Accurate movement tracking is vital for assessing motor recovery in neuro-rehabilitation, yet general-purpose pose estimation models often underperform in clinical environments characterized by constrained motion, limited data, and fixed viewpoints. This paper introduces NeuroYOLO, a data-efficient, fine-tuned YOLOv11-Pose model optimized for vision-based rehabilitation monitoring. Using Bayesian hyperparameter optimization, NeuroYOLO adapts to structured exercise recordings of elderly participants performing assisted upper- and lower-limb tasks, demonstrating that high performance can be achieved without large clinical datasets. The fine-tuned model achieves superior detection precision and temporal stability compared to the default configuration, improving pose mAP@0.5:0.95 from 0.989 to 0.991 and box mAP@0.5:0.95 from 0.961 to 0.985, while reducing inference latency by 36%. These results confirm that targeted, data-efficient domain adaptation can significantly enhance pose estimation reliability under realistic rehabilitation conditions. NeuroYOLO thus establishes a lightweight and deployable foundation for real-time, vision-based patient monitoring and future multimodal rehabilitation frameworks.