ACSAT: An Adaptive Contextual Sentiment-Aware Transformer for Early Detection of Depression in Social Media Text

Objectives: Prompt and correct recognition of depression from social media posts presents an economic, non-invasive method of managing mental health. However, the current models typically perform poorly on very subtle contextual and emotional hints, resulting in not-so-optimal detection, especially about severity levels. Methods: This study introduces ACSAT—an Adaptive Contextual Sentiment-Aware Transformer—that endorses a transformer encoder, a sentiment-aware attention model, and severity-sensitive contrastive learning, which uses margin-weighted separation and balanced sampling to capture both categorical differences and the ordinal progression of depression severity. The model is trained on user posts to label depression into three categories: None, Moderate, and Severe. The model weighs emotionally meaningful tokens adaptively and learns fine-grained distinctions in the linguistic presentation of depressive symptoms. Findings: The ACSAT model was tested on the Mental Health Disorder Detection Dataset from Kaggle; the data was processed using stratified fold cross validation to make for a robust experiment. The model achieved an accuracy of 96%, precision of 94%, recall of 95%, and F1-score of 92%, consistently beats all baseline models such as Support Vector Machines (SVM), Random Forests (RF), and classical NLP classifiers. The enhanced performance is mainly driven by ACSAT's dual sensitivity to both contextual semantics and emotional nuances. All baselines were tuned using the same preprocessing, features, and stratified fold splits with nested cross-validation for hyper-parameter tuning and testing, in order to compare the models fairly without test leakage. Novelty: In contrast to previous models, ACSAT uses contextual embeddings combined with affect-sensitive attention and applies depression severity-specific contrastive learning better to identify subtle variations in effect between social media posts. ACSAT is thus a strong, generalizable, early and scalable depression detection method. Keywords: Depression detection, Mental health prediction, Social media mining, Contrastive learning, Text classification, Psychological disorder detection, Emotion-aware AI