Evaluating Machine Learning Models for Mental Health Diagnostics: A Comparative Analysis and Visual Insights

Gregorius Airlangga (Atma Jaya Catholic University of Indonesia, Jakarta)

KLIK: Kajian Ilmiah Informatika dan Komputer
Vol. 4 No. 4 (2024): Februari 2024

This study addresses the critical challenge of enhancing mental health diagnostics amidst a surge in global mental disorder prevalence. With mental health conditions predicted to become the leading cause of disability by 2030, there is an urgent need for more effective diagnostic methods that transcend the limitations of traditional frameworks, such as subjectivity and clinician bias. Leveraging the capabilities of machine learning (ML) to analyze complex datasets, this research aims to fill the gap in the comparative effectiveness of various ML models, particularly within the context of imbalanced mental health datasets. We systematically evaluated the performance of diverse ML models—including Random Forest, Gradient Boosting, Support Vector Machines, and others—on a rich dataset embodying a wide spectrum of symptoms and diagnoses. Through advanced data preprocessing techniques, such as innovative handling of missing values and categorical encoding, coupled with RandomizedSearchCV for model optimization, we provided a comprehensive analysis of the models' effectiveness. The application of oversampling strategies addressed the challenge of dataset imbalance, ensuring realistic clinical scenario evaluations. The study's findings are presented through detailed model performance metrics and visual analytics, such as symptom distribution visualizations and correlation cluster maps, enhancing interpretability and clinical relevance. The discussion section explores the practical applicability of these findings in clinical settings, acknowledging limitations and outlining future research directions. In conclusion, the study presents a nuanced narrative of ML model selection and performance evaluation complexities. The superior performance of ensemble methods like Random Forest and Gradient Boosting classifiers for certain diagnoses demonstrates the potential of ML in mental health diagnostics. However, the varied performance across models underscores the importance of context-specific model selection, considering the trade-offs between accuracy, interpretability, and computational efficiency. This research contributes significantly to the field of mental health diagnostics by highlighting models with the greatest promise for clinical application and by providing a framework for future advancements integrating ML into mental health diagnostics.