Energy-efficient privacy-preserving AI models for real-time health monitoring in mobile IoT networks for professional sports applications

Mobile Internet of Things (IoT) wearables can now measure temperature, muscle strength, and heart rate in real-time, revolutionizing professional sports. These systems have certain issues that must be addressed before they can be utilized. These limits may include energy efficiency, computation overhead, and privacy. These limits are more severe in edge contexts due to limited resources and the sensitivity of biometric data. High power consumption, communication latency, and privacy issues plague traditional AI health monitoring systems. These systems’ centralized training and inference pipelines exacerbate these issues. These issues make them unsuitable for high-stakes sporting events where fast and encrypted data from devices is crucial. This study proposes an Adaptive Quantized Federated Dual-Network Framework (AQ-FDNF). It combines three distinct features: (i) a hybrid deep neural architecture for dynamic biosignal interpretation that combines lightweight Temporal Convolutional Networks (TCN) and attention-enhanced Gated Recurrent Units (A-GRU); (ii) layer-wise adaptive quantization for energy-efficient inference across heterogeneous mobile nodes; and (iii) a federated privacy-preserving learning protocol enhanced with local differential privacy and secure multi-party aggregation. An architecture that utilizes a real-time energy-accuracy trade-off controller to select the optimal communication channels and update the model at the optimal frequency is described here. Utilizing real-world sports health information, experimental assessments demonstrate that AQ-FDNF outperforms baseline federated and centralized models in terms of accuracy, energy usage, communication latency, and privacy budget. Adversarial inference and model drift are no match for the suggested system’s ability to withstand changing workloads. The framework achieves an overall classification accuracy of 93.8% and task-specific identification rates exceeding 92% for arrhythmias and fatigue, enabling safe federated convergence with low communication costs (approximately 40 KB per round) and model durability even when signal noise varies. Tailoring learning adaptations increased performance by 4.5% across all athlete profiles. In contrast, threat detection detected harmful updates 96.1% of the time.