¹ Physical Education Institute, Luoyang Normal University, Luoyang, 471000, China;

² Physical Education Institute, Zhengzhou University of Science and Technology, Zhengzhou 450064, China;

³ Physical Education Institute, Henan University of Science and Technology, Luoyang 471000, China.

* Correspondence: wj612511@126.com

The deployment of sophisticated deep learning models on battery-operated wearable devices is severely restricted by the von Neumann bottleneck and power constraints inherent in traditional CMOS architectures. This study introduces a neuromorphic hardware accelerator based on Ovonic Threshold Switching (OTS) memristors to enable ultra-low-power edge computing for athletic monitoring. We systematically investigated the impact of doping on Ge₂Sb₂Te₅ (GST) chalcogenides, revealing that carbon doping (C-GST) effectively refines the grain size to approximately 12 nm and stabilizes the amorphous phase. Consequently, the fabricated C-GST devices demonstrate superior reliability, characterized by an endurance exceeding 10⁹ switching cycles, a 10-year data retention at 85 °C, and a robust resistance window greater than 100. By mapping these experimentally characterized devices into a 256×256 crossbar-based hardware model, we evaluated a quantized hybrid CNN-LSTM network for real-time sports motion classification with downstream injury-risk warning. The system achieves a recognition accuracy of 96.2% across eight distinct movement patterns, closely approaching the performance of 32-bit floating-point software baselines. Critically, the in-memory computing architecture reduces inference latency to just 0.41 ms and operates with a total system power of 5.8 mW. This translates to an energy consumption of 2.38 µJ per inference, representing a massive efficiency gain compared to conventional edge GPU solutions. These findings establish C-doped chalcogenide memristors as a viable, high-performance platform for next-generation intelligent wearable electronics.