Deep Learning-based Detection of Depression Severity

Identifying and assessing the severity of depression through social media text offers critical opportunities for enhancing mental health support systems. This paper conducts a comparative study of three deep learning approaches DeProBERTa and a hybrid CNNBiLSTM model aimed at classifying depression severity from comments. A dataset comprising Reddit user posts was preprocessed through sentiment analysis using VADER, TF-IDF feature extraction, and text normalization techniques. Sentiment scores were mapped to severity labels, facilitating supervised model training. Among the models, the BiLSTM architecture demonstrated superior performance, achieving a validation accuracy of 84.10%, compared to DeProBERTa (70.36%). Explainable AI frameworks were utilized to interpret model decisions, highlighting influential linguistic patterns linked to depression levels. The study underscores the advantages of recurrent neural models in capturing complex emotional signals and stresses the role of interpretability in clinical natural language processing applications.