Feature group partitioning: an approach for depression severity prediction with class balancing using machine...

Area of Science:

• Medical Informatics
• Machine Learning
• Mental Health

Background:

• Depression is a growing global health concern with significant mortality impact.
• Machine learning models are increasingly used for depression prediction, but often struggle with multiclass imbalance.
• Addressing class imbalance in multiclass depression severity prediction remains a critical challenge.

Purpose of the Study:

• To introduce and evaluate a novel Feature Group Partitioning (FGP) approach for preprocessing data in depression severity prediction.
• To investigate the effectiveness of synthetic oversampling techniques (SMOTE, ADASYN) in mitigating class imbalance.
• To compare the performance of various machine learning algorithms and ensemble methods on a multiclass depression dataset.

Main Methods:

• Implemented Feature Group Partitioning (FGP) to reduce feature dimensionality.
• Applied Synthetic Minority Over-sampling Technique (SMOTE) and Adaptive Synthetic (ADASYN) for class balancing.
• Utilized heterogeneous ensemble stacking, homogeneous ensemble bagging, and five supervised algorithms, evaluating on training, validation, and testing sets.

Main Results:

• The Feature Group Partitioning (FGP) approach significantly reduced feature dimensionality.
• The stacking classifier combined with FGP and SMOTE achieved the highest balanced accuracy of 92.81%.
• The FGP approach demonstrated improved performance in predicting depression severity and optimized training time.

Conclusions:

• The Feature Group Partitioning (FGP) approach, coupled with SMOTE, is highly effective for multiclass depression severity prediction.
• This method successfully addresses class imbalance and enhances prediction accuracy.
• The optimized training time represents a significant practical advantage for clinical applications.