Using Self-Distillation Strategy to Enhance the Generalization Ability of Nursing Skill Scoring System for Low-Sample Tasks

Automated assessment of nursing skills is important for standardizing clinical training, but it often suffers from overfitting and poor generalization when labeled data are scarce. To address this limitation, this paper proposes a multimodal framework that integrates a cross-modal attention fusion module with a dynamic self-distillation strategy. Video, gesture, and text inputs are processed through dedicated encoders, including TimeSformer, BiLSTM, and RoBERTa, to capture spatiotemporal, kinematic, and semantic features. The cross-modal attention module enables fine-grained interaction among modalities. Self-distillation, with teacher parameters updated every five training cycles, progressively transfers knowledge to enhance robustness in low-sample conditions. The method is evaluated on a nursing-skill dataset aligned with the Nursing Operation Technical Specifications. Results show that the proposed method achieves RMSE of 2.98 using only 10 training samples, which decreases to 0.95 with 50 samples. The model also shows strong cross-validation stability, with RMSE fluctuation within 0.01– 0.03, and real-time efficiency, with a response time of 17 ms. These results indicate that the framework provides a practical solution for data-scarce skill assessment in nursing education.