clinicians identify patients at higher risk of poor outcomes and guide early supportive interventions while more robust and externally validated models are being developed. Research funding bodies should prioritise methodological quality over model proliferation, requiring The Prediction model Risk Of Bias ASsessment Tool-guided risk of bias assessment. Multi-center prospective aging cohort studies should be encouraged through dedicated funding streams. Such investments are necessary to generate the high-quality evidence needed for clinically actionable prediction tools. External validation of existing models must take precedence over developing new ones, and prospective cohort designs are essential to establish temporal relationships. Methodological rigor requires adequate sample sizes with pre-specified calculations, proper handling of continuous variables (e.g., restricted cubic splines), multiple imputation for missing data, and transparent variable selection. Machine learning approaches warrant further investigation but need direct comparison on the same datasets. Interactions among core predictors should be examined, and conceptual overlap (predictors as part of outcome definition) must be avoided. These recommendations are derived directly from this review and provide specific, actionable directions to advance the field toward clinically useful prediction tools for post-stroke frailty. Conclusion This systematic review reveals a significant disconnect between the predictive potential and methodological robustness of frailty prediction models for older adults with stroke. The pooled discriminative ability is modest, yet the consistently high risk of bias undermines clinical applicability. Rather than endorsing any specific model, this review maps the current landscape, identifies key methodological shortcomings (e.g., non-prospective designs, inadequate sample sizes), and provides direction for future research. To address these issues, the field must prioritize methodological rigor over model proliferation. Future studies should comply with The Prediction model Risk Of Bias ASsessment Tool and Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis guidelines, and focus on: (1) prospective, multi-center aging cohort designs; (2) adequate sample sizes with pre-specified calculations; (3) proper handling of continuous variables and missing data; and (4) robust internal and external validation. CRediT authorship contribution statement Lifang Huang: Writing – original draft, Methodology, Funding acquisition, Formal analysis, Conceptualization. Juan Zhang: Resources, Formal analysis, Data curation, Conceptualization. Shanjun Mao: Resources, Formal analysis, Data curation, Conceptualization. Rulei Ding: Methodology, Formal analysis, Data curation, Conceptualization. Lan Luo: Formal analysis, Data curation, Conceptualization. Guilin Shen: Formal analysis, Data curation. Kun Sun: Writing – review & editing, Software, Methodology, Conceptualization. Declaration of competing interest The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Lifang Huang reports financial support was provided by Hunan Provincial Natural Science Foundation of China. Lifang Huang reports financial support was provided by Hunan Nursing Association. Lifang Huang reports a relationship with Hunan Provincial Natural Science Foundation of China that includes: funding grants. Lifang Huang reports a relationship with Hunan Nursing Association that includes: funding grants. If there are other authors, they declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.