Mortality prediction among older people in primary care: A comparison of predictive performance of the frailty index with regression-based and machine learning models

Introduction: The frailty index is widely used to identify vulnerable individuals at risk of adverse outcomes like mortality. However, its predictive performance compared to other mortality prediction models, especially in subpopulations like people with dementia, is not well known. This study aimed to compare frailty index’s performance with regression-based and machine learning models for predicting mortality among community-dwelling older adults, and to test performance in a dementia subgroup.

Methods: We selected 355,958 community-dwelling adults aged 60 years and older from primary care with electronic health records (EHR) linked to mortality registrations. We developed one- and five-year mortality prediction models using a 36-item frailty index and compared discrimination and calibration of a regression model including the frailty index with two types of regression models and two types of machine learning models using single health deficits as predictors. Lastly, we evaluated the models’ performance in 6394 persons with dementia.

Results: The frailty index model showed moderate performance with an AUC-ROC of 0.793 and 0.804 for one- and five-year mortality. The other models, using the single deficits as predictors, reached higher AUC-ROCs up to 0.828 and 0.824, with good calibration. Overall, the models performed worse in the dementia subgroup, with AUC-ROCs between 0.678 and 0.704.

Discussion: Regression-based and machine learning prediction models using single frailty deficits outperform the frailty index in predicting one- and five-year mortality. However, these models can be more complex and less interpretable. We found lower performance for people with dementia, suggesting the models are less applicable in this subpopulation.