A Deep Hybrid Learning Framework with Attention-Enhanced Feature Extraction for BMI Prediction Based on Physical Fitness

Ming Mo¹Wanhong Luo²Qiao HU^3*Jun Wang¹Tianshuo Jiao³Libo Xie⁴Guixiang Wu⁵Ye Yang¹Jinfeng Deng¹Xuyin Xu¹

• ¹Changsha Aeronautical Vocational and Technical College, Changsha, China
• ²Hunan First Normal University, Changsha, China
• ³Hunan University, Changsha, China
• ⁴Hunan Wuxu Network Technology Co., Ltd., Changsha, China
• ⁵Hunan Yanpei Technical School, Changsha, China

This study investigates the relationship between physical fitness and BMI classifications among university students. A total of 6,698 male students aged 18-20 underwent four fitness tests: the 3000-meter run, 30×2 shuttle run, pull-ups, and sit-ups. BMI measurements categorized participants as underweight, normal, overweight, or obese. A hybrid machine learning model combining a one-dimensional convolutional neural network (CNN1D) with an attention mechanism was employed for feature extraction, followed by classification using LightGBM. The CNN1D identified key patterns in fitness data, while the attention mechanism emphasized relevant features for BMI prediction. This hybrid model achieved an accuracy of 94.5% and an F1 score of 0.93, outperforming traditional classifiers such as SVM and XGBoost. Significant correlations were found between fitness test results-particularly the 3000-meter run and pull-ups-and BMI classification, providing valuable insights for personalized health interventions and public health strategies.