Predictive Modeling for Frailty Conditions in Elderly People: Machine Learning Approaches

Area of Science:

• Gerontology and Computational Health Science
• Application of AI in healthcare for aging populations

Background:

• Frailty is a critical age-related condition in older adults, signifying physiological decline and increased vulnerability.
• Despite numerous studies on frailty detection and mortality association, predicting increased risk in clinical settings remains a challenge.
• A universally agreed operational definition of frailty is still lacking.

Purpose of the Study:

• To develop and compare machine learning (ML) models for predicting frailty conditions in older adults.
• To utilize a comprehensive database of clinical and socioeconomic factors for model development.
• To assess the predictive performance of various ML algorithms for different frailty-related outcomes.

Main Methods:

• Utilized an administrative health database of 1,095,612 elderly individuals (≥65 years) with 58 input and 6 output variables.
• Defined six surrogate problems/outputs for frailty and addressed data imbalance using resampling.
• Compared the performance of Artificial Neural Network (ANN), Genetic Programming (GP), Support Vector Machines (SVM), Random Forest (RF), Logistic Regression (LR), and Decision Tree (DT) algorithms.

Main Results:

• ANN and SVM demonstrated superior performance in predicting mortality (accuracy 0.78-0.79).
• Decision Tree classifiers generally showed the lowest accuracy, while ANN, SVM, GP, LR, and RF performed better.
• Predicting emergency admissions with red code was less accurate than predicting fractures or disability across all models.
• SVM achieved the best performance for predicting urgent hospitalization (accuracy 0.73) using 10-fold cross-validation.

Conclusions:

• Developed ML models capable of predicting various frailty conditions, including mortality, urgent hospitalization, disability, fracture, and emergency admission.
• Model prediction performance varied significantly across different frailty outcomes and evaluation metrics.
• Further refinement of high-performing models can lead to decision-support tools for early identification and prediction of frail older adults.