A novel framework for seasonal affective disorder detection: Comprehensive machine learning analysis using multimodal social media data and SMOTE

Abstract

Seasonal Affective Disorder (SAD) is a mood disorder characterized by recurring depressive episodes during specific seasons, particularly in Fall and Winter. With the rise of social media as a platform for self-expression, user-generated content offers valuable insights into mental health trends, presenting an opportunity for data-driven SAD detection. However, existing research faces challenges such as limited multimodal datasets, class imbalance, and the need for real-time predictive models. This study addresses these gaps by curating a unique social media dataset that captures seasonal patterns and employing advanced machine learning techniques for accurate SAD detection. We apply the Synthetic Minority Over-sampling Technique (SMOTE) in two distinct ways—on the training dataset post-splitting and the entire dataset—to assess its impact on model generalization. Our findings highlight Random Forest, LGBM, and XGBoost as the top-performing models, with K-Nearest Neighbors (KNN) achieving the highest accuracy of 97.87 % in the training dataset. Additionally, we optimize computational efficiency to ensure real-time scalability for large-scale social media data processing. This research advances SAD detection by integrating robust dataset curation, class imbalance mitigation, and machine learning optimization, paving the way for more effective mental health monitoring through social media analytics.