Data-driven health deficit assessment improves a frailty index's prediction of current cognitive status and future conversion to dementia: results from ADNI

Abstract

Frailty is a dementia risk factor commonly measured by a frailty index (FI). The standard procedure for creating an FI requires manually selecting health deficit items and lacks criteria for selection optimization. We hypothesized that refining the item selection using data-driven assessment improves sensitivity to cognitive status and future dementia conversion, and compared the predictive value of three FIs: a standard 93-item FI was created after selecting health deficit items according to standard criteria (FI_s) from the ADNI database. A refined FI (FI_r) was calculated by using a subset of items, identified using factor analysis of mixed data (FAMD)-based cluster analysis. We developed both FIs for the ADNI1 cohort (n = 819). We also calculated another standard FI (FI_c) developed by Canevelli and coworkers. Results were validated in an external sample by pooling ADNI2 and ADNI-GO cohorts (n = 815). Cluster analysis yielded two clusters of subjects, which significantly (p_FDR < .05) differed on 26 health items, which were used to compute FI_r. The data-driven subset of items included in FI_r covered a range of systems and included well-known frailty components, e.g., gait alterations and low energy. In prediction analyses, FI_r outperformed FI_s and FI_c in terms of baseline cognition and future dementia conversion in the training and validation cohorts. In conclusion, the data show that data-driven health deficit assessment improves an FI's prediction of current cognitive status and future dementia, and suggest that the standard FI procedure needs to be refined when used for dementia risk assessment purposes.

Keywords: Alzheimer’s disease; Dementia; Frailty; Frailty index; Machine learning; Mild cognitive impairment.

Fig. 1

Data-driven supplement (red boxes) to the standard procedure (green boxes) for creating a frailty index (FI). *For the refined selection of health deficit variables, a false discovery rate (FDR)-adjusted p value < .05 from regression analyses of each variable on cluster belonging is used as selection threshold. †Assess face validity of refined selection against standard criteria and core frailty construct. FI_r (see “Methods” section) was developed using the data-driven supplement to the standard procedure as shown, whereas FI_s and FI_c were developed using standard procedure only

Fig. 2

A Dendrogram showing the hierarchical structure of the subject clustering solution. The blue part shows the first and largest cluster which we coined “fit” due to significantly lower frailty scores in this subgroup (c.fr. Table 3) compared with the second, smaller cluster (yellow), coined “frail”. B Scatterplot showing subjects and their cluster belonging based on the two first principal components (PC, “Dim1”, “Dim2”) from factor analysis of mixed data (FAMD). The percentages denote explained variance of each PC

Fig. 3

A Density plots showing three frailty index (FI) distributions for ADNI1 (development) and ADNI2/GO (validation) cohorts. FI_s = a 93-item FI created according to standard procedure by the authors. FI_r = a 26-item FI created by adding a data-driven supplement to the standard procedure. FI_c = a 40-item FI created according to standard procedure by Canevelli, et al. [26]. B Boxplots illustrating central tendency and variability of the three different FI-variables for cognitively normal (healthy) controls (HC), and subjects living with mild cognitive impairment (MCI) or Alzheimer’s disease dementia (AD). P values are from Wilcoxon rank sum tests comparing diagnostic group differences in FI-scores. C Kaplan–Meier survival curves for sample quartiles calculated for each FI. The survival probabilities indicate the probability of remaining stable MCI at time of follow-up, and vertical lines through each line indicate censoring. D Estimated mean AUC(t) for prediction of AD conversion in subjects with MCI at baseline plotted over 5 years of follow-up for the three rival continuous FI variables in development and validation cohorts