Abstract
Background/Objective: Frailty, a geriatric syndrome characterized by decreased reserve and resistance to stressors in older adults, has been established as a robust predictor of health outcomes. Recently, the Neurocognitive Frailty Index (NFI) was introduced, including 42 physical and cognitive elements that collectively assess an individual's vulnerability to age-related health decline. While this multidimensional approach improves predictive accuracy for cognitive decline, its high dimensionality might be a barrier to widespread adoption. Methods: We employed several machine learning techniques to reduce the dimensions of NFI while maintaining its predictive power. We trained five models: Network Analysis, neural networks, Least Absolute Shrinkage and Selection Operator Regression (LASSO), Random Forest, and eXtreme Gradient Boosting (XGBoost). Each model was calibrated using a dataset from the Canadian Study of Health and Aging, which included various cognitive, health, and functional variables. Results: Results indicated that six variables had minimal impact on outcome. This reduction in dimensionality resulted in a modified NFI scale including 36 elements, while maintaining good predictive performance for cognitive change similar to the original NFI. Conclusions: Our findings support the feasibility of applying machine learning techniques to modify predictive models in neurodegenerative diseases beyond frailty assessment. We recommend exploring the application of this scale using other data. The results also emphasize the potential of machine learning approaches for improving predictive models, highlighting their value as a tool for advancing our understanding of aging and its complexities.