Online support groups offer low-cost and accessible health and mental health support, but low engagement challenges their effectiveness. To encourage engagement and provide evidence-based responses appropriate to participants' needs, we propose an intent detection model for online support groups based on state-of-the-art natural language processing methods. Such a model enables a chatbot that can increase interactions and improve discussion quality. Posts in social media are often short and noisy, especially in group chat. Furthermore, many intents lack data, overlap and/or have specific priorities. We create a human-annotated dataset of posts with intent labels from 45 three-month online support groups for quitting smoking. We then train and examine models to predict the intent behind each post. To reduce the effect of noisy and sparse data, we fine-tune a massive pretrained language model. Also, to represent the unique relationships between intents, we design customized loss functions for training. Empirical evaluations show significant performance improvements with the proposed method; our best model obtains 95.5% accuracy. We also use a fine-grained set of intents and obtain higher accuracy compared to prior models on online health forums and communities. Accurate detection of fine-grained intents opens up new opportunities to improve online self-help support groups.