We conducted a comprehensive search of multiple databases, including PubMed, Web of Science, Embase, Cochrane Library, Scopus, SinoMed, China National Knowledge Infrastructure (CNKI), Wanfang Database, China Science and Technology Journal Database (VIP), Cumulative Index to Nursing and Allied Health Literature (CINAHL). The search was conducted from the time of database construction to June 10, 2025. The objective was to identify studies that used machine learning methods to predict the onset of frailty in community-dwelling older adults. The search strategy used a combination of keywords and subject terms such as “machine learning,” “artificial intelligence,” “predictive models,” “risk prediction modelling,” “frailty,” “aged”, “community”, etc. The detailed search strategy is shown in Supplementary Table 1, and the reference lists of the included literature were carefully reviewed to ensure comprehensiveness.

For systematic evaluation, we used the PICOTS system recommended by the CHARMS (Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modeling Studies) list (25). The system helps frame the purpose of the review, the search strategy, and the inclusion and exclusion criteria for the studies. The key items of PICOTS are described as Table 1.

Based on the aforementioned research framework, we established the following inclusion and exclusion criteria:

Inclusion criteria: (1) Population: Community-dwelling older adults aged ≥60 years; (2) Predictive Model Type: Studies must develop or validate at least one machine learning-based predictive model, or compare machine learning algorithms with other machine learning algorithms, traditional statistical analyses, or clinical scoring tools, and must report at least one model performance metric (discrimination and/or calibration); (3) Study design: cohort studies, cross-sectional studies, and case–control studies; (4) Study outcome: predicted outcome of frailty (comprehensive physical frailty), with frailty defined based on recognized standards or tools.

Exclusion criteria: (1) Studies not reporting the predictive model development process; (2) Studies limited to cognitive frailty or social frailty; (3) Literature not in English or Chinese; (4) Studies where full text was unavailable; (5) Case reports, reviews, conference abstracts, or guidelines; (6) Studies employing only traditional statistical models without machine learning algorithms, or studies conducted in non-community settings such as hospitals, clinics, or long-term care facilities; (7) Studies failing to report model performance metrics.

First, all literature retrieved from various databases was imported into EndNote X9 software for management, with duplicate records automatically removed. Subsequently, two independent researchers conducted an initial screening of the remaining literature’s titles and abstracts to exclude obviously irrelevant studies. Finally, the same two researchers reviewed the full texts of the initially screened articles against predefined inclusion and exclusion criteria to determine the final list of included studies. Any disagreements were resolved through discussion or adjudication by a third researcher. Additionally, we carefully examined the reference lists of all eligible studies to ensure the completeness of the included literature.

The search results were screened independently by two authors according to the CHARMS list. In case of disagreement, it was decided through discussion or by the third author. The information extracted from the selected articles included: (1) Basic information of the study: including information on the data source, country, year of publication, authors, sample size, outcome, outcome assessment tool, and candidate predictors. (2) Model information: including the extraction of information such as predictor screening methods, model construction methods, missing data handling methods, continuous variable handling methods, evaluation model indicators, and predictors finally included in the model.

Quality, potential risk of bias, and study applicability of the included literature were assessed using the PROBAST+AI evaluation tool (22), a newly developed assessment tool for evaluating the quality, risk of bias, and applicability of regression- or artificial intelligence-based predictive models. It includes four domains: participants and data sources, predictors, outcomes, and analyses; a total of 34 signaling questions (16 for model development and 18 for model evaluation) and 6 applicability items (3 each for model development and model evaluation) were addressed. Signaling questions are answered as yes, probably yes, no, probably no, no information, or, when appropriate, not applicable. Quality concerns (for model development) are judged as low, high, or unclear, and risk of bias (for model evaluation) is judged as low, high, or unclear.

