Moving towards the detection of frailty with biomarkers: A population health study

Abstract

Aging adults experience increased health vulnerability and compromised abilities to cope with stressors, which are the clinical manifestations of frailty. Frailty is complex, and efforts to identify biomarkers to detect frailty and pre-frailty in the clinical setting are rarely reproduced across cohorts. We developed a predictive model incorporating biological and clinical frailty measures to identify robust biomarkers across data sets. Data were from two large cohorts of older adults: "Invecchiare in Chianti (Aging in Chianti, InCHIANTI Study") (n = 1453) from two small towns in Tuscany, Italy, and replicated in the Atherosclerosis Risk in Communities Study (ARIC) (n = 6508) from four U.S. communities. A complex systems approach to biomarker selection with a tree-boosting machine learning (ML) technique for supervised learning analysis was used to examine biomarker population differences across both datasets. Our approach compared predictors with robust, pre-frail, and frail participants and examined the ability to detect frailty status by race. Unique biomarker features identified in the InCHIANTI study allowed us to predict frailty with a model accuracy of 0.72 (95% confidence interval (CI) 0.66-0.80). Replication models in ARIC maintained a model accuracy of 0.64 (95% CI 0.66-0.72). Frail and pre-frail Black participant models maintained a lower model accuracy. The predictive panel of biomarkers identified in this study may improve the ability to detect frailty as a complex aging syndrome in the clinical setting. We propose several concrete next steps to keep research moving toward detecting frailty with biomarker-based detection methods.

Keywords: biomarkers; frailty; machine learning.

FIGURE 1

Study population demographics. (a) Female vs. male participants. (b) Race distribution in the ARIC study. The InCHIANTI study is a White European population demographic. *American Indian or Alaskan Indian and Asian populations were not included in the study due to small sample sizes. (c) Education levels. **The ARIC study was missing education information for 3 individuals classified as Robust and 8 individuals classified as Pre‐frail.

FIGURE 2

Study workflow overview of the predictive machine learning model. (a) The predictive clinical and laboratory biomarkers were extracted in phase 1 (b) training was used to select the model hyperparameters and a test set to evaluate the performance of the final model. k‐fold cross‐validation was applied to each problem's data, extending the holdout method until in phase 2 we achieved model performance for prediction of frailty groups. (c) Findings were replicated in the ARIC cohort to test model accuracy.

FIGURE 3

Bubble plot of the importance and log fold change by phenotypes. The size of the bubble is proportional to the importance level of the feature, the larger the bubble the greater effect the feature has on predicting the phenotype. The log fold change becomes negative when the mean value of the feature decreases and positive when the mean value increases.

FIGURE 4

ROC Curve Pre‐frail and Frail Phenotypes Across Race Models in ARIC. The final model's ability to detect frailty status across races in ARIC for (a) pre‐frail and (b) frail phenotypes.