Mamba-SportsNet: High-Frequency Athletic Running Performance Analysis via Gated Cross-Fusion State Space Models

Wearable sensor networks have revolutionized athletic training by enabling in-field monitoring, yet accurate performance assessment is hindered by the difficulty of modeling long-range dependencies in high-frequency, heterogeneous data streams. Traditional Recurrent Neural Networks struggle with memory retention over long durations, while Transformer-based architectures suffer from quadratic computational complexity, restricting their deployment on resource-constrained edge devices. To overcome these limitations, this paper proposes \textbf{Mamba-SportsNet}, a novel multimodal framework based on the Selective State Space Model (Mamba). Leveraging the linear complexity of Mamba ($O(L)$), our architecture efficiently processes synchronized kinematic (IMU) and physiological (sEMG) streams to capture both transient impact events and slowly evolving fatigue patterns. Furthermore, we introduce a \textbf{Gated Cross-Fusion (GCF)} mechanism that models the conditional dependency of mechanical execution on physiological state, allowing the network to dynamically attend to technique degradation induced by metabolic load. Extensive experiments on a dataset of 20 semi-professional athletes demonstrate that Mamba-SportsNet outperforms state-of-the-art baselines, achieving a Ground Reaction Force (GRF) estimation RMSE of \textbf{0.112 BW} and a fatigue classification accuracy of \textbf{94.2\%}. Notably, the model maintains an inference latency of just \textbf{9 ms}, proving its viability for real-time, continuous monitoring in next-generation digital sports applications.