Predictive modelling links exercise dependence to associated psychological and behavioral risk factors

Exercise Dependence (ED) refers to uncontrollable, excessive exercise with harmful effects on life. This study used machine learning to identify behavioral and psychological factors contributing to ED risk. A multi-step procedure was implemented for model construction and validation, utilizing controlled feature selection and bootstrapping. Data were collected over three time points (GR2021-22-23), recruiting 1099 participants (707 males, 64.3%; 392 females, 35.7%) with an average age of 24.8 ± 7.8 years. Based on the Exercise Dependence Scale-Revised (EDS-R), 5.6% (n = 62) were classified as "At Risk" of ED, 50.9% (n = 559) as "Non-Dependent-Symptomatic," and 43.5% (n = 478) as "Non-Dependent-Asymptomatic." Interestingly, 24.0% (n = 264) reported supplement use, 52.5% (n = 577) used analgesics in the last 12 months, 48.1% (n = 528) reported nicotine use, and 88.2% (n = 969) consumed alcohol during the same period. The final model predicted the GR2023 dataset with MAE = 6.90, R² = 0.59, and RE = 9.08%. Predictive performance on the GR2022 dataset was MAE = 5.65, R² = 0.79, and RE = 6.73%, while performance on the GR2021 dataset achieved MAE = 7.60, R² = 0.58, and RE = 7.24%. This study establishes a foundation for developing quantitative risk profiles for ED by analyzing multidimensional constructs and their contributions through interpretable machine learning. The methodology offers insights into how personality, psychological, and behavioral dimensions shape risk attitudes and provides robust predictive tools for assessing ED risk in sports contexts.运动成瘾（Exercise Dependence, ED）指不受控制的过度运动行为，且会对个体生活造成不良影响。本研究借助机器学习技术，识别与运动成瘾风险相关的行为与心理因素。研究采用多步骤流程开展模型构建与验证工作，结合受控特征选择与自助法（bootstrapping）进行分析。数据采集覆盖三个时间节点（GR2021-22-23），累计招募1099名参与者，其中男性707人，占比64.3%；女性392人，占比35.7%，参与者平均年龄为24.8±7.8岁。基于修订版运动成瘾量表（Exercise Dependence Scale-Revised, EDS-R）的评估结果，5.6%的参与者（n=62）被归类为运动成瘾风险人群，50.9%（n=559）为非成瘾但有症状人群，43.5%（n=478）为非成瘾且无症状人群。值得注意的是，24.0%（n=264）的参与者报告存在营养补剂使用行为，52.5%（n=577）在过去12个月内使用过镇痛药物，48.1%（n=528）存在尼古丁使用行为，同期88.2%（n=969）的参与者有饮酒习惯。最终模型在GR2023数据集上的预测性能为：平均绝对误差（Mean Absolute Error, MAE）=6.90，决定系数（R²）=0.59，相对误差（Relative Error, RE）=9.08%；在GR2022数据集上的预测结果为MAE=5.65，R²=0.79，RE=6.73%；而在GR2021数据集上的预测性能则达到MAE=7.60，R²=0.58，RE=7.24%。本研究通过可解释机器学习分析多维构念及其贡献，为构建运动成瘾定量风险评估模型奠定了基础。该研究方法揭示了人格、心理与行为维度如何塑造风险态度，并为运动场景下的运动成瘾风险评估提供了稳健的预测工具。