With the growing prominence of online education, personalized course recommendation has become essential for enhancing learning experiences. While existing recommendation methods primarily rely on course attributes, entity relationships, and user behavior, they often overlook teaching style, a critical factor influencing learner engagement and comprehension. This study proposes a novel course recommendation method that integrates teaching style analysis. Leveraging large language models (LLMs), the method extracts multidimensional teaching style features from multimodal teaching data, including instructional videos, lecture audio, and text. These features are combined with auxiliary course information to construct comprehensive course representations, which are then integrated with user learning records to generate personalized recommendations through a multi-layer perceptron model. Experimental results on real-world data demonstrate that the proposed method outperforms existing approaches, achieving average improvements of 15.58% in precision, 15.17% in recall, and 15.04% in F1 score. Ablation studies confirm the significant contribution of teaching style analysis to recommendation accuracy. This study represents the first attempt to incorporate teaching style features into course recommendation using LLMs, providing new insights for future research on adaptive learning systems.