Hybrid Confidence Model for Depression Detection

This work considers the detection of mental health from social media text, a challenging task due to noisy labels and subtle linguistic cues related to psychological distress. This paper presents a novel hybrid methodology that utilizes large language model (LLM) confidence scores for data filtering and combines classical machine learning with attention-based neural networks to identify depressive comments on Reddit. This approach utilizes Mistral LLM in an initial confidence thresholded labeling, as well as feature engineering that combines sentence embeddings and both sentiment analysis and LLM confidence scores. Comparing several ensembling methods such as logistic regression, XGBoost, hybrid meta-learning and learned attention based fusion networks. Results show that the confidence-guided data filtering greatly enhances model trust, achieving 93 percent accuracy (both precision and recall) with tranditional ensemble method, and providing interpretable feature importance with attention-based fusion. To the best knowledge, this paper was the first study that systematically integrated LLM confidence scores as explicit features in mental health classification with improved performance and interpretability for clinical psychologists.