A scoping review exploring machine learning applications in frailty and cardiovascular disease

Introduction Frailty often coexists with cardiovascular disease (CVD), significantly increasing morbidity and mortality among older adults. Machine learning (ML) methods have the potential to improve frailty prediction, classification, and management. However, the extent of their application to CVD populations living with frailty is unclear. Research Question What ML methods have been used to study frailty in CVD populations? Methods A search strategy was developed for the databases Medline, Embase, CINAHL, IEEE Explore, and SCOPUS. Two reviewers independently screened titles, abstracts, full texts, and extracted data using Covidence. Studies included 1) any measure of frailty; 2) ≥1 ML method; 3) ML analysis involving frailty measures; 4) participants with CVD as a primary study population or separate ML-based frailty analyses for those with comorbid CVD; 5) human data; 6) English-language original research articles published between 2014–2025. Conflicts were resolved through consensus. Results Of 413 records screened, 31 studies met inclusion criteria. Publication years ranged from 2016-2025, with 20 of 31 of studies published in 2023 or later. Mean sample size (SD; IQR) and age across studies were 3122 (7195; 257-1451) and 74.8 (9.4; 66.6-81) years, respectively. Mean percent of female participants were 40.3%. Heart failure was the most common CVD type (18 studies). Thirteen frailty measures were used; the Frailty Phenotype, Frailty Index, and Clinical Frailty Scale were most commonly used (5/31 each). Supervised and unsupervised methods were used in 29 and 4 studies, respectively. The most common algorithms were logistic regression (13), penalized regression (11), and random forest (8). ML was used mainly for prediction (29), with mortality (12), frailty (8), and hospitalization (7) as frequent outcomes. Common internal validation methods were k-fold cross-validation (17) and train-test split (8). Only 2 studies used external validation. ML performance metrics included AUC (20), sensitivity (12), and specificity (8). Sixteen studies included explainability elements. Conclusions The use of ML in frailty and CVD research is expanding but remains focused on traditional ML methods. In addition, heart failure as the primary CVD subtype studied indicates that new ML research should be applied to other CVDs. Greater adoption of advanced algorithms, external validation, and including varied CVD types is needed to enhance clinical utility of these techniques.