Radar-based health monitoring provides a continuous, privacy-preserving, and cost-effective solution for mobility and gait analysis. However, improving RF-based skeleton estimation accuracy is crucial for applications such as early disease detection, post-treatment monitoring, and emergency response. Real-time performance depends on computational complexity and latency. This study evaluates accuracy versus real-time performance trade-offs in skeleton estimation using various RF data representations and proposes accuracy enhancement through Kinematic Cycle Consistency (KCC) loss, enforcing biomechanical constraints and temporal coherence. Specifically, we compare CNN-BiLSTM using joint RF representations against models directly processing raw RF data. Results on simulated and experimental datasets, validated against a Vicon optical system, provide insights into real-time radar-based pose estimation.