This research seeks to examine how teenagers develop self-esteem using advanced deep learning techniques, and create a reliable prediction model by merging insights from psychology and data science. Utilizing data from the National Youth Mental Health Survey, which includes teens’ basic information and scores from various psychometric questionnaires, the model merges the benefits of CNN and LSTM to effectively forecast self-esteem scores of adolescents. The research results show that family background and psychometric questionnaire scores are key factors affecting the development of adolescent self-esteem, especially anxiety scores and self-esteem scores that contribute the most to predicting results. In addition, through cross-validation and confusion matrix analysis, we evaluated the generalization ability and classification performance of the model, verifying the significant advantages of deep learning methods in processing large-scale, multi-dimensional mental health data. The research results are of great significance in practical applications. They can assist the education system in monitoring students’ mental health status in real time and offering data support for mental health intervention, as well as laying a foundation for the government and relevant institutions to develop more scientific mental health policies. Simultaneously, parents and community groups can utilize this information to enhance their assistance and direction of the positive development of adolescents.