Development and validation of a physical frailty phenotype index-based model to estimate the frailty index

Abstract

Background: The conventional count-based physical frailty phenotype (PFP) dichotomizes its criterion predictors-an approach that creates information loss and depends on the availability of population-derived cut-points. This study proposes an alternative approach to computing the PFP by developing and validating a model that uses PFP components to predict the frailty index (FI) in community-dwelling older adults, without the need for predictor dichotomization.

Methods: A sample of 998 community-dwelling older adults (mean [SD], 68 [7] years) participated in this prospective cohort study. Participants completed a multi-domain geriatric screen and a physical fitness assessment from which the count-based PFP and the 36-item FI were computed. One-year prospective falls and hospitalization rates were also measured. Bayesian beta regression analysis, allowing for nonlinear effects of the non-dichotomized PFP criterion predictors, was used to develop a model for FI ("model-based PFP"). Approximate leave-one-out (LOO) cross-validation was used to examine model overfitting.

Results: The model-based PFP showed good calibration with the FI, and it had better out-of-sample predictive performance than the count-based PFP (LOO-R², 0.35 vs 0.22). In clinical terms, the improvement in prediction (i) translated to improved classification agreement with the FI (Cohen's k_w, 0.47 vs 0.36) and (ii) resulted primarily in a 23% (95%CI, 18-28%) net increase in FI-defined "prefrail/frail" participants correctly classified. The model-based PFP showed stronger prognostic performance for predicting falls and hospitalization than did the count-based PFP.

Conclusion: The developed model-based PFP predicted FI and clinical outcomes more strongly than did the count-based PFP in community-dwelling older adults. By not requiring predictor cut-points, the model-based PFP potentially facilitates usage and feasibility. Future validation studies should aim to obtain clear evidence on the benefits of this approach.

Keywords: Frailty index; Frailty phenotype; Frailty scales; Geriatrics; Prediction.

Fig. 1

Multivariable associations (black lines or points) of physical frailty phenotype criterion predictors (expressed on their natural scales for interpretability) with Frailty Index. Predicted mean frailty index values were calculated from a Bayesian beta regression model using thin-plate splines for continuous predictors and the monotonic effects approach for ordinal predictors. For all predictors, ribbons are 95% (light blue), 80% (medium blue), and 50% (dark blue) credible intervals

Fig. 2

Visual assessment of model calibration for frailty index (FI). Predicted FI were derived from a model using the count-based physical frailty phenotype (PFP) as the only predictor (left panel) and a model using non-dichotomized PFP criterion predictors (right panel). Solid line represents the identity line. Dotted line represents a lowess smoother through the data points, showing good calibration (linear relation) between observed and predicted FI values for the model-based PFP