MSBA-LSTM: a multi-scale bidirectional attention long short-term memory network for early risk prediction from multimodal wearable sensor time-series data in sports

Multimodal wearable sensor time-series data present fundamental challenges for sequential risk prediction, including high dimensionality, heterogeneous sampling rates, and cross-scale temporal dependencies. Existing recurrent architectures are constrained by unidirectional information flow, single-scale temporal receptive fields, and insufficient dynamic weighting across heterogeneous sensor channels. To address these limitations, we propose the Multi-Scale Bidirectional Attention Long Short-Term Memory Network (MSBA-LSTM), a hybrid feature-engineered deep learning framework integrating three synergistic modules: (i) a stacked Bidirectional LSTM (BiLSTM) encoder that exploits both forward and backward temporal context at every timestep; (ii) a Multi-Head Self-Attention (MHSA) module with 8 parallel heads ( \(d_k\) =32), residual connection, and Layer Normalization that dynamically focuses on semantically critical timesteps; and (iii) a multi-scale feature extraction strategy that captures biomechanical dynamics at a 0.5 s window (impact-level), time-domain and frequency-domain patterns at 1.0 s (movement-level), and physiological state at 5.0 s (fatigue-accumulation level), spanning three biomechanically meaningful temporal granularities. The model is trained with AdamW, cosine annealing scheduling, weighted binary cross-entropy, and SMOTE oversampling under stratified five-fold cross-validation, with additional subject-independent leave-one-subject-out (LOSO) validation to assess generalization to unseen athletes. On a self-collected dataset of 295 labeled sessions, MSBA-LSTM achieves 93.2% accuracy, 92.6% F1-score, and AUC-ROC of 0.967 under window-level cross-validation, and maintains an accuracy of 88.5% under the stricter LOSO protocol, outperforming standard LSTM by 5.9%, 6.8%, and 4.9 percentage points respectively, and surpassing the strongest deep-learning baseline (InceptionTime) by 3.4 percentage points in accuracy and 0.026 in AUC-ROC, with all pairwise improvements statistically significant ( \(p<0.001\) , paired Wilcoxon signed-rank tests). Ablation studies validate the contribution of each module, and SHAP analysis provides feature-level attribution scores together with sport-category-wise importance variations that confirm the interpretability of the learned representations.