This study proposes a teaching-improvement model for midwifery education in higher vocational colleges that integrates artificial-intelligence-based assessment with classroom interaction analysis. The model is intended to enhance teaching quality and students’ learning experience by using AI techniques to capture dynamic information that is difficult to obtain through conventional end-of-course assessment. A dual-path architecture combining ELMo-BiLSTM-CRF with a Transformer encoder is employed to analyze classroom interaction data across multiple modalities, including speech, text, and behavioral information, thereby supporting emotion recognition, engagement assessment, and feedback on instructional strategies. The study further incorporates a closed-loop mechanism in which AI-generated assessment results are linked to instructional adjustment so that classroom interaction, learner affect, and teaching decisions can be analyzed within a unified framework. The experimental sections report model-performance analyses, group comparisons, and multimodal visualization results. In classroom interaction emotion recognition, the proposed dual-path model achieved an accuracy of 89.7%, a recall of 90.3%, and an F1 score of 90.0%, outperforming the comparative approaches reported in the study. Classroom engagement in the AI-enhanced group reached 85.9%, compared with 68.4% in the control group, while the Emotion Improvement Index increased from 0.21 to 0.43. Additional analyses illustrate how the framework can capture changes in learner emotion and engagement, examine the influence of regulatory mechanisms, and extract information relevant to teaching improvement. Overall, the results support the potential of AI-assisted classroom analytics as a decision-support tool for data-informed improvement in vocational midwifery education.
