Characterization of Models for Identifying Physical and Cognitive Frailty in Older Adults With Diabetes: Systematic...

Physical frailty and cognitive frailty are increasingly recognized as critical geriatric syndromes among older adults with diabetes, contributing to adverse outcomes such as disability, hospitalization, and mortality. Early identification of individuals at high risk is therefore essential for timely prevention and intervention. Although a growing number of prediction models have been developed for this population, evidence regarding their methodological rigor, predictive performance, and generalizability remains fragmented.

This study aims to evaluate and characterize existing models for detecting or predicting physical frailty and cognitive frailty in older adults with diabetes.

PubMed, Embase, Web of Science, China National Knowledge Infrastructure (CNKI), Wanfang, and VIP databases were searched from their inception to December 2025. Retrospective, cross-sectional, and prospective studies that developed or validated models predicting frailty or cognitive frailty in older adults with diabetes were included. The Prediction Model Study Risk Of Bias Assessment Tool (PROBAST) was used to assess risk of bias and applicability. Random effects meta-analyses using the Hartung-Knapp-Sidik-Jonkman method were conducted to synthesize model performance, including the pooled area under the receiver operating characteristic curve (AUC). Heterogeneity was explored through subgroup and sensitivity analyses. Small study effects were evaluated using funnel plots, the Egger test, and the Deeks funnel plot asymmetry test.

A total of 24 studies comprising 32 diagnostic models were included. The overall pooled analysis demonstrated an AUC of 0.851 (95% CI 0.820-0.882) with a 95% prediction interval of 0.710-0.992, sensitivity of 0.810 (95% CI 0.740-0.850), and specificity of 0.850 (95% CI 0.810-0.890). Statistical comparisons in the modeling approach revealed that logistic regression models achieved a significantly higher pooled AUC (0.850) compared with machine learning models (0.785; P=.003). Similarly, retrospective studies demonstrated superior performance, with an AUC of 0.900 compared with 0.843 for cross-sectional studies (P=.03). Conversely, no significant differences were observed across subgroups stratified by data source (P=.42), patient characteristics (P=.77), validation methods (P=.16), or specific outcomes (P=.94). The most common predictors identified were depression, age, and regular exercise; however, all included studies were assessed as having a high risk of bias.

To our knowledge, this review provides the first comprehensive synthesis of models for risk stratification of physical frailty and cognitive frailty in older adults with diabetes. The findings indicate that existing models demonstrate satisfactory discrimination; specifically, CIs confirmed a robust average effect, while prediction intervals suggested that performance in future settings, though variable, is likely to remain acceptable. However, clinical utility is currently constrained by high risk of bias and limited external validation. Future research must prioritize rigorous, prospective, multicenter studies adhering to standard reporting guidelines (eg, TRIPOD [Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis]) to establish valid, generalizable, and clinically actionable prognostic instruments.