Dual-Modality Depression Severity Prediction Using PHQ-9 and Emotion-Aware Text Modeling with Transformers
Abstract
Depression is currently known to have over 280 million victims in the world, yet it has not been adequately diagnosed because it uses subjective and inaccessible diagnostic methods. The current study introduces a new two-modality model that combines the survey data on the structured PHQ-9 with the unstructured emotional self-reports to forecast the degree of depression. We use traditional machine learning classifiers and BERT model based on transformers that runs on a mental health-specific dataset. The proposed model demonstrates a classification accuracy of 87.6 percent with the use of BERT and 78 percent with the use of Random Forests with structured inputs through the methods of ensemble and comparative analysis. High-level statistical tools such as ANOVA, Pearson correlation, and PCA show that such significant predictors as sleep hours and suicidal thoughts are present. PCA plots and confusion matrices are other visualizations that can be used to enhance model interpretability. The system will provide a scalable and interpretable early mental health diagnosis tool that will be ideally applicable in under-resourced settings.