JMIR Publications

ABSTRACT

Background:

Frailty affects approximately 10–15% of community-dwelling older adults globally and is associated with increased risk of falls, hospitalization, and mortality. Conventional frailty assessment relies on clinician-administered instruments such as the Fried Frailty Phenotype, which are time-intensive and show only moderate inter-rater reliability. Skeleton-based and sensor-based gait analysis—encompassing computer-vision pose estimation, depth-camera skeleton tracking, and inertial measurement unit (IMU)-based motion capture—has emerged as a candidate technology for objective, scalable frailty screening. However, the methodological quality, diagnostic performance, and clinical applicability of these approaches have not been comprehensively evaluated using a structured risk-of-bias framework.

Objective:

This systematic review synthesizes evidence on the diagnostic accuracy, methodological quality, and clinical applicability of technology-based gait analysis systems for frailty prediction in community-dwelling and institutionalized older adults.

Methods:

Following the PRISMA 2020 statement, we searched PubMed/MEDLINE, Embase, IEEE Xplore, Web of Science, and the Cochrane Central Register of Controlled Trials for studies published between January 2018 and April 2026. The lower bound was set to coincide with the publication of two enabling technologies (ST-GCN and YOLOv3, both 2018). Two reviewers independently screened records against predefined PICOS criteria. Risk of bias and applicability concerns were assessed using the QUADAS-2 tool across four domains: Patient Selection, Index Test, Reference Standard, and Flow and Timing. Because of clinical and methodological heterogeneity across the small number of eligible studies, a narrative synthesis was performed rather than quantitative meta-analysis

Results:

Of 371 records retrieved (after de-duplication), 7 studies (total N = 2,226 older adults) met inclusion criteria. Sensing modalities comprised computer-vision skeleton extraction (n = 2), IMU-based motion capture (n = 4), and marker-based 3D motion capture (n = 1). Reported overall classification accuracy ranged from 79% (in a study with dual-dataset external validation) to approximately 97.5% (in a study that excluded the frail class from training due to small sample size). Frail-class sensitivity varied from 22.1% to 95.6% and tracked methodological choices—class imbalance handling, participant- vs sample-level data partitioning, and external validation—rather than algorithmic differences. Only 1 study was judged at low risk of bias across all 4 QUADAS-2 domains; 4 studies were judged at overall high risk of bias and 2 raised some concerns. Convergent biomechanical signatures of frailty included reduced gait speed, reduced ankle plantar flexion and range of motion, reduced knee heel-strike angle, and increased hip toe-off angle.

Conclusions:

Skeleton-based and sensor-based gait analysis show promise as adjunctive tools for frailty screening but do not yet meet the methodological threshold for first-line clinical deployment. The evidence base is limited by reliance on internal validation, small frail-class sample sizes, cross-sectional designs, and absence of cross-population testing. Future research should treat participant-based stratified data partitioning, external dataset validation, and explicit reporting of frail-class sensitivity as minimum methodological standards before clinical translation. Digital health stakeholders deploying such systems in geriatric care should plan for privacy-preserving inference, demographic generalizability testing, and longitudinal validation as prerequisites for routine use. Clinical Trial: CRD420261373600