Depression Risk Classification Using Machine Learning: A Model Performance Study

Abstract

This study presents a comparative evaluation of multiple machine learning algorithms for depression risk classification using a publicly available mental health survey dataset. Rather than predicting clinical depression, the target variable is formulated as a risk proxy derived from social weakness indicators to support screening-oriented analysis. A quantitative experimental framework is employed to compare Logistic Regression, Random Forest, Support Vector Machine, and Extreme Gradient Boosting under consistent preprocessing and data partitioning conditions. Model performance is evaluated using complementary metrics, including accuracy, recall for High-risk cases, and the area under the receiver operating characteristic curve (ROC-AUC). Threshold optimization based on ROC analysis is applied to align model outputs with screening objectives that prioritize sensitivity. The results demonstrate that Logistic Regression and Support Vector Machine consistently achieve superior or comparable performance across all evaluation dimensions, including high overall accuracy, near-perfect sensitivity for High-risk detection, and strong discriminative capability. In contrast, more complex ensemble and distance-based models show mixed outcomes, indicating diminishing performance gains from increased algorithmic complexity. These findings highlight that simple and interpretable models can effectively support depression risk screening using survey-based data, offering a practical balance between predictive performance, transparency, and computational efficiency.