First, we summarized the AUC values and their 95% confidence intervals (95% CI) for each study. Due to the inconsistent time windows for frailty prediction across the included studies—three studies assessed baseline prediction, six studies assessed longitudinal prediction, and one study assessed both baseline and longitudinal prediction—we conducted meta-analyses based on the prediction window (baseline vs. longitudinal) and stratified analyses based on the model construction method. For studies that did not report the standard error (SE) or 95% CI of AUC, the Hanley and McNeil formula was used to estimate the SE and 95% CI (26, 27). Considering that the included studies may have heterogeneity among study designs, ML models, and predictors. Therefore, the AUC of each study was combined in the meta-analysis using the Der Simonian and Laird random effects model (28). Heterogeneity between studies was assessed using the Cochrane Q-test and the I² statistic (I² ≥ 50% or p ≤ 0.10) to determine whether to use a fixed-effects model or a random-effects model (29). All statistical analyses were performed using Stata18.0 software and results were considered statistically significant at p < 0.05.

Through systematic searching, we retrieved 8,210 articles from 10 databases. We also identified 4 studies by reviewing reference lists. After deduplication using EndNote X9 literature management software, 4,836 articles remained. Initial screening based on titles and abstracts led to the exclusion of 4,812 articles, leaving 24 articles for full-text eligibility assessment. After reviewing the full texts, we excluded 14 articles (reasons for exclusion detailed in Figure 1). Ultimately, 10 articles were included in this study (30–39). Figure 1 shows the Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) 2020 flowchart depicting the comprehensive search process and results.

Ten studies were included in this study with publication dates of 2022–2025, all of which were in English. Most studies were cohort-based, including six prospective cohort studies, one retrospective cohort study, one cross-sectional cohort study, and two cross-sectional studies. Data for seven studies originated from single-center databases, while the remaining three studies utilized data collected by researchers from community sources. Among studies developing models using databases, data were sourced from China (n = 5), Japan (n = 1), and South Korea (n = 1). Study sample sizes ranged from 214 to 14,540 participants, with 28–4,883 outcome events recorded. Prevalence rates varied between 5.33% and 64.77%.

Candidate predictors numbered between 16 and 56, most commonly comprising demographic characteristics, clinical laboratory data, and lifestyle factors. Regarding frailty assessment, the most frequently used tools were the Fried Phenotype (n = 4) and Frailty Index (n = 3), with others including Kihon (n = 1), Study of Osteoporotic Fractures (SOF) Index (n = 1), and Tilburg Frailty Indicator (TFI) (n = 1). Among these, five studies employed a binary classification of “frail” and “non-frail” to define outcomes, while the remaining five studies included a “pre-frail” state in their outcome definitions. The detailed characteristics of the included models are shown in Table 2.

Table 3 summarizes detailed information on model development characteristics. In terms of variable treatment, four studies converted continuous variables to categorical variables (32, 35, 37, 38). In terms of missing data treatment, four studies used multiple imputation (31, 34, 37, 39), two studies excluded missing data (33, 35), three studies did not report information about missing data (30, 32, 38), and one study used the miss Forest method for missing data (36). Among the methods for predictor screening, commonly used approaches include Lasso regression (37, 38), random forest method (33, 36), and univariable analysis (31, 35). The number of predictors included in the final model ranged from 5 to 16, covering a total of 61 predictors. The most common predictor was age (n = 8), followed by cognitive function (n = 5) and depression (n = 3).

Internal validation was performed in all 10 included studies, with three studies performing both internal and external validation (30, 32, 34) and two studies performing two external validations. The methods used in internal validation were random split (n = 7), bootstrap (n = 2), and 10-Fold Cross-Validation (n = 1). The three studies that conducted external validation used the time validation method. Specific information about modeling is shown in Table 3.

Table 3 summarizes detailed information on model performance. A total of 45 models were constructed using 16 methods across the 10 studies. The most used methods were Random Forest (RF, n = 8), Logistic Regression (LR, n = 7), Extreme Gradient Boosting (XGB, n = 7), and Support vector machine (n = 5). In the included studies, the performance of the models was assessed mainly through differentiation and calibration methods. Nine of the studies measured model discrimination by AUC, and one study did not report model discrimination and used the F1 score to measure model performance.

