With the increase of social pressure and the popularity of social media, the mental health problems of adolescents are becoming increasingly prominent, and timely and effective mental health warning has become an important issue that needs to be solved urgently. To solve this problem, this paper proposes an artificial intelligence-based adolescent mental health warning system, focusing on the deep learning algorithm based on the Long Short-Term Memory (LSTM) network. Firstly, social media data is collected and preprocessed to extract potential mental health risk characteristics. Then, the LSTM model is used to perform time series modeling on the extracted emotional states and behavioral patterns to identify changes in the mental states of high-risk individuals. Finally, personalized features and historical data are combined to achieve early warning based on logistic regression, so that relevant personnel can be reminded to conduct psychological intervention in a timely manner. The experimental results show that the LSTM model has a prediction accuracy of 70% for 200 data points, and increases to 80% for 600 data points. In terms of warning response time, the LSTM model is 0.9 s for 1,000 data points, which is significantly better than other algorithms. In the above data conclusions, the early warning system proposed in this paper is efficient and real-time in identifying adolescent mental health risks, and can effectively meet practical application needs.