Abstract Body (Do not enter title and authors here): Background: Although frailty affects clinical outcomes and complicates management in elderly patients with heart failure (HF), its evaluation remains inconsistent due to the lack of objective and accessible assessment tools. We sought to develop a machine learning-based automatic rating of the clinical frailty scale (CFS) in elderly patients with HF.
Methods: We prospectively examined 839 elderly (≥75 years) patients with symptomatic chronic HF (mean age 81.7±6.7 years; left ventricular ejection fraction 56 [interquartile range [IQR] 42-65]%) from 26 Japanese sites between January 2019 and November 2024. Patients were allocated to the derivation (n = 523, 10 sites) or validation (n = 316, 16 sites) cohort according to the institution at enrollment. CFS assessments were determined through a modified Delphi method involving 10 independent trained cardiologists as the gold standard. We used ViTPose, a vision transformer–based pose estimation model, to analyze videos of patients standing up and walking along a 4-meter L-shaped path, and extracted 128 gait parameters. We developed a predictive model for the CFS using the Light Gradient Boosting Machine (LightGBM) algorithm. Hyperparameters were optimized via nested cross-validation with Optuna in the derivation set. To evaluate the performance of this model, we calculated accuracy, balanced accuracy, and Cohen's weighted kappa (CWK) between the predicted and actual CFSs. We also computed Shapley Additive Explanations (SHAP) values to identify key predictive parameters.
Results: The LightGBM model demonstrated excellent agreement rate in both the derivation (accuracy 0.84; balanced accuracy 0.83; CWK 0.86) and validation (accuracy 0.79; balanced accuracy 0.75; CWK 0.85) cohorts. No cases were misclassified by more than two CFS level (Figures 1 and 2). In the SHAP analysis, peak gait speed (mean absolute SHAP value: 0.127), total gait time (mean absolute SHAP value: 0.095), and swing phase time (mean absolute SHAP value: 0.049) were identified as the important predictors of CFS (Figure 3).
Conclusions: We developed a machine learning model to provide objective and reliable CFS assessment for elderly patients with HF, potentially standardizing frailty evaluation in clinical practice.