Hybrid Deep Learning Healthcare AI Framework for Real-Time Human Pose Estimation and Remote Patient Monitoring to Support TKR Physiotherapy

Total Knee Replacement (TKR) rehabilitation critically depends on precise physiotherapy exercise execution, and the rise of patient volumes and constrained clinical resources limit continuous supervision. This study presents an Artificial Intelligence (AI) framework for real-time assessment and feedback of TKR exercises using deep learning–based human pose estimation to empower remote rehabilitation. We investigate three architectures: a Dense Convolutional Neural Network (DCNN) incorporating frame decoupling for robust joint tracking; a pruned Generative Adversarial Network (Sparse GAN) optimized for computational efficiency; and a novel hybrid model that embeds the DCNN as a discriminator within the GAN model. A diverse dataset of over 10,000 annotated video clips, sourced from clinical environments and public repositories, was processed with OpenCV, and joint annotations were generated using OpenPose. Models were trained and evaluated on standard metrics (i.e. Precision, Recall, F1-score) alongside runtime and memory usage benchmarks. The hybrid architecture achieved the highest classification performance with 86.01% F1-score, which demonstrates the synergetic benefits of combining rich feature extraction with generative refinement, though it incurred elevated computational costs. The Sparse GAN provided faster inference suitable for mobile deployment, with only a marginal decrease in F1-score. The standalone DCNN provided a balance between accuracy and speed, but it did not match the hybrid’s precision. These results highlight a fundamental trade-off between model complexity and real-time usability in AI-driven therapeutic monitoring. The hybrid model is optimal for clinical settings where accuracy is paramount, while the Sparse GAN provides a practical solution for resource-constrained and edge-based applications. Future work will explore model compression, hardware acceleration, and edge-computing strategies to further optimize performance. By demonstrating the viability of advanced pose estimation techniques in a physiotherapy context, this research contributes to the broader discourse on the use of AI in healthcare for scalable, autonomous rehabilitation tools across several medical and wellbeing domains.