<p>In recent years, people have been using social media platforms to express their feelings and share their mental health struggles openly and anonymously. This surge has motivated many researchers to take advantage of social media as a valuable resource of data to detect severe depression. However, existing approaches have significant limitations as they rely on datasets with a wide disparity between depressive and non-depressive instances, ignoring depression-mimicking expressions (such as stress, anxiety, sadness, sarcasm, and complaints) that are frequently misclassified as severe depression due to their linguistic overlap. This, in turn, leads to false alarms about inaccurate cases of depression, undermining our confidence in the detection system. We present the first study that aims to detect severe depression while simultaneously differentiating it from other depression-mimicking expressions. We curated and annotated a new dataset and refined existing ones to capture these depression-mimicking expressions. We implement and compare the performance of state-of-the-art large language models (LLMs), with RoBERTa emerging as a top-performing model, achieving an AUC of 98.37% on the validation set and 98.53% on the separate test set. Notably, fine-tuning of the models led to an impressive average AUC increase in around 40% over their original baseline versions, significantly enhancing the models’ ability to distinguish severe depression from depression-mimicking expressions. The fine-tuned RoBERTa model generalized well to an external dataset, increasing AUC from 0.65 to 0.97 and significantly reducing false alarms. The significant improvement highlights the effectiveness of fine-tuning LLMs on carefully curated data, reducing false alarms, and boosting the model’s reliability and applicability in practical settings.</p>