From Motion to Prevention: A TimeGAN‐Enhanced Hybrid BiLSTM‐CNN‐Attention Model for Predicting Falls and Preventive Interventions

Injury prevention, with a specific focus on fall and sports‐related injuries, remains a paradigm for public health and sports communities. With advances in sensor technology enabling the capture of extensive motion information, there has been a growing need for effective machine learning models to accurately predict the risk of injury. This paper describes ACL‐Net, a hybrid deep learning architecture specifically designed for time‐series classification based on data captured from wearable devices. ACL‐Net integrates BiLSTMs to understand motion patterns over time, 1D‐CNNs for local pattern recognition, and an attention mechanism that emphasizes key motion points. ACL‐Net achieved superior performance, with training accuracy of 97.8%, subject‐independent LOSO accuracy of 91.5%, and strong results under random splits (up to 94.2%). It also yielded a Cohen’s Kappa and Matthews Correlation Coefficient of 0.91 each, a precision of 93.8%, a recall of 92.9% and an F1‐score of 93.3%. This outperforms the individual performances of many baselines, including LSTM, CNN, logistic regression and decision trees. The synergetic result between the new architecture and augmentation using TimeGAN offers credible discriminative capability and identification of risky movements. Unlike existing CNN‐LSTM‐Attention frameworks, ACL‐Net integrates TimeGAN‐based temporal augmentation specifically tailored for wearable inertial data, enabling improved robustness under severe class imbalance.