Depression, a global mental health concern of increasing urgency, has been further exacerbated by recent crises such as the pandemic. Detecting depression from short texts on online platform is a significant challenge due to the informal and diverse nature of online communication. Existing studies use complex features to characterize depressive words, but choosing optimal features remains a challenge. This study introduces a new FS method, class-specific mutual information (CSMI), designed to enhance the accuracy of depression detection from short text data. The proposed method extracts unigram, segment-based bigram, and segment-based trigram features. These features are efficient in preserving semantic meaning. Subsequently, the CSMI scores for every feature are computed based on its relevance to the depressive and non-depressive classes, selecting the top N% ranked features (10, 20, 30, 40, 50%) to improve differentiation between depressive and non-depressive content. Evaluation involves three publicly accessible short text datasets and two classifiers, demonstrating that the proposed technique surpasses traditional methods in (Accuracy and F1-scores). These results underscore the practical application of CSMI in enhancing mental health monitoring, offering a robust tool for early detection and intervention efforts. By leveraging social media data effectively, mental health professionals can better identify at-risk individuals and deliver timely support. Future research directions include real-time implementation and extension to other mental health conditions, promising broader impacts in addressing the escalating mental health crisis.