Comparison of predictive models for knee pain and analysis of individual and physical activity variables using interpretable machine learning

Abstract

Background

Knee pain is associated with not only individual factors such as age and obesity but also physical activity factors such as occupational activities and exercise, which has a significant impact on the lives of adults and the elderly.

Objectives

The purpose of this study was to construct a model for predicting knee pain using individual and physical activity variables and to determine the relationship between knee pain and individual and physical activity variables.

Design

Observational study.

Methods

A total of 19 variables related to individual and physical activity were used to create a knee pain prediction model. Model composition variables were selected using recursive feature elimination with cross validation. The performance of the model was evaluated using test data, and the relationship between knee pain and predictor variables was analyzed using SHapley Additive exPlanations (SHAP).

Results

The CatBoost model showed the highest performance. And, activity limitation was identified as the most influential predictor, followed by weekly physical activity, body image, weight change, occupational type, age, BMI, and housing type.

Conclusion

Knee pain prediction models built with individual and physical activity variables can exhibit relatively high predictive performance, and interpretable machine learning models can provide valuable insight into the complex relationships between individual and physical activity variables and knee pain.