Due to the increase in the online user base and their Internet usage and activity, a humungous amount of content is generated daily on the Internet, especially on online social networks. Though this user-generated content can be collected easily from these websites and online social networks using their corresponding Web APIs, however, annotating and obtaining ground truth labels for every user post (which is required to develop any social-computing system) still remains a challenging and cost-intensive process. Data annotation is even more difficult for healthcare applications where these ground truth labels must be provided by domain specialists. These challenges related to data annotation of user-generated content inhibit the research and development of social-computational linguistic systems for any real-world use case. One of the crucial applications of user-generated content is public healthcare surveillance. The recent research in this domain has mainly utilized deep learning neural networks, especially pre-trained Transformer-based Large Language Models, e.g., BERT. BERT has now become the state-of-the-art for most of the NLP tasks. However, the accuracy and reliability of any machine learning driven decision-making system heavily depends on the quality of training data. Any supervised learner is only as good as the annotated data used as input to train them. Most of the previous research studies have primarily worked on improving the classification accuracy via algorithmic enhancements. In contrast, our research emphasizes on reducing the data annotation costs for cost-efficient categorization of user-generated content to develop decision-making systems for any social-computational task, e.g., sentiment analysis, opinion mining, content moderation, topic identification, etc. In this research, through our proof of concept using Deep Active Learning and Large Language Models, we demonstrate that it is possible to achieve high/comparable accuracy with as few as 10% of samples from the entire dataset compared to using the entire dataset for model development. For our POC, we have used mental healthcare datasets to demonstrate the use of BERT with Active Learning Loop to detect various mental health disorders from online user-generated content.