Predicting Adherence to Physical Activity in Individuals With Cancer Using Machine Learning and National Health and Nutrition Examination Survey Data

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Background: Physical activity (PA) is a core component of cancer survivorship care, yet adherence to PA guidelines remains low. Correlates of PA adherence reflect a complex interplay of multiple factors, many aligning with domains addressed in physical therapy (PT) but not integrated within a PT-informed analytic framework. Machine learning (ML) approaches may support modeling of complex relationships by accommodating nonlinear interactions. This study aimed to develop and compare ML models to classify PA adherence among adults with cancer using National Health and Nutrition Examination Survey (NHANES) data and a PT-informed feature selection framework. Methods: This retrospective cross-sectional study included 1169 adults with a history of cancer from NHANES 2021-2023. From ~400 variables, PT-informed screening, redundancy reduction, and feature engineering yielded 45 features. Logistic regression, random forest, gradient boosting, XGBoost, and support vector machine (SVM) models were developed and compared. Discrimination was assessed using the area under the receiver operating characteristic curve (AUC). Permutation feature importance was applied to the best-performing model with a reduced-feature sensitivity analysis. Results: The SVM demonstrated the strongest performance (AUC = 0.749). Influential variables included self-rated health, sedentary time, educational attainment, difficulty walking or climbing steps, body mass index, gender, number of rooms in the home, systolic blood pressure, income-to-poverty ratio, and anxiety. A reduced SVM using the top 10 variables showed comparable discrimination (AUC = 0.757). Conclusion: ML models demonstrated moderate discrimination in classifying PA adherence among adults with cancer. Parsimonious models using clinically accessible variables may support rehabilitation workflows by prompting targeted screening and timely intervention.