To address poor generalization in lying-posture recognition for healthcare monitoring—caused by scarce data and category misalignment in real-world deployment—we propose a clustering-based transfer-learning framework. First, we analyze the limitations of conventional transfer strategies, such as fine-tuning and shallow-layer freezing, as well as domain-adaptation techniques, when the source and target domains differ (e.g., mattress thickness of 5 cm and 20 cm). To mitigate accuracy loss due to category misalignment, we introduce Structured Regularized Deep Clustering (SRDC), which enhances cross-domain feature alignment and discrimination. Experiments demonstrate that SRDC achieves a mean accuracy of 84.36%, even when some categories are absent, significantly improving model generalization in semi-supervised cross-domain settings. This approach provides a robust and scalable solution for real-world healthcare applications.