Machine Learning-Based Frailty Prediction and Classification in Community-Dwelling Older Adults: A Systematic Review of Validation, Explainability, and Implementation Readiness

Le résumé fourni par la source

Background and Objectives: Frailty is a multidimensional vulnerability in older adults; the Fried phenotype and Frailty Index are clinically informative but labor-intensive, limiting scalability for community screening. Machine learning (ML) can model heterogeneous, high-dimensional data, but real-world adoption is constrained by heterogeneity in definitions, predictors, validation strategies, and explainability. We systematically synthesized ML-based studies of frailty prediction and classification in community-dwelling older adults, examining validation rigor, explainability, and implementation readiness. Methods: This systematic review followed PRISMA 2020 and was registered in PROSPERO (CRD420251081555). PubMed, Embase, Web of Science, and Scopus were searched on 4 July 2025, with a supplementary IEEE Xplore and ACM Digital Library search conducted on 12 May 2026. Eligible studies included community-dwelling adults aged ≥60 years, ML-based frailty prediction or classification, sample ≥ 1000, and publication in a peer-reviewed journal indexed in the Web of Science Core Collection; hospital-based studies were excluded. Risk of bias and reporting quality were assessed with PROBAST (Prediction Model Risk of Bias Assessment Tool) and TRIPOD (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis); implementation readiness was assessed with the RE-AIM (Reach, Effectiveness, Adoption, Implementation, Maintenance) framework and a Technology Readiness Level (TRL)-style rubric. Findings were synthesized narratively. Results: Fourteen studies (development cohorts 1230–86,133 participants) were included; the supplementary IEEE/ACM search identified 42 records but yielded no additional eligible studies. Classification of current frailty status (n = 7) yielded AUROCs (area under the receiver operating characteristic curve) of 0.70–0.98, with the highest values likely reflecting partial label overlap with frailty components; incident prediction (n = 6) yielded internal AUROCs of 0.70–0.81 and same-cohort temporal AUROCs of 0.58–0.85; independent external validation was uncommon. Only 2 of 14 studies had both low overall risk of bias and low applicability concern (PROBAST); the field is concentrated at TRL 4–6, with no study at TRL 7 or higher and none documenting Implementation or Maintenance domains of RE-AIM. Conclusions: ML-based frailty models show heterogeneous discrimination and limited readiness for routine community use. Priorities include standardized task-type-specific definitions, independent external validation, calibration and decision-curve reporting, transparent predictor disclosure, and prospective implementation evaluation.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.