Abstract: The current study explored the long-term impact of the DeepSeek-integrated blended teaching model on medical students' academic English writing self-efficacy in the context of the digital transformation of medical education,. Through a quasi-experimental design (N=124), the subjects were divided into an experimental group (DeepSeek-integrated teaching) and a control group (traditional teaching), and a three-stage longitudinal tracking (pretest-immediate posttest-delayed posttest) was used to measure the changes in self-efficacy. The study found that: (1) The experimental group had significant improvements in terminology accuracy confidence (d=0.72, p<0.01) and paper structure control (d=0.65, p<0.05); (2) Qualitative analysis showed that 23.7% of over-assisted users were at risk of DeepSeek dependence; (3) The delayed posttest showed that the improvement in efficacy was sustained (β=0.41, p<0.001), but metacognitive training was required. The study showed that DeepSeek showed the characteristics of “limited empowerment” in medical English writing, and its long-term mechanism depended on the teaching design of human-computer collaboration. This study provides empirical evidence for the application threshold of DeepSeek in medical education.

Key words: DeepSeek integration, Medical academic English writing, Self-efficacy, Blended teaching, Long-term mechanism research, Medical education