Online videos are vital for health education and medical decision-making, but their comment sections often spread misinformation, causing anxiety and confusion. This study identifies stress-inducing comments in oral health education content, aiming to improve mental health outcomes, educational effectiveness, user experience, and scalability. This study uses RoBERTa, a state-of-the-art language model, to advance Natural Language Processing (NLP) research and enable real-time feedback in social media environments. The RoBERTa-base configuration, with 12 transformer blocks, attention heads, and a 50,265-token vocabulary, was fine-tuned using optimized hyperparameters. The workflow includes data ingestion, token normalization, special character handling, embedding generation, transformer encoding, classification head processing, output generation, and evaluation metrics. This framework aims to enhance online health education discourse and establish automated comment moderation systems. The RoBERTa model achieved 75.00% overall accuracy in classifying panic and anxiety-inducing comments, with 74.76% precision and 0.800 recall for positive cases. While the model performed well in identifying relevant comments, its accuracy in panic and informative categories requires improvement. This study demonstrates the potential of RoBERTa-based deep learning for classifying dental-related comments, providing clinical insights and identifying areas for refinement. Although the model shows promise in detecting anxiety-inducing content, further optimization is needed.