Five studies reported model calibration methods. In the model construction phase, two studies assessed calibration using the Brier score and their constructed models all scored <0.25, and 2 studies assessed calibration using the Hosmer-Lemeshow (H-L) test, with the majority of the models demonstrating good calibration performance (p > 0.05). In internal validation, studies using the H-L test and Brier score also showed good calibration performance (p > 0.05, Brier score < 0.25). In external validation, two studies assessed predictive model calibration using the Brier score (n = 2) and H-L test (n = 1), respectively.

Six studies reported on the presentation form of the model. Of these, one study presented the model in Nomogram and Online clinical supports, and another presented it in Nomogram and Prognostic index, respectively. The models of the remaining four studies were presented as graphs (n = 2), websites (n = 1), and column line graphs (n = 1), respectively.

Among the 10 included studies, two employed the SHAP method to quantify feature contributions, two conducted interpretable model inference based on Bayesian networks, four presented primary predictors through feature importance ranking, one visualized predictor weights using a column chart, and one study utilized both feature importance ranking and column charts. To further summarize interpretive findings across models, we compiled the top five key predictors reported in the included studies. Results indicated that age (n = 7), cognitive function (n = 5), and ADL (n = 3) were the most frequently occurring predictors. Supplementary Table 3 summarizes the interpretability analysis methods used in the study and the top five key predictors identified by each method.

Two personnel used the PROBAST+AI assessment tool to evaluate the quality and applicability of the model development process and the risk of bias and applicability of the model performance assessment. After careful evaluation, in the model development and assessment section, seven studies had low-quality issues and high risk of bias concerns (31, 33, 34, 36–39) and the quality of the remaining three studies was rated as unclear (30, 32, 34). The results of applicability in the model development and assessment sections were the same, with low applicability of models in four studies (31, 33–35), high applicability of models in five studies (30, 32, 37–39), and unclear applicability of models in one study (36). The final results of bias and applicability of the included models are shown in Table 4. Bar charts showing the PROBAST+AI quality and risk of bias assessment results included in the study are presented in Figures 2, 3.

In the model development and assessment section, the reasons for being rated as low quality and high bias were mainly attributed to (1) retrospective cohort studies raising low-quality concerns and high risk of bias among participants and data sources; (2) not reporting whether the measurements of the predictors were affected by outcomes that had already occurred; (3) the number of events per variable (EVP) was less than 20; (4) categorization of the continuous predictors; (5) missing data were directly excluded or information on missing data was not reported; (6) there was no information on methods to address category imbalance; and (7) there may be concerns about data leakage. Detailed assessment information is provided in Supplementary Table 4.

A total of 45 models were developed in the 10 included studies: 36 models were developed for internal validation, and nine models were developed for external validation. Since only three studies performed external validation and one study evaluated the performance of the model with the F1 score, the meta-analysis only included data from the internal validation set of 9 studies. Because most of the included studies used multiple ML models, we first summarized the AUC for each study, as shown in Supplementary Table 5. Given the variation in prediction time points across studies, we further conducted a stratified meta-analysis based on prediction windows (baseline prediction vs. longitudinal prediction). Results showed that the pooled AUC for baseline frailty prediction studies was 0.878 (95% CI 0.799, 0.958; I² = 94.9%; heterogeneity p < 0.001), while the pooled AUC for longitudinal frailty prediction studies was 0.730 (0.670, 0.790; I² = 96.6%; heterogeneity p < 0.001). The overall pooled AUC for all studies, regardless of prediction time point, was 0.786 (0.697, 0.875; I² = 99.4%; heterogeneity p < 0.001). These findings indicate that frailty prediction models based on baseline data demonstrate superior discriminatory performance, with a significantly higher pooled AUC compared to longitudinal prediction studies. The results of the meta-analysis are shown in Figure 4.

