# Multi-level determinants of physical activity and sports participation among adults during COVID-19 pandemic: an interpretable machine learning approach

## TL;DR

This study uses machine learning to identify factors influencing physical activity and sports participation during the pandemic, revealing insights for public health policies.

## Contribution

The study introduces interpretable machine learning to analyze multi-level determinants of physical activity and sports participation during the pandemic.

## Key findings

- Random Forest best predicted physical activity with an AUC of 0.613 and identified 10 key factors.
- XGBoost best predicted sports participation with an AUC of 0.772 and identified 12 key factors.
- Common factors included exercise suitability and BMI, while distinct factors varied by community and individual levels.

## Abstract

Both physical activity (PA) and sports participation (SP) are considered important for the promotion of health among adults in the post-disease outbreak period. In the context of the COVID-19 pandemic, the study applied the Socio-ecological Model, with a total of 45 factors on four levels: individual characteristics, individual behaviors, interpersonal relationships, and community environment. The aim was to apply interpretable machine learning algorithms in the examination of common and distinct determinants of PA and SP with the purpose of deriving specific insights relevant to public health policy.

To examine the comparable but different patterns of behavior regarding PA and SP, this research used the Chinese General Social Survey of 2021 with a sample of N = 2,717 participants. Eight machine learning models were designed with the aid of Python coding, including the following models: Logistic Regression, Support Vector Machine, Decision Tree, Random Forest (RF), Adaptive Boosting, Gradient Boosting Decision Tree, eXtreme Gradient Boosting Model (XGBoost), and Light Gradient Boosting. As part of evaluating these models' performance, Accuracy, Area Under the Curve (AUC), and the F1-score results were used after executing the grid search on the models' respective variables. The Permutation Feature Importance method was used to quantify factor importance and identify key factors, and Partial Dependence Plots were generated to interpret the direction of these influences.

Results showed that the best algorithm for predicting PA was the RF with an AUC of 0.613 and that it selected 10 key factors. Additionally, the best algorithm that predicted SP was XGBoost with an AUC of 0.772, and it selected 12 key factors. Common influencing factors during the COVID-19 pandemic include suitability for exercise and recreational lifestyle, with BMI category also playing a significant role. Distinctive factors of PA were primarily related to the community environment (e.g., fresh food outlets and neighborhood care), reflecting its dependence on environmental contexts. In contrast, distinctive factors of SP were more concentrated at the individual characteristics (e.g., education level and socioeconomic status) and behaviors level (e.g., learning and health examination), highlighting the role of personal initiative and the accumulation of socio-cultural and economic capital.

The Socio-ecological Model effectively delineated commonalities as well as differences in determinants of PA and SP across adults during the COVID-19 pandemic. Interpretable machine learning aided in identifying and ranking multi-level determinants, offering a nuanced insight into the relative importance across levels of ecology. These findings provide data-driven insights for future disease outbreaks, facilitating the targeted allocation of intervention resources to key influencing domains.