The significance of mental health issues such as depression and anxiety has become an increasingly prominent global concern, with Indonesia showing an increasing prevalence of these conditions. Social media platforms, such as Facebook and TikTok, provide valuable data sources for detecting mental health problems as their users often share their emotional state on these platforms. This study proposes a BERT-based deep learning approach, utilizing the IndoBERT model, capable of providing multi-label classifications of depression and anxiety using text contents from Facebook and TikTok in the Indonesian language. The study follows the KDD (Knowledge Discovery in Databases) process as a methodological framework, which includes stages such as data selection, pre-processing, transformation, data mining, and interpretation/evaluation. This approach supports a structured and thorough investigation involving data handling and model fine-tuning. The final IndoBERT model showed strong results, with an overall accuracy of 0.8185 and a Hamming Loss of 0.0971, while Per-label performance showed that the model identified anxiety more effectively compared to depression, potentially due to label imbalance. The F1-Score for Anxiety was 0.8167, while for Depression it was 0.7127. Further analysis using bigram frequency was conducted on classified texts to identify recurring lexical patterns. Results revealed thematic keywords such as “tidak bisa” and “orang tua” as dominant expressions in both depression and anxiety categories, suggesting patterns of emotional distress and family concerns often observed in Indonesian social media. These findings reinforce the model’s reliability and offer a deeper understanding of how mental health topics are expressed in digital conversations.