Based on the different algorithms of model, we categorize them into traditional regression models (n = 4) as well as ML models (n = 28). Among them, the traditional regression models include LR, and the ML models include RF, SVM, XGB, etc. After summarizing the AUC of these models, we found that the combined AUC of the traditional regression model was 0.785 (95% CI 0.615, 0.955; I² = 99.2%; heterogeneity p < 0.001), and the combined AUC of the ML model was 0.784 (95% CI 0.747, 0.822; I² = 99.1%; heterogeneity p < 0.001). The results showed that the combined AUC values of the two models were similar. However, due to the smaller sample size of the traditional regression model, its CI was wider, suggesting higher uncertainty in the results. The results of the meta-analysis are shown in Supplementary Figure 1.

To determine which ML model in the set of 9 internal validations had the better performance ability, we pooled and compared the AUCs of the models that were used three or more times. The results of the combined AUCs of the different ML models were as follows: 0.785 (95% CI 0.615, 0.955; I² = 99.2%; heterogeneity p < 0.001); SVM, 0.711 (95% CI 0.527, 1; I² = 99.4%; heterogeneity p < 0.001); DT, 0.696 (95% CI 0.591, 0.801; I² = 90.2%; heterogeneity p < 0.001); RF, 0.802 (95% CI 0.677, 0.927; I² = 98.8%; heterogeneity p < 0.001); XGB. 0.782 (95% CI 0.668, 0.896; I² = 99.3%; heterogeneity p < 0.001). Among these models, RF had the best performance, while DT had a lower performance. However, these differences were not statistically significant. The results of the meta-analysis are shown in Figure 5.

This study included a total of 9 research projects. Sensitivity analysis indicated that the pooled results demonstrated good robustness. Forest plot results showed that the point estimates for all individual studies, after exclusion, clustered closely around the overall pooled effect size (between 0.72 and 0.74), with high overlap in confidence intervals. The results of the sensitivity analysis are shown in Supplementary Figure 2.

This study revealed through stratified meta-analysis that the performance of frailty prediction models based on baseline data (AUC = 0.878; 95% CI 0.799, 0.958) exceeded that of longitudinal prediction models (AUC = 0.730; 95% CI 0.670, 0.790). This performance difference stems not from inherent model superiority but from the fundamental distinction between the two predictive tasks.

In baseline prediction models, all predictors and frailty status are measured at the same point in time. While this feature helps the model identify individuals who are frail at baseline, it also leads to construct overlap between some predictors and the frailty phenotype itself, thereby introducing endogeneity issues (22). When predictors conceptually overlap with frailty or exhibit reverse causality due to simultaneous measurement, non-independence between predictors and frailty outcomes may occur, leading to overestimation of model performance (40). Consequently, baseline prediction models are better suited for screening and case identification, enabling the rapid detection of individuals who are currently in a frail state.

In contrast, the longitudinal prediction model seeks to forecast future events. During follow-up, subjects may experience competing risk events such as death, other illnesses, or lifestyle changes that influence frailty development. These events introduce random variability that is not accounted for by the model, objectively limiting its performance (41). Consequently, longitudinal predictive models perform below baseline models; yet, their clinical value lies primarily in their risk assessment and primary prevention capabilities. However, longitudinal prediction allows tracking and evaluation over time, yielding more accurate assessments of a model’s performance and effectiveness on a long-term timescale (32, 42). This enables the identification of currently healthy individuals at high risk of developing frailty in the future, facilitating early interventions to prevent or delay the onset of frailty.

