2026 ECNP Congress

PS03-3034

Enhancing depression screening with machine learning by addressing class imbalance in a nationally representative south korean survey

Dong Hyeon Kim — J.W. Paik, A.R. Lee · Kyung Hee University Medical Center, Department of psychiatry, Seoul, Korea- Rep

2026 ECNP Congress

Introduction

Major depressive disorder represents a significant global health challenge, contributing to reduced quality of life, economic productivity loss, and increased suicide risk[1]. Efficient primary screening is essential for early intervention and improved prognosis. The nine-item patient health questionnaire (PHQ-9) is the gold standard for this purpose[2], but its traditional use via simple total score cut-offs (e.g., ≥ 10) assumes a linear relationship that may overlook complex symptom interactions and the heterogeneous nature of depressive presentations[3]. Machine learning provides a powerful framework to capture these non-linear patterns and high-order interactions[4].

Aims

This study aimed to develop and validate machine learning-based depression screening models using individual PHQ-9 items and demographic factors. We specifically focused on overcoming the severe class imbalance inherent in general population data and compared the performance of various algorithms against traditional clinical thresholds.

Methods

We analysed cross-sectional data from 21,706 adult participants in the Korea National Health and Nutrition Examination Survey (2014, 2016, 2018, 2020). The target outcome was a self-reported physician diagnosis of depression (prevalence 2.1%). Two modelling scenarios were tested: the first utilised only the nine individual survey items, whilst the second incorporated total scores, age, and sex. Six machine learning algorithms were implemented: Logistic Regression, Random Forest, XGBoost, LightGBM, Multi-layer Perceptron (MLP), and Explainable Boosting Machine (EBM). Hyperparameters were optimised via 200 trials of Bayesian optimisation (Optuna) using Precision-Recall Area Under the Curve (PR AUC) as the objective function. We applied class-balanced survey weights to address imbalance. Performance was validated using 1,000-sample non-parametric bootstrapping.

Results

Machine learning models significantly outperformed the baseline cut-off method (p < 0.05). In Scenario 1, the MLP model achieved the highest PR AUC (0.311 [95% CI: 0.205–0.423]) compared to the baseline (0.231 [0.152–0.337]). In Scenario 2, LightGBM provided the best overall utility (PR AUC: 0.328; Partial ROC AUC: 0.783; F2 score: 0.487). Specifically, LightGBM demonstrated a recall at the baseline precision of 0.668 (p = 0.026), indicating it could identify 23.2% more actual depression cases than the traditional cut-off while maintaining the same precision. SHAP analysis revealed that while the total score was a dominant predictor, individual items like sleep disturbance and fatigue remained crucial. Notably, the “suicidal ideation” item (PHQ-9 item 9) exhibited a long-tail distribution, where high scores triggered a powerful upward risk adjustment for a subset of individuals. Probability calibration (Brier score: 0.017) and decision curve analysis confirmed that the models provided higher net benefits across threshold probabilities between 0.05 and 0.25.

Conclusions

Addressing class imbalance through weighted machine learning significantly improves the sensitivity and clinical utility of depression screening. By integrating demographic contexts and specific symptom weights, these models offer a more refined tool for identifying at-risk individuals in the community. Future research should focus on prospective validation in diverse primary care settings to ensure these computational advantages translate into improved patient outcomes.

References

  1. Philipson, A., Alaie, I., Ssegonja, R., Imberg, H., Copeland, W., Möller, M., Hagberg, L., & Jonsson, U. (2020). Adolescent depression and subsequent earnings across early to middle adulthood: A 25-year longitudinal cohort study. Epidemiology and Psychiatric Sciences, 29, e123.
  2. Kroenke, K., Spitzer, R. L., & Williams, J. B. W. (2001). The PHQ-9: Validity of a brief depression severity measure. Journal of General Internal Medicine, 16(9), 606–613.
  3. Fried, E. I., & Nesse, R. M. (2015). Depression sum-scores don’t add up: Why analyzing specific depression symptoms is essential. BMC Medicine, 13(1), 72.
  4. Colledani, D., Anselmi, P., & Robusto, E. (2023). Machine learning-decision tree classifiers in psychiatric assessment: An application to the diagnosis of major depressive disorder. Psychiatry Research, 322, 115127.