The emotional feedback mechanism in physical education is of great significance to improving teaching effectiveness and student participation. This study optimized the mechanism through machine learning algorithms, constructed an emotion recognition model based on multimodal data, and combined convolutional neural networks with recurrent neural networks to achieve efficient recognition of students’ emotional states. The study designed a weighted loss function and attention mechanism to improve the accuracy and real-time performance of the model in scenarios such as classroom teaching, after-school training, and psychological counseling. The application results show that the model can accurately capture emotions such as positivity, fatigue, and anxiety, and generate personalized feedback, effectively assisting teachers to adjust teaching strategies and enhance students’ learning experience. Case analysis further verifies the differences in the applicability of the emotional feedback mechanism in different scenarios, and provides theoretical support and practical basis for the intelligent development of physical education. This study shows the broad prospects of machine learning technology in the field of education and provides a new path for optimizing teaching quality and student mental health.