ENHANCING BIPOLAR DISORDER DETECTION USING UNSUPERVISED CLUSTERING-DRIVEN OPTIMIZED XGBOOST: A GAUSSIAN MIXTURE MODEL AND SMOTE-BASED HYBRID APPROACH

Bipolar Disorder (BD) is a serious mental health condition that is often difficult to diagnose correctly due to overlapping symptoms and imbalanced clinical data. Early and accurate diagnosis can help patients receive timely treatment and support. In this study, we developed a hybrid machine learning approach that integrates Gaussian Mixture Model (GMM), Synthetic Minority Over-sampling Technique (SMOTE), and an optimized XGBoost classifier to improve the detection of BD. First, GMM was applied to uncover hidden clusters of patients based on medical and psychometric features, generating informative cluster assignments as additional features. Next, SMOTE was used to address the imbalance between bipolar and non-bipolar cases by generating synthetic samples for the minority class. Finally, an optimized XGBoost model was trained on the enriched and balanced dataset, leveraging both original clinical features and GMM-derived cluster representations for improved predictive power. The GMM-SMOTE-XGBoost hybrid model achieved an accuracy of 93.27% on an independent test set. Comparative analysis against conventional classifiers such as Logistic Regression(LR), Decision Tree(DT) and Support Vector Machine(SVM), and shows that this hybrid approach consistently outperformed all baselines, particularly in sensitivity and overall diagnostic reliability. This work demonstrates the potential of combining unsupervised clustering, synthetic data balancing, and optimized gradient boosting to enhance psychiatric disorder classification. Such hybrid frameworks could serve as valuable clinical decision-support tools, improving diagnostic accuracy and robustness in mental health diagnostics