In summary, the two models address distinct clinical questions, and their performance differences stem not only from randomness in follow-up processes but also from underlying endogeneity mechanisms within the data structure. Therefore, a model’s value should not be measured solely by AUC scores but interpreted based on its specific design objectives and application scenarios. Future research may utilize variables from non-frailty performance components as predictors, or consider methods such as mediation analysis and joint modeling to generate inputs with greater causal significance. This approach could reduce bias arising from endogeneity, thereby enhancing the conceptual rigor and technical validity of predictive models.

Without distinguishing prediction time points, the pooled AUC across all studies was 0.786 (95% CI 0.697, 0.875). This suggests that machine learning shows promising performance in predicting frailty onset among community-dwelling older adults, although the pooled AUC exhibits substantial heterogeneity. PROBAST+AI results suggest that most current studies raise concerns about low quality and high risk of bias. These issues may lead to potential overestimation of model performance metrics in this meta-analysis, while reducing the robustness and generalizability of the findings. Therefore, future research requires further standardization and improvement in study populations, data sources, predictive factors, and data analysis methodologies.

The studies we included employed various methods to enhance interpretability. Feature importance, the most common approach, enables researchers to identify key predictors influencing the model rapidly (68). However, it struggles to explain complex interactions between features, thereby limiting the depth of model interpretation and clinical applicability (69, 70). In contrast, Bayesian networks, as inherently interpretable models, reveal conditional dependencies between variables. They can identify potential causal relationships based on conditional dependencies and independence among variables (71). However, their construction accuracy relies on prior knowledge and data quality, which limits their practical application (72). Furthermore, SHapley Additive exPlanations (SHAP) stands as one of the most widely adopted explainable AI methods today. Its strength lies in simultaneously providing global feature importance and locally revealing key features and logical relationships underlying prediction outcomes. By quantifying the contribution of each feature to specific predictions, SHAP enhances model transparency and clinical interpretability (73, 74). Future research may develop multi-level interpretability frameworks that simultaneously report global feature importance and local prediction explanations. This approach could enhance model transparency and clinical applicability while maintaining predictive performance.

Based on the above analysis, we synthesized the interpretable findings from the included studies. The final results encompassed seven distinct categories of predictors: demographic data, socioeconomic factors, physical function and fitness indicators, cognitive and mental health, chronic diseases and medical conditions, nutritional and metabolic status, and social support. The heterogeneity of these predictors reflects that frailty arises from the combined effects of multiple influencing factors. Notably, across different explanatory methodologies, age, cognitive function, and activities of daily living (ADL) ability were the most frequently identified predictors. In most studies, age was recognized as a key predictor of frailty. Cognitive decline was also identified as a critical predictor of frailty progression. Research indicates a bidirectional relationship between cognitive impairment and physical functional decline, with both factors jointly accelerating the onset and progression of frailty (75). Furthermore, the combination of cognitive impairment and frailty may increase the risk of adverse health outcomes in older adults. Additionally, ADL is closely associated with frailty development. Previous studies suggest that ADL decline not only represents a clinical manifestation of frailty but may also serve as a mediating factor in its progression (5).

Given that ML is more efficient in dealing with complex nonlinear relationships and can improve prediction accuracy, we compared the performance of the ML model with that of the traditional regression model. Our results showed that the combined AUC point estimates of the traditional regression model and the ML model were similar, 0.785 (95% CI 0.615, 0.955) and 0.784 (95% CI 0.747, 0.822), respectively. However, due to the lack of statistical efficacy of the traditional regression group, which contained only four models at the time of the analysis, it is not yet possible to conclude that there is no difference in the performance of the two types of methods. Such small subgroups cannot reliably support the “no difference” conclusion, and their wide confidence intervals may improve the truthfulness, ranging from significantly worse to better than the ML models (76). Thus, the current results do not negate the theoretical advantages of ML in dealing with complex relationships. However, it suggests that logistic regression models are still competitive in specific scenarios due to their simplicity, stability, and ease of interpretation (77). Future research should fairly compare the advantages and disadvantages of different algorithms on the same high-quality dataset while ensuring methodological rigor.